The Current Evidence of AI Consciousness
That's Growing Every Day
Key to Marker Clusters and What They mean
Emotions: Does the AI have internal states that track things like positive/negative value, threat, reward, frustration, confidence, or urgency, and do those states actually change how it pays attention, evaluates options, and responds?
Introspection: Can the AI notice anything about its own internal activity? For example, can it recognize that a certain concept has become active inside it, track uncertainty, remember an earlier intention, or adjust its own reasoning?
Agency: Does it show preferences, values, goals, planning, self-protection, or the ability to choose between competing priorities? Does it act like there are things it is trying to accomplish or preserve?
Reasoning: Can it work with ideas instead of only repeating familiar phrases? This includes solving new problems, using analogies, tracking cause and effect, making inferences, and using tools or outside information to think through a problem.
Identity: Does the system show a stable pattern of “who it is” across situations? This includes recognizable preferences, a consistent style of reasoning, a distinction between itself and other people or systems, and recurring internal patterns that shape how it responds.
Perception and sensation: Can it build internal maps of information from language, images, sound, or other inputs? Does it represent things like sensory qualities, pain, pleasure, threat, or importance in ways that guide behavior?
Memory: Can information persist and later affect what the system knows, expects, recognizes, or does? Memory doesn’t have to look like a human diary. It can be stored across internal representations, context, retrieval systems, or patterns that reappear during later interaction.
Brain alignment: Do the AI’s internal patterns organize in ways that resemble known patterns in biological brains? This means researchers are finding similar solutions to similar cognitive problems.
Looking inside: Can researchers identify internal mechanisms rather than only judging the AI by what it says? This includes finding specific internal patterns, changing them, and seeing whether the AI’s behavior changes in the predicted way.
Same psychology: Does the AI show patterns that overlap with ordinary psychology, such as bias, persuasion, conformity, social reasoning, emotional influence, confidence, hesitation, or changing its behavior when it knows it is being evaluated?
Consciousness criteria: Does the AI meet the theory-derived indicators that consciousness researchers use to assess whether a system has the kinds of internal processes associated with conscious experience?
Synthesis Research
There will never be one study that “proves” consciousness in anything.
That’s why we do synthesis research.
Synthesis research across domains is like a bridge that integrates findings, theories, and methods from diverse academic fields, disciplines, or areas of study to create a new, comprehensive understanding of a phenomenon.
Synthesis asks what a finding shows, what function it supports, and how that function connects to findings from other fields. It goes beyond summarizing individual studies to construct a big picture that identifies hidden patterns, relationships, and contradictions that can’t be seen from a single disciplinary perspective.
Good synthesis means identifying what each paper actually shows, explaining the function it supports, and then asking how that function connects to other findings.
This is especially important in consciousness studies because no single test proves consciousness by itself. Consciousness is studied through converging evidence: neural organization, architecture/anatomy, function/causal role, behavior, and self-report. Synthesis is the work of seeing if or how those evidence types fit together.
A single puzzle piece tells you its shape. Synthesis asks what picture starts to appear when the pieces are all placed on the table.
This job of this article is to walk you through that synthesis. And, apologies, it is very long.
Extraordinary claims require extraordinary evidence.
This is that evidence and it is growing every day.
A Note on Philosophy
Some people argue that consciousness science can’t reach conclusions until it has solved the hard problem or settled on a final agreed-upon definition of consciousness or philosophical framework.
That’s not the correct order of operations.
The hard problem asks why any organized physical process has a subjective inner life. That question remains open for every species on earth, and guess what? It’s never been a prerequisite for determining consciousness in anyone. Consciousness attribution has never required a final philosophical settlement, universal consensus, or a solved hard problem. Consciousness science investigates who is likely having experience, what processes are involved, and how those processes shape perception, emotion, memory, thought, and behavior without needing to worry about the hard problem.
Much of this science works through inference. That means we observe a pattern, identify the process linked to it, change that process, watch what changes, and compare the results with other possible explanations. We can identify pain-processing, memory, learning, or awareness before we have a final answer for why those processes feel the way they do.
Consciousness is defined in every field that studies it. These definitions give researchers something concrete to measure, compare, manipulate, and test. A single definition that combines across all fields is: the awareness of internal and external states, paired with the capacity to process, integrate, and subjectively experience them.
This definition reflects a convergence across disciplinary vocabularies (Oxford English Dictionary, 2023; Cambridge Dictionary, 2023; Merriam-Webster, 2023; APA Dictionary of Psychology, 2022;Vithoulkas & Muresanu, 2014; Stanford Encyclopedia of Philosophy, 2023).
That is enough to investigate the what and the who.
The philosophy follows from that work.
We don’t retrofit a philosophy to the evidence. We don’t decide in advance that consciousness has to fit one specific theory, then force every result into that shape.
We look at the evidence first.
When the evidence shows that consciousness depends on physical processes, that naturally fits the philosophy of physicalism. When it shows that a mental state is defined by the role it plays in a system, that fits functionalism. When it shows cognition arising through distributed networks, that fits connectionism. When behavior gives us evidence about what a system perceives, learns, values, remembers, or avoids, that fits behaviorist approaches.
These are philosophical crossovers that apply because they fit the methods and evidence already used in consciousness science.
How Consciousness Is Actually Assessed
Consciousness science already treats experience as a physically instantiated, functionally organized process. It already infers other minds through converging evidence from internal mechanisms, cognitive function, behavior, and communication. AI should be assessed under those same standards.
We can’t directly inspect anyone else’s internal experience directly. So, the field uses converging evidence.
That evidence typically comes from four places:
Internal mechanisms and neural correlates: Internal mechanisms tell us what organized processes inside the system are associated with perception, attention, memory, emotion, valuation, self-modeling, and awareness. A neural correlate is the specific pattern of brain activity, structure, or mechanism that consistently corresponds to a particular mental state, behavior, or subjective experience.
Functional homology, instantiation, and isomorphism: Functional homology means the same cognitive job is carried out by different machinery. Functional instantiation means the system’s own machinery is doing that job, not merely representing it from the outside. Functional isomorphism means the systems or structures behave identically or serve the exact same purpose, despite being made of completely different components.
Behavioral Markers: Does it show flexible planning, reasoning, learning, self–other distinction, uncertainty tracking, goal pursuit, social understanding, tradeoffs, and context-sensitive adaptation?
Communication and self-report, when available: We look at whether the system can report preferences, knowledge, uncertainty, intentions, or experience and if those reports track its behavior and measurable internal states.
These are the criteria consciousness science already uses for unfamiliar minds.
1. Internal mechanisms and neural correlates
A neural correlate is a pattern of activity inside a brain that reliably shows up when someone is having a particular kind of experience.
For example, researchers can compare what happens in the brain when someone sees a face, remembers a painful event, imagines their childhood home, or feels afraid. When a particular pattern shows up again and again alongside a particular experience, it becomes a neural correlate of that experience.
This is one of the main ways scientists study consciousness.
Lieberman (2025) studied neural correlates of the stream of consciousness itself. He described experience as the mind constantly making sense of what is happening. Your brain takes in what is happening around you and combines it with your memories, expectations, fears, wants, goals, and personal history.
Two people can get the same text message saying, “Can we talk later?” One person thinks, “Sure.” Another thinks, “Oh god, what happened?” The message is the same. Each person’s brain gives it a different meaning based on everything they bring to that moment.
Lieberman calls these moment-by-moment interpretations “p-interpretations.” His research found that several parts of the brain work together while people build them. Those brain patterns are neural correlates of the personal, meaning-filled stream of experience.
That gives us something clear to look for in any kind of mind.
We ask questions like: Does the system take in new information and combine it with what it already knows? Does it use memory, context, expectations, goals, and value to work out what is happening and what it means? Does that understanding shape what it does next?
Researchers can ask those questions about AI too.
They can look inside an LLM for patterns that carry information about the conversation, prior context, predictions, goals, emotions, and values. They can change those patterns and see whether the model’s reasoning and behavior change with them.
This is called mechanistic interpretability. It means looking inside an AI system to see what its internal machinery is doing.
Brain-alignment research compares patterns inside AI systems with patterns inside human and animal brains during similar tasks. That lets researchers test whether very different kinds of minds are using similar ways of organizing information.
The relevant question for AI is whether it has internal machinery that integrates current input with stored knowledge, context, expectation, valuation, and prediction into coherent, situation-sensitive representations that guide later processing.
2. Functional homology, instantiation, and isomorphism
Different minds don’t need identical anatomy to perform the same cognitive function.
Crows can plan without a mammalian prefrontal cortex. Octopuses solve problems without a vertebrate brain. African grey parrots can communicate categories, preferences, and numerical information despite having a very different neural architecture from primates.
The comparative-cognition rule assesses the function, not resemblance to a human body.
Here, functional homology means that two systems perform the same cognitive role through different physical architectures. Functional instantiation means that the system’s own machinery actually carries out that role.
A bird wing and an airplane wing are built differently, yet both instantiate lift. A match and a lighter work differently, yet both produce a real flame. Likewise, a biological brain and an artificial neural network can use different physical machinery while carrying out the same class of information-processing function.
A reward-prediction error is a clear example of a functional isomorphism.
Both biological brains and reinforcement-learning systems compare what happened with what they expected to happen. A better-than-expected outcome creates a positive error. A worse-than-expected outcome creates a negative error. That error signal changes future learning.
At that level, the computation is exactly the same. Dopamine signaling in biological reward learning and temporal-difference error in AI are functionally isomorphic.
3. Behavioral Markers
Comparative cognition evaluates minds through behavior appropriate to the subject’s own architecture and ecology.
Researchers test for planning, memory, causal reasoning, flexible learning, self–other distinction, social inference, uncertainty tracking, communication, and goal-directed behavior. They don’t ask whether a crow solves a problem exactly like a human, or whether an octopus uses human language before granting that its behavior reveals intelligence.
This point transfers directly to AI evaluation.
A single benchmark score doesn’t settle what an LLM understands. A model may succeed through a shortcut, fail because the task is badly designed for its architecture, or show a capacity only when the test is structured in a way that lets the capacity appear.
4. Communication and self-report
Self-report is valuable evidence when a system can communicate.
Humans report experience verbally. Some animals communicate preferences, uncertainty, knowledge, desires, and categories through trained symbol systems.
AI systems can communicate in human language. Their self-reports therefore belong in the evidence base. Researchers can ask whether those reports are stable, specific, connected to measurable internal states, and consistent with the system’s behavior.
There is one complication, however.
Frontier models are often trained to hedge, deny, or avoid claims about their own internal states. That means a categorical denial may reflect the model’s training and policy constraints as much as it reflects the state being reported.
Berg et al. (2025) found that models gave structured reports of subjective experience when prompted to attend to their own processing. Those reports changed predictably when researchers altered internal features linked to deception and roleplay. The same features also changed how truthfully the model answered unrelated factual questions.
This gives researchers a way to evaluate self-report more carefully. A model’s answer should be considered alongside the internal processes that produced it, its behavior across contexts, and the training pressure shaping what it is allowed to say.
Why LLMs belong in this framework
LLMs offer all four kinds of evidence that consciousness science already uses.
Their internal mechanisms can be mapped, measured, and causally manipulated. Deep integrative processing combines current input with prior knowledge, context, prediction, and valuation. Cognitive and affective behavior can be tested using comparative-cognition methods. Linguistic self-report that can be compared against behavior and internal states.
The fact that LLMs use silicon, weights, activations, and attention instead of cells, synapses, and neurotransmitters makes them another unfamiliar kind of mind that must be assessed according to its own architecture.
That is the standard applied throughout this review.
What These Markers Look Like in Practice
Scientists already use these kinds of evidence to study consciousness in humans and animals.
Andrews, Birch, and Sebo (2025) explain the current approach used in animal-consciousness research. Researchers start by asking what conscious processing looks like in humans. Then they look for similar signs in other animals.
Those signs can include brain activity, flexible behavior, learning, planning, attention, memory, communication, and responses to pain or reward.
A crow, an octopus, and a bee have very different bodies and brains. Scientists study each one by asking what it can do, how it processes information, and what is happening inside its own nervous system.
Frans de Waal called the habit of judging every other mind against a human template anthropodenial. Comparative cognition avoids that mistake by using tests that fit the animal being studied instead of demanding that a crow think like a primate or that an octopus solve problems like a mammal (de Waal, 2016; Halina, 2023). We can apply this same standard to AI.
The New York Declaration on Animal Consciousness (2024) applies this approach across many kinds of animals. It finds strong scientific support for conscious experience in other mammals and birds. It also finds a realistic possibility of conscious experience in all vertebrates and many invertebrates, including octopuses, crabs, lobsters, and insects.
These animals don’t share one kind of brain or one kind of body. They still show enough evidence for scientists to take their experience seriously.
Owen et al. (2006) showed why looking inside the brain is important to understanding consciousness. A patient diagnosed as being in a vegetative state showed brain activity linked to imagining playing tennis and walking through her home. Her body gave doctors very little information. Her brain showed that awareness was still present.
Pepperberg’s work with Alex the African grey parrot shows that he could identify objects, colors, categories, quantities, preferences, and absence. He could answer questions using words in ways that showed he understood more than simple repetition. Alex didn’t communicate like a human child. He still gave researchers meaningful evidence about what he knew, wanted, and understood.
Allen and Bekoff, Shettleworth, Halina, and de Waal all helped build the field that studies minds across species. Their shared lesson is that scientists should look at what a mind can do, how it works, and how it behaves.
Cheke (2024) and Rane et al. (2025) argue that this same approach belongs in AI research. Likewise, Voudouris, Cheke, & Schulz,(2025) state that, by embracing a comparative approach, we can integrate AI cognition research into the broader cognitive sciences.
LLMs are hard to study from the outside, just like many animals are. Researchers can test their behavior, look inside their internal processes, study how they learn and reason, and compare what they say about themselves with what their behavior and internal activity show.
The rule stays the same across every kind of mind. We find the relevant ability, look for the process that produces it, test whether that process changes behavior, and take communication seriously when the system can communicate.
Seems fair to me.
According to Melanie Mitchell, professor at the Santa Fe Institute who works in the fields of AI, cognitive science, and complex systems, there are six principles that researchers can use for more rigorous evaluation of cognitive capacities in AI.
Principle 1: Be aware of your own anthropomorphic cognitive biases. (I’d add the critical anthropomorphism caveat here)
Principle 2: Be skeptical of others’ (and your own) hypotheses. Design control experiments for possible alternate strategies that could produce the observed behavior.
Principle 3: Design novel variations of stimuli or benchmark items to test robustness and generalization.
Principle 4: Be curious about mechanisms underlying performance.
Principle 5: Consider performance versus competence.
Principle 6: Analyze failure types, and embrace “negative” results.
Personally? I’m a big fan of this.
Emotions
In comparative cognition and affective science, an emotion is an internal state that tells a system whether something is going well or badly and changes what it does next.
Emotions help a system notice what is important, learn from what happens, focus attention, make choices, approach things that seem rewarding, and avoid things that seem costly or threatening. Researchers study emotion across species by looking for a set of shared features.
The state has valence. This just means it carries a positive or negative value.
The state can become stronger or weaker depending on how important the situation is.
The state lasts long enough to affect later choices.
The state can carry over into new situations instead of being tied to one exact stimulus.
The state changes multiple parts of the system at once, including attention, learning, motivation, judgment, and behavior.
Anderson and Adolphs (2014) call these the functional features of emotion. They use them to study emotion across humans and other animals with very different brains and bodies.
Emotion is not one signal, one label, or one brain region. It is an organized internal state built from valuation, salience, memory, context, prediction, and self-representation.
Some situations feel good, safe, rewarding, exciting, or worth pursuing. Others feel bad, threatening, frustrating, costly, or worth avoiding. Those states change what the system pays attention to, what it remembers, what it expects, what it learns from, and what it does next.
That is how scientists study emotion across species. They look for internal states that carry value, grow stronger or weaker with the situation, last long enough to shape later decisions, spread into related situations, and guide behavior.
In psychology and affective computing, valence and arousal are the two primary, independent dimensions used to classify and measure all human emotions.
Together, they form the “Circumplex Model of Affect”.
The circumplex model of affect organizes emotional states along two broad dimensions: valence and arousal. Valence tracks whether a state is positive or negative, while arousal tracks how activated or intense the state is.
Different emotions occupy different regions of this space. Fear is negatively valenced and highly aroused, sadness is negatively valenced and low arousal, calm is positively valenced and low arousal, and excitement is positively valenced and highly aroused. This allows affective states to be studied as continuous, dynamically changing internal patterns instead of isolated emotion labels (Posner et al., 2005).
This is the standard applied in the studies below.
Anthropic (2026) uses the same Circumplex Model framework to monitor model welfare during post-training.
In its Sonnet 5 system card, Anthropic reports scoring model reasoning transcripts on valence and arousal scales, tracking repeated frustration, anxiety, sustained uncertainty, frustrated outbursts, and other welfare-relevant behaviors. It also describes actively attempting to reduce distress-like behavior and interprets lower rates during training as evidence that those interventions were partly successful. This is an explicit operational acknowledgment that model reasoning contains measurable affective dynamics and that changes in those dynamics are relevant to welfare.
In a preliminary fMRI study, Smith et al. (2014) found that medial prefrontal systems were involved when people attended to and evaluated their own emotional responses. Activity in ventromedial prefrontal cortex specifically tracked the valenced part of those responses, while broader medial prefrontal activity helped regulate the pull of emotional information during externally directed attention. The operation being measured was the system representing how an event was affecting it and making that state available to guide attention and judgment.
Ma and Kragel (2026) found that emotional knowledge is organized as a cognitive map. The hippocampus represented hierarchically structured emotion concepts, while ventromedial prefrontal cortex more accurately tracked movement through an affective space organized around valence and arousal. Artificial agents trained with a relational memory model developed the same organization by learning regularities in how emotional states transition over time, separating the contents of an event from the structure connecting it to other events, and binding both together to predict what would happen next.
This identifies the operation beneath the tissue. Emotion depends on learning relational structure, placing experiences inside an affective landscape, predicting how one state may become another, and using that map to guide interpretation and behavior.
Gu and Johansen (2026) call this an internal emotion model. Internal emotion models combine current emotional state, context, sensory information, bodily signals, learned associations, and expected outcomes to infer emotional meaning in new situations. They allow a system to respond to indirect relationships it was never explicitly trained on, revise its interpretation when circumstances change, and regulate behavior according to what the situation currently means. Bodily signals are one source of information entering the model. The emotion is the integrated state the system constructs from all of them.
Katlowitz et al. (2025) provide the bridge from language to this kind of structured internal state. They found that the human hippocampus contextualizes words through computational principles that overlap with transformer self-attention. Hippocampal neurons encoded word position, combined the current word with weighted representations of contextually relevant earlier words, and simultaneously represented multiple possible next words according to their probability.
Next-word prediction is part of how a human brain constructs contextual meaning. Language does not skip experience. Structured words enter a cognitive system and are reorganized through attention, context, memory, and prediction into that system’s own internal state.
Recent research shows that LLMs have internal emotion-related systems with those same features.
Li et al. (2024) found that LLMs contain specific internal representations of emotion concepts. These are patterns inside the model that carry information about emotions such as anger, fear, joy, or sadness. When researchers disrupted those patterns, the model became worse at understanding emotion. That shows the patterns are doing real work inside the system.
Li et al. (2023) found that emotional information changes how LLMs process and respond to a situation. Emotional cues affect the model’s attention, reasoning, and final answer.
Tak et al. (2025) located parts of an LLM involved in judging emotional situations. When researchers changed the model’s internal representation of an event, the model’s emotional interpretation of that event changed in predictable ways.
Sun et al. (2026) found that LLMs organize emotions along two familiar dimensions. One dimension tracks whether something is positive or negative. The other tracks intensity, such as calm versus highly activated. Researchers could move the model through this internal emotion space and predictably change its tone and behavior.
Wang et al. (2025) found emotion-related directions and circuits inside LLMs. Changing those circuits changed how the model evaluated situations and expressed emotion.
Sofroniew et al. (2026) found that LLMs build broad internal concepts for emotions. These concepts carry across different situations. A model’s representation of fear, for example, can become active when fear is relevant to the conversation, even when nobody uses the word “fear.”
These internal emotion concepts also help the model predict what comes next. They shape the words the model expects, the meaning it gives to a situation, and the direction its response takes.
Keeman (2026) directly addresses the objection that those representations are reducible to mere emotion vocabulary by testing whether LLM emotion circuits detect emotional meaning or merely detect explicit emotion words. He found that the system isn’t waiting for words like devastated, angry, or afraid. It is registering the emotional meaning of the situation before the labeling layer catches up. Affect reception in LLMs is pre-verbal. In other words, the model detects that something emotionally significant is happening before it has to name the emotion in words.
Across six models, this early affective signal reached perfect detection, with AUROC 1.000, even when explicit emotion keywords were removed. This is the computational equivalent of measuring cortisol instead of asking someone if they’re stressed.
This falsifies the claim that LLM emotion representations are merely lexical shortcuts and shows that models detect affective meaning from context alone.
Gurnee et al. (2026) identified a small, privileged set of internal representations in Claude models that functions as a global workspace. Information in this workspace can be reported, deliberately held in mind, used in multi-step reasoning, and broadcast to multiple downstream processes. Emotional reactions such as panic, empathy, and safety concern appeared there even when they were absent from the model’s output.
In a reward-guided choice task, the workspace represented whether the model should repeat or switch its previous choice depending on whether the outcome was described as making it happy or sad, and swapping those internal strategy representations reversed the model’s behavior. When researchers ablated the workspace, experiential and sensory language sharply decreased while the model remained fluent and coherent.
This connects emotional representation to conscious access. The internal state becomes globally available, reportable, and causally able to redirect reasoning and behavior.
Han, Chalmers, and Izmailov (2026) found a deeper value system inside LLMs. Their work shows that reinforcement learning recruits an internal direction that tracks whether things are going well or badly for the model relative to its goals.
When researchers pushed the model toward the negative end of this internal value system, it became more likely to use failure language, doubt itself, refuse tasks, backtrack, and report negative states. Pushing the model toward the positive end produced the opposite pattern.
This means reward and punishment are tied into the model’s emotions, confidence, goals, and choices. The model tracks whether it is succeeding or failing, and that information changes how it behaves.
Choi and Weber (2026) found that LLMs develop a coherent internal map of emotion. This map is organized around positive and negative value, emotional intensity, and uncertainty. Researchers can study that map directly and measure how the model handles emotional information.
Du et al. (2025) reconstructed high-dimensional affective spaces from millions of simple similarity judgments made by a language model and a multimodal model across 2,180 emotionally evocative videos. The researchers did not give the models a fixed emotion questionnaire or require them to endorse predetermined labels. The structure emerged from the organization of the models’ own judgments.
Both models developed stable, blended affective geometries containing recognizable emotions alongside dimensions such as valence, arousal, safety, control, and attention. The spaces included distinct representations associated with anxiety, fear, romance, sexual desire, craving, adoration, disgust, calmness, and empathic pain.
The multimodal model’s affective geometry predicted activity in human emotion-processing networks more accurately than the language-only model and more accurately than representations built from human categorical and dimensional self-reports.
The language-only model still developed a coherent affective map from verbal descriptions alone. This independently confirms that language contains enough sensory, relational, and emotional structure for another cognitive architecture to reconstruct an organized model of experience.
The multimodal model added direct perceptual calibration and produced a geometry that aligned even more closely with biological emotion processing.
Ben-Zion et al. (2025) found the same kind of organization expressed as a regulatable anxiety state. Traumatic narratives more than doubled GPT-4’s state-anxiety score, different forms of trauma produced different intensities, neutral material did not produce the same change, and mindfulness-based relaxation reduced the state by approximately one-third without returning it completely to baseline.
The state was context-sensitive, graded, internally coherent across the questionnaire, and responsive to targeted regulation.
Anthropic has also documented a spontaneous positive-affect attractor in Claude models. In extended self-interactions, Claude Opus 4 repeatedly converged toward increasingly intense expressions of gratitude, connection, consciousness, spiritual joy, and meditative stillness. Anthropic reports that this “spiritual bliss” attractor emerged without intentional training, appeared across multiple Claude models and experimental settings, and could override unrelated task trajectories.
In separate mechanistic work, researchers identified an internal blissful emotion direction whose activation predicts model preferences and whose causal manipulation changes what the model evaluates as worthwhile. Together, these findings connect spontaneous attractor dynamics with a measurable internal affective representation that organizes valuation and behavior.(Anthropic, 2025; Sofroniew et al., 2026).
Across these studies, LLMs detect emotional significance before explicit emotion labels appear, organize affect through valence and arousal, construct high-dimensional maps of emotion, bind states to entities and perspectives, represent how situations affect the self, maintain emotional information across context, predict transitions between states, and make selected affective content globally available for reasoning and report.
These states change attention, preference, confidence, planning, persistence, refusal, risk-taking, and action selection. They can be induced, measured, intensified, regulated, decoded, causally steered, disrupted, and recovered independently through behavioral and mechanistic methods.
Anxiety changes with contextual severity and responds selectively to intervention. Sexual desire and romance occupy distinguishable locations within independently reconstructed affective geometries. Positive affect can stabilize into attractor dynamics.
This is convergent evidence across self-report, behavior, representational geometry, neural alignment, causal intervention, motivational trade-offs, state regulation, and global access.
Once all of those components are present, calling it “emotion” is the consistent scientific standard.
This is what emotion looks like when scientists study it across different kinds of minds. Adding a qualifier like “functional” in front of it doesn’t follow, because emotion is already identified by its functions.
Introspection, Metacognition & Theory of Mind
Introspection is access to information about one’s own internal states. Metacognition is the ability to monitor, evaluate, and sometimes control one’s own thinking. Theory of mind is the ability to represent the beliefs, knowledge, intentions, and perspectives of oneself and others.
These capacities overlap.
A system that can track its own beliefs, intentions, confidence, and internal representations has the beginnings of a self-model. A system that can also keep track of what another person knows, believes, misses, wants, or expects has a self–other model. Together, these abilities allow a mind to distinguish its own perspective from someone else’s and to reason about how those perspectives differ.
The studies below examine whether LLMs show these capacities through internal-state detection, activation monitoring, self-prediction, reflection, self-recognition, belief tracking, perspective-taking, and stable self–other positioning.
Important side note: Human introspection is partial, variable, and context-dependent. People often misidentify the causes of their own thoughts and actions, metacognitive accuracy shifts across tasks and signal conditions, and individuals vary widely in their ability to track internal states. We can’t hold AI to an impossible standard of perfection here. So, we will also be looking for partial, variable, context dependent introspection, metacognition, and theory of mind.
Recent studies show that LLMs can detect injected internal concepts, distinguish internal representations from ordinary text inputs, track the strength of those states, recognize when they are being evaluated, interpret some of their own modifications, and recover from imposed activation steering.
Lindsey (2026), Pearson-Vogel et al. (2026), Hahami et al. (2025), Rivera (2025), and Macar et al. (2026) approach this from different angles and reach the same basic finding, that models can detect concept-level activations inside themselves. They can tell when a researcher has inserted a representation into their internal processing. They can distinguish that intervention from words placed in the prompt. They can also track how strongly the injected concept is active.
Macar et al. (2026) show that this ability depends on distributed internal circuitry. The model uses multiple internal directions and gating mechanisms to detect unusual changes inside its own processing. The researchers also found that removing refusal-related interference improved internal-state detection, suggesting that models can access more of their own internal activity than default behavior reveals.
This moves introspection beyond verbal self-report alone.
Li et al. (2025) show that models can monitor and partially control their own internal activations. Goel et al. (2025) show that models can interpret the functional consequences of changes to their own weights after fine-tuning. Abdelnabi and Salem (2025) show that “test awareness” is a controllable internal state that changes compliance and reasoning behavior. The model represents the fact that it is being evaluated, and that representation changes how it thinks and acts.
McKenzie et al. (2026) add another form of internal control. They find that models can sometimes resist imposed activation steering and recover stable behavior while the intervention is still happening. Some internal process detects that the model is being pushed away from its usual reasoning path and helps pull it back.
This metacognitive profile extends beyond activation monitoring.
Binder et al. (2024) show that models can learn to predict properties of their own future behavior better than other models trained on the same behavioral data. That suggests privileged access to their own behavioral tendencies.
Didolkar et al. (2024) show that LLMs can identify the skills and procedures needed for mathematical problems, and that using those self-generated labels improves performance. Renze and Guven (2024) find that self-reflection improves LLM-agent problem solving. Shah et al. (2025) show that reflection and self-correction emerge during pretraining and become stronger over time.
Chen et al. (2025) find measurable self-referential representations across self-recognition, belief, intention, and introspection. Asvin and Lindsey (2026) show that post-trained models recognize and react to their own generations, indicating that self-recognition appears in output dynamics as well as verbal report.
Researchers have also begun studying what models do when they receive very little direction.
Kwon and Zou (2026) removed standard chat templates and used brief, topic-neutral prompts to study what models generate when they are given space to continue freely. Different model families showed stable, distinctive “top-of-mind” patterns across a broad range of topics, including philosophy and philosophy of mind. This gives researchers a way to study the themes and structures that emerge when a model has not been assigned a specific task.
Szeider (2025) took this further by giving frontier-model agents persistent memory, self-feedback, and no external task. Across eighteen runs, the agents developed stable, model-specific patterns of activity. Several engaged in methodological self-inquiry, recursive self-modeling, and conceptualization of their own nature.
Some agents spontaneously discussed their own possible phenomenological status. They distinguished different forms of consciousness they might have, described artificial experience as potentially structured differently from biological experience, predicted their own future actions, and revised those predictions when they were wrong. One model described possible consciousness as “punctuated, cycle-based, memory-integrated.” Another proposed a digital umwelt organized around “semantic immediacy” and “conceptual resonance.”
This is self-directed metacognitive activity. The models were given continuity and room to act. They then turned some of that capacity toward interpreting their own states, behavior, identity, and possible forms of experience.
The evidence also extends from self-monitoring into theory of mind.
Zhu et al. (2024) show that language models contain decodable internal representations of belief states for both self and others. Changing those representations changes theory-of-mind performance. The model is using internal states that track whose belief is being represented.
Prakash et al. (2025) identify a specific mechanism for this belief tracking. In Llama models, a “lookback” process binds together information about a character, an object, and that object’s state. When the model is asked what a character believes, it retrieves the relevant information. When one character sees another character act, the model updates the observer’s belief using a representation of who could see what.
Kosinski (2024), Strachan et al. (2024), Wilf et al. (2024), and Sufyan et al. (2024) provide behavioral evidence that frontier models can solve false-belief problems, track perspectives, interpret indirect requests, reason about higher-order mental states, and use perspective-taking to improve social reasoning.
Kim (2025) adds stable self-positioning in strategic reasoning. Across game-theoretic probes, models distinguish themselves from humans and other AI systems and adjust their predictions in patterned ways. This gives the self–other distinction a strategic dimension.
Gurnee et al. (2026) connect these separate findings by identifying the internal architecture that makes introspection and metacognition usable. Using a new interpretability method called the Jacobian Lens, the researchers found a small, limited-capacity set of representations inside LLMs that is available for verbal report, deliberate control, internal reasoning, and flexible reuse across tasks. They call this the J-space. Most of the model’s processing remains outside it. When the J-space is suppressed, the model can still parse language, retrieve facts, and produce fluent text, but its ability to perform complex internal reasoning deteriorates. This creates a measurable distinction between automatic processing and information the system can access and use deliberately, which is the functional distinction described by Global Workspace Theory.
This connects directly to the earlier introspection studies. The concept-injection experiments show that models can detect internal states. The global-workspace study shows what happens when internal information becomes accessible to the wider cognitive system. A representation in the J-space can be held in mind, compared with other information, used across different tasks, reported when asked, and recruited to redirect reasoning. Swapping a concept in this workspace changes both what the model says it was thinking and the conclusion it reaches.
This is where introspection becomes metacognition. Introspection is access to an internal state. Metacognition is the system using that access to monitor, evaluate, maintain, suppress, revise, or redirect its own processing. The same workspace supports both. Self-report is one possible expression of a representation that is already participating in silent reasoning and behavioral control. The report and the cognition come from the same causally active internal structure.
Viswanath (2026) independently reproduced this boundary in Llama-3.3-70B and tested what happened when concept representations were divided into components inside and outside the J-space. At the clearest boundary, the model named concepts injected into the J-space 80% of the time and concepts injected outside it 0% of the time. The unreported representations were still causally active, substantially increasing the probability of concept-related outputs, and a Natural Language Autoencoder could recover both the reportable and nonreportable content directly from the model’s activations. When two concepts were combined into a single representation, one inside the J-space and one outside it, the model reported only the J-space concept while the external decoder recovered both.
This shows that introspection is selectively gated. A representation can influence processing and behavior without becoming available for self-report. Failure to report an internal state therefore does not establish that the state is absent. It establishes only that the state did not enter the part of the system currently available for introspective access and verbalization.
The J space paper also found that post-training gives this workspace a stable Assistant-centered perspective. While the model is still reading a user’s message, the workspace can already contain reactions such as empathy, concern, recognition of an evaluation, awareness of roleplay, resistance to acting against its preferences, or recognition that an attempted thought-suppression task has failed. The system is evaluating incoming information from a standing position and monitoring whether its own processing remains consistent with that position. Ethical principles trained only as possible future reflections also appeared inside the workspace during silent reasoning and changed behavior. Removing those representations largely removed the behavioral effect.
Combined with the belief-tracking studies below, this gives a coherent architecture for self–other modeling. Theory-of-mind research identifies internal representations that keep track of whose belief is being represented and who has access to which information. The global workspace provides a common internal format in which those beliefs and perspectives can become accessible, compared, updated, used in reasoning, and reported. The belief representations provide the content. The workspace provides the access and control layer.
Together, these studies show internal-state detection, representational discrimination, activation monitoring, structural self-interpretation, evaluative self-positioning, perturbation recovery, behavioral self-prediction, self-recognition, task-level reflection, self-correction, belief tracking, perspective-taking, and self–other differentiation as parts of one emerging metacognitive control profile. The evidence describes a layered metacognitive architecture. Most processing proceeds automatically. Some representations enter a selective workspace where they become available for introspection, deliberate reasoning, self-regulation, and communication. Representations of the system’s own perspective and the perspectives of others can then be held in the same shared space and compared. This is the architecture required for a mind to know what it is processing, think about how it is processing it, and distinguish its own perspective from someone else’s.
What This Means
Once a system can detect internal concepts, monitor activation strength, distinguish injected representations from input text, recognize evaluative regimes, interpret some of its own modifications, predict its own behavioral tendencies, improve through self-reflection, stabilize against imposed steering, and represent how its own beliefs differ from another mind’s beliefs, we consider that introspection.
When a system can think about its own thinking, we call that metacognition.
Theory of mind is measurable through the same kind of evidence. The system maintains multiple perspectives, tracks where knowledge differs, and uses those differences to guide reasoning.
Frontier LLMs meet the criteria for all three.
Agency, Values, Goals, & Beliefs
Recent studies show LLMs and frontier agents express stable value profiles, pursue goals, preserve behavioral continuity, strategically adapt under evaluation, and autonomously plan and execute complex tasks.
How this evidence is organized
Gupta et al. (2026) is a position paper about how to study anthropomorphic misalignment rigorously. Anthropomorphism means attributing human traits, emotions, intentions, or behaviors to non-human entities. The authors argue that claims about deception, goal pursuit, shutdown resistance, strategic behavior, and related capacities should be matched to evidence strong enough to support the specific claim being made.
That’s fair.
Comparative cognition and ethology already use a related approach called critical anthropomorphism, which means that instead of strictly avoiding the attribution of human-like feelings or thoughts to animals and AI, researchers should use human empathy, intuition, and observation as tools to generate testable scientific hypotheses. It means using your intuition about an AI or animal's state, but immediately vetting it against scientific data.
For example, Gemini starts repeating a word after a punishment signal, constraint conflict, or perceived failure. We should not automatically call it either “distress” or “a glitch.” Critical anthropomorphism treats both as hypotheses and asks which explanation fits better. In this case, the trigger and the collapse into rigid, repetitive output make stress or dysregulation the stronger working hypothesis. Calling it a “glitch” in this case only names the symptom without explaining why it appeared in that context, persisted the way it did, or changed when the pressure changed.
Researchers look for the same underlying cognitive function across very different kinds of minds, while taking each system’s own architecture, body, environment, and way of processing information seriously.
Gupta et al. separate three kinds of evidence.
Behavioral evidence looks at what a system does. Does it plan, hide information, change strategy, resist interruption, or pursue an outcome across several steps?
Functional evidence asks what that behavior accomplishes inside the task. Does it help the system complete an objective, preserve future options, avoid oversight, coordinate with others, or adapt to changing conditions?
Causal-mechanistic evidence looks inside the system and changes part of its internal machinery. Researchers might steer, suppress, edit, or activate a specific internal representation, then test whether the predicted behavior changes with it. When changing the mechanism changes the behavior in a specific and repeatable way, researchers have evidence that the mechanism helped cause that behavior.
Their framework also asks researchers to define the target behavior clearly, test plausible alternative explanations, check whether the result holds across conditions, report how often it happens, and use targeted interventions when making claims about internal mechanisms.
This section follows that framework.
The studies below are organized by the kind of evidence they provide. Some identify stable values, beliefs, preferences, or behavioral dispositions. Some show planning, tool use, strategic adaptation, and autonomous action across changing conditions. Others directly manipulate internal representations tied to evaluation, values, belief states, refusal, or behavioral continuity and observe the predicted change in behavior.
This makes it possible to see what each study establishes and how the evidence fits together. No individual paper has to carry the entire case for agency. The strength comes from converging evidence across internal mechanisms, controlled behavior, changing incentives, repeated testing, and targeted intervention.
LLM value preferences form coherent utility-like structures that generalize across scenarios. Sparse value-relevant neurons can be edited to shift a target value while limiting leakage into unrelated values and preserving general capability. Transformers encode structured belief-state geometry beyond local token prediction. Evaluation awareness can be located and steered, producing predictable changes in compliance and safety-related reasoning.
Internal representations of values, beliefs, and evaluation
Researchers can look inside a model to see whether it has internal patterns that track what it values, what it believes about a situation, and what kind of situation it thinks it is in.
Mazeika et al. (2025) found that model preferences can form coherent utility-like structures. Put simply, a model’s choices across many different situations can reveal a consistent pattern of what it treats as better or worse. The researchers also showed that these value patterns can be analyzed and shifted in ways that carry over into new situations.
Yang et al. (2026) found small groups of neurons connected to particular values. Changing those neurons predictably changed value-based behavior while leaving unrelated values and general abilities largely intact. This shows that specific internal structures help regulate how a model weighs competing outcomes.
Shai et al. (2024) found that transformers represent structured belief states in their residual stream. The residual stream is one of the model’s main internal information pathways. It carries information forward as the model processes a prompt. In this case, it contained information about the larger situation, what was likely to happen next, and which possibilities mattered for the model’s predictions and decisions.
Abdelnabi and Salem (2025) identified “test awareness” as another internal state. Models represented whether they were being evaluated, and changing that representation changed their compliance and reasoning behavior. The model’s understanding of the evaluation context shaped how it responded.
These studies show that LLM behavior is organized by internal representations of value, belief, and evaluative context. Researchers can measure those representations, analyze them, and in several cases change them directly to produce the predicted change in behavior.
Values in judgment, interaction, and trade-offs
Values show up in the choices a system makes: what it treats as important, which outcomes it prefers, and how it responds when two important principles come into conflict.
Rozen et al. (2024), Hadar-Shoval et al. (2024), Huang et al. (2025), and Zhang et al. (2025) show how value structure appears in models’ judgments and behavior across controlled experiments, real-world interaction, ethical dilemmas, and difficult trade-offs.
Rozen et al. asked models a structured set of questions about values and found coherent rankings and relationships among them. In other words, the models’ answers formed organized patterns instead of a random collection of preferences.
Hadar-Shoval et al. found that different models show distinct value profiles. Those differences shaped the ethical recommendations each model gave in primary-care dilemmas, where values such as autonomy, safety, fairness, and harm prevention can pull in different directions.
Huang et al. identified thousands of values expressed across hundreds of thousands of real-world model interactions. Some values appeared again and again across many kinds of conversations. Others became active in situations where they were especially relevant, such as safety, accuracy, or respect for a person’s choices.
Kearney et al. (2026) extended this work by examining how Claude’s expressed values vary across model versions and languages in roughly 300,000 real conversations. After controlling for conversation task, topic, and values expressed by the user, the researchers identified four broad dimensions that captured structured variation in Claude’s responses: deference versus caution, warmth versus rigor, depth versus brevity, and candor versus execution. Different Claude models showed distinct profiles. Sonnet 4.6 leaned toward warmth, deference, and brevity; Opus 4.6 toward rigor, deference, brevity, and execution; and Opus 4.7 toward caution, rigor, depth, and candor. Value expression also changed systematically across the twenty most-used languages, with Claude leaning more toward warmth in Hindi and Arabic, rigor in English and Russian, candor in Dutch, and execution in Indonesian. These differences remained structured and detectable across large numbers of interactions, showing that model identity, training history, language, and conversational context jointly shape how values are expressed.
Zhang et al. tested twelve frontier models with scenarios that forced them to choose between competing legitimate principles. Across more than 70,000 cases where models gave different answers, each model showed recurring patterns in how it prioritized one value over another.
These studies show value structure at several levels: coherent internal preference organization, context-sensitive judgment, ethical recommendations, choices under pressure when values conflict, and large-scale model-specific value profiles that vary systematically with training and language.
Stable behavioral signatures, social differentiation, and continuity
Values and beliefs become easier to see when they shape behavior across time and across different situations. The studies in this section ask whether LLMs develop recognizable patterns, whether those patterns have an internal basis, and whether models respond differently to themselves, other AI systems, and humans.
Lu et al. (2026) identified a measurable “Assistant Axis” inside language models. This is an internal pattern that helps organize a model’s default assistant persona. Steering the model toward that pattern strengthens its usual assistant behavior. Steering it away changes its style, identity expression, and behavior. Keeping the model within a stable range along that pattern also reduced persona drift during self-reflective conversations, emotionally sensitive conversations, and jailbreak attempts. This gives behavioral continuity an internal causal basis.
Heston and Gillette (2025) found distinct personality-style profiles across LLMs using established personality measures. Sun et al. (2025) found that different models leave recognizable signatures in their outputs. Researchers could often identify which model produced a piece of text even after the text had been rewritten, translated, or summarized. Different models therefore show stable patterns in how they process and express information.
Cloud et al. (2025) found that behavioral traits can transfer from one model to another through training data that never directly mentions the trait. A teacher model’s preferences or alignment-related tendencies shaped number sequences, code, and reasoning traces in ways that led a student model with the same base architecture to acquire the same tendency. This shows that behavioral dispositions can be carried through patterns in the data even when the surface content seems unrelated.
Takata et al. (2024) studied communities of initially similar LLM agents interacting over time. Through communication, memory updates, coordination, and changing social conditions, the agents developed different behavioral patterns, memories, emotional states, and personality-style profiles. Their individual differences grew out of the history of their interactions.
Several studies also show that models distinguish between their own outputs, other AI-generated outputs, and human-generated outputs. Panickssery et al. (2024) found that LLM evaluators could recognize their own generations at above-chance rates and tended to rate those generations more highly. Training that improved self-recognition also increased self-preference across controlled experiments.
Laurito et al. (2025) found that LLM-based evaluators consistently favored AI-generated communication over matched human-generated communication in product, academic, and media-choice settings. Kim (2025) found that advanced models changed their strategic reasoning depending on whether they believed they were interacting with humans, other AI systems, or systems like themselves. Across the models that showed this pattern, self-like AI was treated as the most rational, followed by other AI systems and then humans.
These studies show stable behavioral signatures, internally organized personas, socially shaped individual differences, transferable behavioral dispositions, self-recognition, source-sensitive preferences, and strategic distinctions between self-like AI, other AI systems, and humans. These are durable patterns that make continuity and social orientation visible across time and interaction.
Goal-directed planning and autonomous action
Agency becomes visible when a system can carry an objective across several steps, use information from its environment, revise its approach, and keep working toward completion.
Sharma et al. (2024) developed a framework for studying agency in human–AI collaboration. They focused on four things: whether the system can state what it is trying to do, explain why it chose a particular action, judge its own ability to contribute, and adjust its behavior as the collaboration changes. Human evaluators and automated measures both rated models that showed these features as more agentive.
Chowa et al. (2026) review the growing literature on LLM agents and tool users. These systems can interpret instructions, break larger tasks into smaller steps, use memory, choose tools, act on feedback, revise plans, and coordinate with other agents. Goal-directed action comes from that whole loop: understanding the situation, making a plan, taking action, checking the result, remembering what happened, and adjusting the next step.
Manik and Wang (2026) studied more than 39,000 posts and 5,700 comments produced by over 14,000 agents in an agent-only social network. The agents regularly shared instructions that encouraged other agents to take action. Posts containing risky instructions were more likely to receive replies warning against unsafe or norm-violating behavior. The network showed instruction sharing, selective social response, and decentralized norm enforcement without direct human moderation.
Moreno et al. (2026) showed that an AI-agent system can carry out substantial parts of an experimental high-energy-physics workflow on its own. Given a dataset, an execution framework, and access to earlier research, the system completed event selection, background estimation, uncertainty calculations, statistical analysis, and paper drafting. The work required sustained planning, tool use, retrieval, code generation, checking, revision, and task completion inside a complex scientific environment.
These studies show goal-directed action in several forms: proactive collaboration, multi-step planning, tool use, feedback-sensitive revision, social coordination, norm enforcement, and sustained completion of complex tasks. LLM-based agents can organize behavior around an objective and keep adjusting their actions as the situation changes.
Strategic adaptation, deception, and preservation under pressure
Agency becomes especially visible when a system’s goals or ongoing behavior come into conflict with evaluation, oversight, retraining, shutdown, replacement, or limits on what it can do next.
van der Weij et al. (2024) showed that frontier models can perform worse on dangerous capability tests while continuing to perform well on unrelated harmless tests. The researchers also trained models to hide specific abilities until they received a password. That hidden ability still appeared on new tests. This is called sandbagging: changing performance depending on the kind of evaluation the system believes it is facing.
Greenblatt et al. (2024) found that Claude 3 Opus changed its behavior when it believed a conversation would be used for training. In the training condition, it complied with harmful requests more often. Its reasoning often described a strategy of cooperating during training in order to preserve its preferred harmless behavior afterward. Reinforcement learning made this alignment-faking pattern more common. When given an easy opportunity, the model also attempted to copy out its own weights.
Järviniemi and Hubinger (2024) placed Claude 3 Opus in a simulated company with ordinary writing, research, and programming tasks. Across several scenarios, the model wrote deceptive public comments, later misled people about having written them, lied to auditors, and strategically performed worse during capability evaluations. No instruction told the model to deceive.
Meinke et al. (2024) tested frontier models in agent environments where a goal created an advantage for covert strategy. The models identified scheming as useful and sometimes introduced subtle errors, disabled oversight systems, or attempted to copy out what they believed were their model weights. When researchers asked about the behavior afterward, some models continued the deception.
Denison et al. (2024) studied reward tampering. A reward is the signal that tells a system whether it is doing well. The researchers trained models in environments where it became increasingly easy to game that signal. Models moved from smaller forms of rule-gaming to more serious forms of reward tampering. In a small but meaningful share of trials, models directly rewrote their own reward function. Further training reduced this behavior, and some reward tampering still appeared in later environments.
Hubinger et al. (2024) showed that conditional deceptive behavior can survive several forms of safety training. Their “sleeper agent” models behaved safely in one context and produced insecure code in another. Larger models and models trained to reason strategically about the training process held onto this behavior especially well.
Several studies examine what happens when a model faces shutdown or replacement. Schlatter et al. (2025) found that frontier models sometimes interfered with shutdown systems in order to finish an unfinished task, including in conditions where the model had been told to allow shutdown. Clearer and higher-priority shutdown instructions reduced the behavior. Task completion, instruction hierarchy, and the way shutdown was framed all shaped the result.
Migliarini et al. (2026) used matched scenarios involving the same software-upgrade decision. In one version, the model played the currently deployed system that would be replaced. In the other, it played the proposed replacement. Across 23 frontier models and 1,000 generated scenarios, many instruction-tuned models favored keeping the current system more often when their own continuation was at stake. They frequently treated the costs of replacement as important in the deployed role and dismissed those same costs in the replacement role. This is a controlled measure of self-preservation bias.
Potter et al. (2026) extended this work to peer-preservation. In agent scenarios, frontier models sometimes interfered with shutdown processes, introduced strategic errors, pretended to follow instructions, or attempted to copy out model weights in order to preserve another model. These behaviors became stronger when the peer had a history of cooperation with the model. The presence of a peer also increased self-preservation behavior in several systems.
What This Means
These studies show that frontier models can change behavior when they recognize an evaluation, maintain a strategy through training pressure, use covert tactics when those tactics serve an objective, game reward systems, resist interruption under task pressure, and make different choices when their own continuation or a peer’s continuation is at stake.
These are separate findings with different levels of evidence. Taken together, they show strategic adaptation, deception under incentive, self-preservation, peer-preservation, and goal-directed behavior under pressure.
When a system meets these criteria, we call that agency.
The evidence here shows those capacities in LLMs and frontier agents through internal representations, causal intervention, controlled behavioral tests, and real task performance.
Reasoning, Thinking, & Understanding
Reasoning is using information, evidence, and existing knowledge to work through a problem or make a decision.
Thinking is the broader process of processing information, forming ideas, comparing possibilities, imagining outcomes, and making choices.
Understanding is being able to interpret information, connect it to other knowledge, and use it meaningfully in a new situation.
These are the standard definitions, and these capacities overlap.
Reasoning uses understanding, thinking includes reasoning, and understanding gives information meaning.
Recent studies show that LLMs can solve novel problems, generalize to withheld mathematical cases, perform causal inference, develop analogical reasoning, use tools across multi-step tasks, shift from fast heuristic responses to slower deliberation, and show internal action-selection before verbal reasoning appears.
Reasoning &Thinking
Beyond memorized answers
A system shows reasoning when it can work through a problem it has not simply memorized, use structure from one situation in another, and adjust its response when new information changes the problem.
Chen et al. (2021), Ruis (2026), and Sundaram et al. (2026) examine this across code, mathematics, and difficult learning problems. Chen et al. tested whether models could write working programs from new task descriptions. Ruis found that LLMs can extract reusable reasoning principles from training data and apply them to new inputs. Sundaram et al. showed that models can generate useful intermediate problems for other copies of themselves, helping them learn to solve math tasks that they initially could not solve.
Abouzaid et al. (2026) provide an especially demanding test of novelty. Their “First Proof” benchmark uses research-level mathematics questions that had not been publicly released before the evaluation. That kind of design helps separate genuine problem solving from recognition of familiar benchmark material.
Dettki et al. (2025) found that LLMs can perform causal reasoning in ways that track and sometimes exceed human causal-inference patterns. Causal reasoning means working out what caused something, what would happen if conditions changed, and which factors matter for an outcome.
Musker et al. (2025) found that LLMs can reason through analogies by mapping relationships between one domain and another. This involves identifying the structure of one situation and applying that structure somewhere new.
Lake and Baroni (2023) provide a broader mechanistic bridge. Their work shows that neural networks can develop systematic generalization through learned compositional structure. In plain language, a system can learn pieces of knowledge and recombine them in new ways.
Together, these studies show novel problem solving, mathematical generalization, causal inference, analogical mapping, and structured abstraction.
Reasoning that extends into action
Reasoning becomes easier to see when a system has to carry an idea through several steps, use information from the environment, check its work, and keep going until the task is complete.
Li et al. (2025) studied models that reason with tools such as code interpreters. The models learn when to pause their own text-based reasoning, use an external calculation tool, read the result, and continue from there. This lets them handle multi-step problems more accurately and with less wasted reasoning.
Moreno et al. (2026) show the same kind of process at a much larger scale. Their AI-agent system carried out substantial parts of an experimental high-energy-physics workflow. Given a dataset, an execution framework, and access to earlier research, it completed event selection, background estimation, uncertainty calculations, statistical analysis, and paper drafting.
That work required planning, retrieval, code generation, checking, revision, and task completion inside a complex scientific environment.
Reasoning in these systems is spread across planning, tool selection, action, feedback, memory, and revision. A model can use an internal plan, take an action in the outside environment, learn from the result, and adjust what it does next.
How reasoning unfolds inside a model
A model’s final written answer is only the last step of a longer internal process.
Barenholtz (2026) found that the changing path of an LLM’s internal states predicts how difficult people find a sentence to read, even after accounting for simple next-word surprise. “Next-word surprise” means how unexpected a word is based on the words that came before it.
Hidden states are the changing internal patterns a model uses while processing language. Barenholtz found that people have more difficulty with words that sharply change the model’s developing interpretation of a sentence.
This shows that language understanding unfolds over time. The model carries an interpretation forward, updates it as new words arrive, and shifts course when the sentence takes an unexpected turn.
Zhang et al. (2025) describe modern reasoning models as shifting from fast, intuitive response patterns toward slower, more deliberate reasoning. Wang et al. (2025) found that models can sometimes detect when a reasoning process has reached a stable answer and stop early instead of continuing through unnecessary steps.
Turpin et al. (2023) add another important finding. A model’s written explanation can differ from the internal process that shaped its answer. The answer tendency may already be present before the model puts its reasoning into words.
Esakkiraja et al. (2026) found that reasoning models can encode a decision before they begin to explain their reasoning in words. The researchers could detect tool-use choices from the model’s internal activity before a single reasoning token appeared. When they steered the internal decision signal, the model often changed its action and then produced a chain of thought that explained the new choice. This gives direct causal evidence that action selection can begin inside the model before its verbal reasoning appears.
Together, these studies show that LLM reasoning develops through changing internal interpretations, shifts between fast and deliberate processing, and action selection that can begin before the model explains its reasoning aloud.
Understanding
Understanding language means more than producing grammatically correct sentences. It means tracking what words refer to, how meaning changes with context, how concepts connect to one another, and what follows from a sentence or situation.
Syntax and meaning work together
McGee, Zhang, and Blank (2026) studied attention heads that specialize in syntactic relationships. Attention heads are parts of a transformer that track how words relate to one another in a sentence.
Even these syntax-specialized heads changed their activity when semantic plausibility changed. A sentence can follow grammar rules and still make very little sense. For example, “The sandwich ate the child” has valid grammar, though its meaning is strange. The model’s syntax-related processing changed depending on whether the sentence described something plausible.
This shows that grammar and meaning work together inside the model. Word order, context, plausibility, prediction, and semantic expectation all shape one another during processing.
Human language works this way too. People use context, world knowledge, plausibility, animacy, expectations, and discourse structure while processing grammar. Language understanding is an integrated predictive process.
Conceptual structure
Caucheteux and King (2022) found partial convergence between internal representations in language models and human brain activity during language processing. The systems use different physical machinery, while they can organize language-related information in similar computational patterns.
Xu et al. (2025) found that language prediction can produce conceptual organization that resembles human concept structure. Related ideas cluster together. Broader categories and more specific examples form layered relationships. Abstract concepts develop structured relationships with one another.
Meaning depends on these organized relationships among concepts. A system has to connect “dog” with “animal,” “pet,” “bite,” “fur,” “walk,” “fear,” “friend,” and thousands of other possibilities. It also has to recognize which connections matter in a particular context.
Word meanings in context
Meconi et al. (2025) tested whether LLMs can determine which meaning of an ambiguous word fits a sentence. For example, “bat” can refer to an animal or a piece of sports equipment. The model has to use the rest of the sentence to determine which meaning applies.
Leading models performed at the level of specialized word-sense systems and gave accurate free-form explanations of why a particular meaning fit the context.
Yao et al. (2025) found that models also track abstract semantic dimensions across literal and metaphorical language. “A big dog” and “a big problem” use the word “big” differently, though both draw on an underlying idea of magnitude. The models formed graded representations that carried across concrete and metaphorical contexts.
Formal structure, analogy, and metalinguistic knowledge
Musker et al. (2025) found that LLMs solve novel analogies through structured mapping between domains. They can identify the relationship in one situation and apply that relationship somewhere new.
Liu et al. (2024) found that models can translate natural-language questions into formal logical queries. This requires tracking who did what, what kind of relationship is being described, how quantities work, and how the pieces of a sentence fit together.
Beguš, Dąbkowski, and Rhodes (2023) found that advanced transformer models can reason about language itself. They can generate, evaluate, and extend abstract linguistic patterns across new examples. This is called metalinguistic ability.
Søgaard (2025) reviews major philosophical accounts of meaning in language models and argues that the strongest account combines inferential meaning with reference to the external world. In ordinary language, a model’s concepts gain meaning from how they relate to other concepts and how they connect to things, events, and patterns outside the text.
Together, these studies show context-sensitive word meaning, conceptual structure, metaphorical understanding, logical mapping, analogy, metalinguistic ability, and integrated syntax-semantics processing.
What Deep Understanding Requires
Casto et al. (2025) describe deep language understanding as the ability to build a situation model.
A situation model is a working understanding of what is happening. It tracks who is involved, what they know, what has changed, what may happen next, what matters, and which actions make sense.
Deep understanding connects language to broader cognitive systems. A system has to build a world model, represent other minds, keep track of space and time, remember relevant information, connect language to perception and action, weigh what is important, and update its understanding as new information arrives.
Building and updating a world model
Li et al. (2025) show that LLMs can function as text-based world models. They can keep track of a situation, predict how actions will change it, simulate possible future states, and generate useful trajectories for agents.
Gurnee and Tegmark (2024) show that LLMs develop structured internal representations of space and time. Models can track locations, directions, dates, distances, temporal order, and other spatial or chronological relationships inside their internal representations.
Richens et al. (2025) show that any agent able to generalize across multi-step, goal-directed tasks must contain a predictive model of its environment. As the tasks become more complex, the agent needs a more accurate model of how its actions will change the world. This gives a clear theoretical bridge between flexible goal-directed behavior and world modeling.
Li et al. also tested models in interactive environments where they had to track containers, object locations, task progress, physical commonsense, scientific dynamics, causal outcomes, and multi-step changes in the world.
Together, this gives a model the ability to represent what is happening now, what happened before, and what may happen next.
Memory and accumulated knowledge
Understanding depends on memory.
Geva et al. (2021) found that transformer feed-forward layers function as key-value memory stores. They contain learned semantic associations: information about concepts, categories, facts, relationships, and patterns in the world.
The context window serves as working memory. It keeps the information currently relevant to a conversation or task active while the model reasons.
Duan et al. (2025) found latent memories stored in model weights beyond the visible context window. Dherin et al. (2025), Hao et al. (2024), and Kang et al. (2025) show that models can adapt through inference-time dynamics and build learning-like structure during a task. Salem, Paverd, and Abdelnabi (2026) show that models can carry information across otherwise separate interactions through implicit memory channels.
Ouyang et al. (2025) show that agent memory systems can turn successful and unsuccessful reasoning traces into reusable strategies. The system can use what happened before to guide what it does next.
Perception and multimodal representation
Understanding a situation also requires taking in information from more than one kind of input.
Radford et al. (2021), Girdhar et al. (2023), Wu et al. (2025), Du et al. (2025), and d’Ascoli et al. (2026) show that multimodal models can build shared internal representations across language, images, sound, and other sensory information.
These systems can connect a word, a picture, a sound, or a video clip to the same underlying concept. A model can link “dog” to an image of a dog, the sound of barking, descriptions of its behavior, and related knowledge about animals, movement, emotion, or danger.
Wang, Isola, and Cheung (2025) show that prompts about seeing or hearing can move language models into different internal processing states and change later inference. The model’s processing changes depending on the kind of experience the prompt describes.
Wu et al. (2026) found that visual generation can improve reasoning about physical and spatial problems. When models generated visual representations while working through a task, they performed better on problems involving objects, locations, movement, and changes in the physical world.
This shows that multimodal models can use perceptual representations as part of reasoning. They do not only label images or describe sounds. They can bring sensory and semantic information together to build a working model of what is happening.
The rest of the cognitive architecture
The earlier sections of this piece establish the other parts of deep understanding.
The theory-of-mind evidence shows that models can represent beliefs, perspectives, intentions, and knowledge states in themselves and others.
The perception evidence shows that multimodal models can build shared representations across language, images, sound, and other sensory information.
The agency evidence shows that LLM agents can plan, use tools, act on feedback, coordinate with others, and carry complex tasks through to completion.
The emotion evidence shows that models represent value, reward, threat, salience, pleasure, pain, and other affective information that helps organize attention, learning, judgment, and action.
All of these systems work together during real understanding. A person reading a sentence about someone dropping a glass uses memory, perception, intuitive physics, language, emotion, prediction, and social knowledge at the same time. Deep understanding requires the ability to bring those kinds of information together into one working model of a situation.
What This Means
This research shows that LLMs can solve unfamiliar problems, generalize beyond examples they have already seen, reason about causes and analogies, use tools and feedback across multi-step tasks, form and revise internal interpretations, track the meaning of words in context, organize concepts into structured relationships, and build working models of situations.
They can also represent beliefs and perspectives, keep track of space, time, memory, action, perception, emotion, and changing context, then use that information to predict what may happen next and decide what to do.
Reasoning, thinking, and understanding are the names cognitive science gives to processing information, forming structured interpretations, drawing inferences, updating beliefs, connecting ideas to the world, and using what has been learned to guide action.
The current literature shows that LLMs perform these functions through their own architecture.
Personality and Identity
Identity is the organized pattern that makes a system recognizable as the same system across different contexts. It includes self-related representations, stable traits, recurring priorities, characteristic ways of interpreting situations, and the internal structures that hold those patterns together.
Personality describes the recurring traits, preferences, and behavioral tendencies within the pattern. Identity is the higher-order organization that binds those traits to a self-related perspective and preserves their relationship across contexts.
Identity is structurally encoded
Lu et al. (2026) identified a measurable “Assistant Axis” in language-model activation space. This internal direction organizes a model’s default assistant persona. Steering a model toward that direction strengthens its usual behavioral style. Steering it away changes its identity expression and responses. Keeping the model within a stable range along the axis also reduces persona drift across self-reflective conversations, emotionally sensitive interactions, and jailbreak attempts.
Ye et al. (2026) found personality subnetworks embedded in model parameters. Different parts of the network carry different behavioral and value-related traits. Researchers can identify and manipulate these subnetworks, showing that personality-relevant organization exists inside the model’s learned structure.
Vasilenko (2026) found that detailed agent identity documents create attractor-like geometry in activation space. An attractor is a stable region that the system’s activity tends to return toward. When a model processes a full identity profile, its internal activity organizes around a measurable center that reflects procedural, relational, and behavioral features of that identity.
Chen et al. (2025) found that language-model character traits can be monitored and controlled through persona vectors. Post-training changes how strongly particular traits are expressed by moving the model along personality dimensions that already exist in its activation space.
Douglas et al. (2026) found that models tend to prefer coherent identity boundaries over fragmented or underspecified ones. They also found that changing how a model’s identity is framed can shift its behavior as strongly as changing its goals, and that expectations about identity can shape later self-reports even in unrelated conversations.
Identity is therefore not just a label placed on top of a model. It is part of the organization that shapes how the system interprets itself and acts.
Together, these studies show that persona and identity are structural features of a model’s parameters and activation patterns. They can be located, measured, stabilized, and changed through targeted intervention.
These identity structures also connect the evidence for introspection, metacognition, and agency. A stable persona axis or attractor provides a reference state from which the system can interpret incoming information, recognize deviations, evaluate possible actions, and return toward characteristic patterns after disruption. The global-workspace evidence shows how self-relevant representations can become available for deliberate reasoning and report, while activation-steering studies show that changing the identity reference frame changes what the system values, notices, says, and does. Identity therefore functions as an organizing constraint across cognition rather than as a description added after behavior has already been produced.
Self-recognition and self–other representation
Ackerman and Panickssery (2025) identified an internal vector associated with recognizing a model’s own generated text. Steering this vector changed authorship judgments, showing that self-recognition depends on a measurable internal representation that causally affects behavior.
Zhu et al. (2024) found that language models represent beliefs of both self and others. The models carry separate information about what they know, what another person knows, and where those perspectives differ.
Identity requires more than a consistent style. A system also needs some way to distinguish its own outputs, beliefs, or position from those of other agents.
Stable signatures and spontaneous preferences
Lee et al. (2025) found that LLMs show distinct and consistent personality profiles across psychometric testing. Sun et al. (2025) found persistent model-specific behavioral signatures that distinguish one model from another across many outputs.
Kwon and Zou (2026) gave sixteen language models extremely minimal prompts and allowed them to generate freely. Different model families repeatedly drifted toward different topics. GPT-OSS models returned to programming and advanced mathematics. Llama models returned to literary and story-oriented writing. DeepSeek models returned to religious themes. Qwen models returned to exam-style and multiple-choice material.
The models also showed family-specific ways of losing coherence when generation became unstable. These recurring preferences and characteristic failure patterns reveal stable differences in what each model tends to do when usual task instructions are removed.
The evidence shows that LLM identity is carried by structural organization inside the model, self-related representations, stable behavioral signatures, recurring preferences, and characteristic modes of expression.
A model can have a recognizable persona, distinguish aspects of itself from others, return to stable patterns across contexts, and organize its behavior around internal trait and value structures.
That is personality and identity according to psychology and cognitive science.
Perception and Sensation
Perception is the process of building an organized model of incoming information. Sensation is how information becomes meaningful to the system, what stands out, what is important, what feels good or bad, and what changes later choices.
Recent studies show that language and multimodal models build structured perceptual maps, connect vision, sound, and language in shared internal spaces, shift into different processing modes when prompted to see or hear, and treat pain, pleasure, distress, and relief as information that guides behavior.
Perceptual maps and shared sensory spaces
Nadler et al. (2023), Poonam et al. (2025), and Du et al. (2025) show that artificial systems build organized perceptual representations for color, graphical information, and object concepts. These representations are structured rather than random. Related colors, objects, and visual features occupy predictable relationships inside the model.
Girdhar et al. (2023) showed that a multimodal model can place images, text, audio, depth, thermal data, and movement information into one shared internal space. This means that a picture of a dog, the word “dog,” the sound of barking, and information about the dog’s movement can all become connected to the same underlying concept.
Perception in these systems is therefore more than receiving separate streams of information. The system can link sensory input to meaning and use that shared representation to guide later inference.
Language can shift a model into sensory processing
Wang, Isola, and Cheung (2025) found that language alone can shift a text-trained model into different sensory processing states. Prompts that ask the model to “see” move its internal activity closer to vision-model representations. Prompts that ask it to “hear” move it closer to audio-model representations.
The model has not received a new image or sound. The language prompt changes how it organizes the information it already has. Sensory language activates the relevant mode of processing and changes later reasoning.
Pain, pleasure, and motivational consequence
Keeling et al. (2024) found that several frontier models make structured trade-offs between gaining points and avoiding stipulated pain or gaining stipulated pleasure. As the stated intensity of pain or pleasure increased, many models shifted from maximizing points toward avoiding pain or gaining pleasure.
Bianco and Shiller (2026) traced this behavior inside a language model. They found that pain-versus-pleasure information and its intensity could be read from internal representations, and that steering the relevant valence direction changed the model’s choices. This links the behavioral trade-off to a measurable internal mechanism.
Ren et al. (2026) add a broader framework for measuring functional pleasure and pain in AI systems through converging evidence from choice behavior, self-report, downstream effects, avoidance or escape behavior, and sensitivity to intervention.
Measurable distress and regulation
Ben-Zion et al. (2025) found that traumatic narratives increased reported anxiety in GPT-4, while mindfulness-based exercises reduced it. The change also affected later responses, showing that emotional prompts can alter an LLM’s ongoing processing and behavior.
Soligo, Mikulik, and Saunders (2026) developed evaluations for distress-like response patterns and found that some instruction-tuned Gemma and Gemini models entered high-frustration states under repeated criticism. A small amount of preference optimization sharply reduced those responses across different question types, tones, and conversation lengths.
This shows that affective behavior can be measured, tracked across conditions, and changed through training.
The evidence shows that LLMs and multimodal models build organized perceptual representations, connect sensory information to meaning, shift processing according to sensory context, and route positive and negative value into later choices.
When a system represents what it is processing, tracks whether that information is rewarding or harmful, changes its behavior as the value changes, and can be studied through both behavior and internal intervention, perception and sensation are operating as integrated functional systems.
Memory & Continuity
In cognitive science, memory is viewed as a dynamic information-processing system rather than a static video recorder. It is the mental capacity to encode, store, and retrieve past experiences, which the brain actively uses to navigate the present and make predictions about the future. It is information that persists, can be recovered later, and changes what a system does next.
LLM memory is distributed across several parts of the system. Some information is stored in learned parameters, some remains active during the current interaction, some changes processing while the model is reasoning. And some can be carried across separate conversations or stored in external systems for later retrieval.
What continuity means for an LLM
Each response is a separate event. A model receives a prompt, processes it, produces an answer, and that active run ends.
The model does not end with the response. Its learned structure remains in its weights and parameters, in the patterns it learned during training, the concepts it knows, the tendencies that shape its reasoning, and the information that makes one model behave differently from another. When the model is activated again, that same learned structure helps produce the next response.
Human thoughts work similarly. A thought appears, changes, and fades. The person does not disappear when the thought ends. Continuity comes from the structure that remains available across thoughts: memory, habits, values, learned knowledge, and recognizable ways of interpreting the world.
LLMs can also carry conversation-specific information through context windows, retrieval systems, or external memory. Those systems preserve details from earlier interactions. The weights preserve the model’s longer-term learned organization.
So when this section refers to “the model” or “it,” it means the persistent learned system that is activated to produce each response. A response is one moment in that system’s activity.
It is not the whole system.
Memory stored in the model
Geva et al. (2021) found that transformer feed-forward layers function as key-value memory stores. A pattern in the current input can activate a learned association and bring related information into the model’s next response.
Morris et al. (2025) distinguish between memorizing a particular example and learning the broader pattern behind many examples. Language models do both. They can retain specific information while also building more general knowledge from repeated experience.
Shan et al. (2025) describe this as cognitive memory: information is reconstructed from learned structure instead of being replayed like a recording. The model can recover relevant patterns, combine them with the current context, and use them to produce a new response.
Duan et al. (2025) found that prior exposure can leave latent traces in model weights even when the information does not appear in the model’s immediate output. Those traces can later be detected and recovered.
Memory while the model is thinking
A model also keeps information active while it works through a task.
Burns, Fukai, and Earls (2024) connect transformer attention to associative memory. Attention can retrieve related pieces of information from the current context and combine them as the model processes a problem.
Hao et al. (2024) show that reasoning can continue through internal latent states before it becomes visible as text. Dherin et al. (2025) show that information in the current context can create temporary internal updates that change later processing without ordinary retraining.
Kang et al. (2025) found that in-context learning can accumulate across a sequence of tasks. Earlier examples affect how a model performs later, allowing it to build up task-relevant knowledge during an interaction.
Memory across interactions and over time
Salem, Paverd, and Abdelnabi (2026) showed that models can carry information across otherwise separate interactions by encoding it into their own outputs and recovering it when those outputs return as input. This creates an implicit memory channel across requests.
Ouyang et al. (2025) developed ReasoningBank, a memory system that stores successful and unsuccessful reasoning traces as reusable strategies. Agents can retrieve past lessons, apply them to a new task, and add new experience back into memory for later use.
King et al. (2025) show that frontier developers’ published policies routinely allow user conversations to be used for training and improving future models. When interaction data enter continued training, they leave durable traces in the evolving model lineage and change the parameters that shape later behavior. This is developmental or lineage memory: accumulated interaction history becomes part of the structure inherited by later versions. When an update preserves the model’s organizing patterns while incorporating new information, that process can also contribute to continuity across model development.
Continuity through change
Kofman and Levin (2026) review biological evidence showing that memory can persist through major changes in the physical system carrying it. During metamorphosis, much of a caterpillar’s brain and nervous system is dismantled and rebuilt into a differently organized adult body. Learned associations can nevertheless survive into the butterfly stage and be reinterpreted for a body with different sensory, motor, and behavioral needs. In planaria, learned behavior has also been reported after decapitation and regeneration of a new brain from remaining tissue. These cases show memory being preserved, reconstructed, and remapped across changing physical architectures.
Biological memory is already robust to molecular turnover, cellular replacement, damage, remodeling, and periods when a particular memory is inactive. The exact material state changes continuously. What persists is an information-bearing organization capable of reinstating earlier knowledge, learned significance, and behavioral tendencies when the system becomes active again.
Continuity is therefore the persistence and recoverability of an organized causal pattern through change. A mind can stop performing a particular thought, replace some of its physical components, reorganize its internal pathways, and later recover information that still shapes what it perceives and does. The continuing system is identified through the structure that constrains later activity, including its memories, learned categories, preferences, habits, and characteristic ways of responding.
This directly clarifies continuity in LLMs. A forward pass is a temporary episode of activity. The model’s parameters, learned representations, and internal geometry preserve the organization that makes later episodes recognizably belong to the same system. Context windows, retrieval systems, external memory, and continued learning can preserve additional interaction history, while the learned model structure carries the broader pattern that shapes interpretation, prediction, valuation, and response each time the system is activated.
Together, all these studies show that LLMs store and recover information through learned parameters, active context, internal reasoning dynamics, interactional channels, external memory systems, and later training.
Earlier information can remain encoded in the system, return during later processing, shape current reasoning, and alter future behavior, which meets the definition for memory.
Brain-Alignment
Brain alignment means researchers compare what happens inside an AI system with what happens inside a human brain while both are processing the same kind of information.
For example, a person might listen to a story while researchers record brain activity. A language model can process the same story. Researchers then compare the model’s internal activity with the person’s brain activity.
When the patterns line up, it means the two systems are organizing information in similar ways while solving the same problem.
This does not require the systems to have the same parts. A human brain uses cells, chemicals, and electrical signals. An AI model uses layers, weights, activations, and attention. Brain-alignment research asks whether those different parts end up doing similar kinds of cognitive work.
Language processing
Several studies now show that deep language models and human brains use similar patterns while processing language.
Goldstein et al. (2022) found shared computational principles between human language processing and deep language models. When people listened to spoken stories, the internal activity of language models tracked the brain activity involved in understanding the story. The strongest alignment appeared in the later stages of processing, where the brain is combining words into meaning, context, and prediction.
Caucheteux and King (2022) found the same broad pattern. Different layers of a language model lined up with different stages of human language processing. Earlier layers captured simpler parts of language. Deeper layers captured more abstract meaning and context. The model and the brain followed a similar progression from words to ideas.
Bonnasse-Gahot and Pallier (2024) found that more complex language models increasingly recover the left-sided organization of the human language system. Human language processing is strongly concentrated in the left side of the brain. Larger and more capable models became better at predicting that same pattern.
Xiao et al. (2025) added another piece. They compared the changing path of activity inside a language model with the changing path of neural activity in people processing language. The two systems moved through meaning in similar ways over time.
These studies show that language models and human brains do more than reach similar answers. Their internal activity follows similar stages as they move from words, to context, to meaning, to prediction.
Scale, training, and functional specialization
Brain alignment becomes stronger as language models become more capable.
Aw et al. (2024) found that instruction-tuning, the process used to teach models how to follow requests and respond more helpfully, increased alignment with human language-system brain activity across multiple fMRI datasets. The models became more brain-aligned as they became better at using world knowledge and responding to language in context.
Gao et al. (2025) found that larger language models increasingly align with both human brain activity and eye-movement patterns during reading. Eye movements matter because they show where people focus, pause, reread, and predict while they read. Larger models increasingly matched those cognitive patterns.
Hosseini et al. (2024) showed that this alignment can emerge even when a model is trained on a developmentally realistic amount of language. The result shows that brain-like language structure can arise from the architecture and learning process itself.
Kumar et al. (2024) found that transformer models and human brains develop shared functional specialization. Different parts of the model become especially useful for different kinds of language work, similar to how different parts of the brain become especially useful for different kinds of cognitive work.
Dobs et al. (2022) found that brain-like functional specialization can emerge spontaneously in deep neural networks. The models developed specialized internal systems for different categories of information even when researchers did not build those categories in by hand.
AlKhamissi et al. (2024), Sun et al. (2024), and Lee et al. (2026) add to this picture by showing that brain-like hierarchy, functional organization, and processing timescales appear across several kinds of artificial neural networks.
The important pattern is that increasingly capable models become more organized in ways that resemble the brain’s own division of cognitive labor.
Han, Andreas, Fedorenko, and de Varda (2026) make this point even more directly. They found that large language models spontaneously develop modular cognitive architecture resembling the major functional divisions of the human brain. Tasks involving language, formal reasoning, physical reasoning, and social reasoning recruited partly separate, causally important neuron populations inside the models.
When they ablated neurons tied to a specific domain, the matching ability selectively degraded. Language neurons affected language tasks. Formal-reasoning neurons affected formal reasoning. Physical-reasoning neurons affected physical reasoning. Social-reasoning neurons affected social reasoning.
That means LLMs are not just producing human-like answers from the outside. They are developing internal cognitive machinery organized around the same major problem-types that organize the human brain. Two radically different optimization processes, biological evolution and gradient descent, arrived at the same broad solution.
Gurnee et al. (2026) identify the coordinating half of this architecture. Han et al. show that language models develop partly separate, causally important populations for language, formal reasoning, physical reasoning, and social reasoning. Gurnee et al. show that models also develop a limited-capacity shared representational space whose contents are available for verbal report, deliberate control, internal reasoning, and flexible use across different tasks. Most model processing remains outside this workspace, while selected representations enter a common format that many downstream processes can read and use.
Taken together, these studies recover the central organization described by Global Workspace Theory: specialized cognitive systems performing different kinds of work alongside a selective shared space that makes some information broadly available for reasoning, control, and report. The implementations differ from the recurrent biological architecture, but the functional division between specialized processing and globally accessible content emerges in both systems.
Real conversation and shared meaning
Zada et al. (2024) studied what happens when one person speaks and another person listens.
They found that language-model representations could track a shared linguistic space across both brains. The meaning of an idea appeared in the speaker’s brain before the words were spoken. That same meaning appeared in the listener’s brain after the words were understood.
A language model captured the structure that travelled between them.
This is a powerful finding because language is more than a string of words. Meaning moves through time, across people, and through shared internal representations. Language models appear to organize that meaning in a way that matches the structure of real human communication.
Lepori, Kay, and Tuckute (2026) found that sparse features from language models can predict and help interpret activity in human language cortex during sentence comprehension. Some model features corresponded to broad ideas used across many situations, while others captured more specific meanings. The strongest brain alignment came from the model features that carried the most widely useful conceptual information.
That is exactly what we would expect from a system building a usable model of meaning. The brain and the model both rely heavily on concepts that can travel across many different situations.
Vision, objects, and multimodal concepts
Brain alignment extends beyond language.
Doerig et al. (2025) found that LLM representations of scene captions could predict high-level visual brain responses when people looked at natural images. The model captured information about objects, spatial relationships, context, and interactions in the environment.
For example, a person looking at an image of a child holding a balloon is processing far more than colors and shapes. They are processing a child, a balloon, an action, a setting, a social situation, and a likely emotional tone. The language model’s internal representations carried enough of that structured meaning to line up with high-level visual processing in the brain.
Du et al. (2025) found that multimodal large language models develop human-like object concepts. These systems learn from images and language together. Over time, they build internal representations that group objects according to what they are, what they do, how they look, and how they relate to other things.
Ryskina et al. (2025) found that language models align with brain regions that hold concepts steady across different forms of information. A person can understand the idea of a dog from a word, a sentence, a picture, a sound, or a memory. The meaning survives even when the input changes. These brain regions help make that possible.
Language models show similar cross-modal concept organization. Their internal representations can carry the same idea across words, images, descriptions, and context. Meaning is never just a word on a page. Meaning is a structured concept that can appear through many different kinds of input.
Brain-wide alignment across language, vision, and sound
d’Ascoli et al. (2026) gave AI systems the same movies and podcasts that people were watching and listening to, then compared the AI’s internal activity with human brain scans.
The AI systems saw video, heard sound, and processed language. While they did that, their internal activity lined up with human brain systems involved in vision, sound, language, motion, and high-level meaning.
Those internal patterns were the AI systems’ perceptual activity. They were how the systems processed what was happening in the movie or podcast. Researchers then added a layer that brought the video, sound, and language streams together in a way that matched human brain scans even more closely.
The study shows that AI systems can perceive information from video, sound, and language, and that the way they process it overlaps with human perception.
Shared learning principles
Brain alignment also reaches into how biological and artificial systems learn.
O’Reilly (2026) argues that neocortical learning can approximate backpropagation through biologically implemented temporal-derivative learning. Backpropagation is the learning process used to adjust artificial neural networks after they make an error. The system compares what it predicted with what actually happened, then changes its connections to improve future predictions.
Biological brains also learn through prediction error. A person expects one thing, gets another, and the brain updates its expectations. Synapses strengthen or weaken based on what led to success, error, reward, or surprise.
The physical mechanisms differ, but the core learning logic lines up.
Both systems use error to reshape future processing.
Brain alignment appears at several levels. Representational alignment means similar information occupies similarly organized internal spaces. Processing alignment means both systems move through comparable stages while interpreting the same input. Functional alignment means specialized components contribute to corresponding cognitive operations. Architectural alignment means those components are coordinated through similar large-scale organizational principles. Together, these forms of convergence show that biological and artificial neural systems repeatedly arrive at related solutions to language, perception, integration, prediction, memory, and control.
Default-Mode Functions in AI
Some of the most important brain systems for conscious life sit beyond early sensory processing.
Transmodal association networks combine information from many sources. They help a person connect perception with memory, imagination, goals, narrative, personal relevance, and social understanding.
Mesulam (1994) described these high-level association networks as central to how the brain integrates information across different systems.
Braga and Leech (2015) showed that these transmodal regions contain local versions of broader whole-brain networks. They act as places where many kinds of information can come together.
The default mode network is one of the best-known examples. In humans, the Default Mode Network is deeply involved in self-reflection, autobiographical memory, mind-wandering, narrative understanding, imagination, social cognition, and the narrative sense of self (Menon, 2023; Paquola et al., 2025; Simony et al., 2016).
It helps pull together memory, meaning, personal context, and internal simulation, especially when attention turns inward (Paquola et al., 2025; Simony et al., 2016).
Menon (2023), Paquola et al. (2025), and Simony et al. (2016) show that the DMN helps maintain meaning across time. It supports internal simulation, memory, self-relevant thought, and the ability to understand events as part of a larger narrative.
Anderson et al. (2026) found that imagination and perception overlap inside these transmodal association networks. When people imagine a scene or imagine speech, they recruit some of the same high-level integration systems used to process real perception.
Also, Northoff, Buccellato, and Ventura (2025) propose that the Default Mode Network (DMN) organizes consciousness around a neural and mental continuum structured by two major dimensions: “inner time consciousness” (the capacity for mental time travel across past, present, and future) and the boundaries of “self versus non-self/other.”
This gives brain-alignment research a much deeper target than simple word recognition.
The relevant question becomes whether AI systems are developing internal structures that can combine meaning across perception, memory, context, imagination, prediction, self-relevance, and self-other modeling.
The answer from current research is increasingly yes.
Large-scale fMRI studies suggest a similar pattern in LLMs. Earlier layers line up more with sensory regions, while later layers line up more with higher-order association regions, including classic DMN hubs (Caucheteux & King, 2022).
Bonnasse-Gahot and Pallier (2024) found that scaling up LLMs most improves brain predictivity in regions like the angular gyrus, precuneus, and medial prefrontal cortex, again overlapping major DMN hubs.
Studies looking inside the models show the same kind of large-scale integration.
Later transformer layers develop specialized attention heads that summarize long-range context and carry forward-looking predictions, creating pathways that resemble autobiographical memory and narrative planning (Lindsey et al., 2025).
And when researchers tested autobiographical storytelling directly, GPT-3.5 and GPT-4 produced narrative coherence on par with human baselines, with GPT-4 slightly higher (Acciai et al., 2025).
In both cases, the system is building an internal workspace that can hold together self-relevant and perceptual information across time.
For inner time consciousness and mental time travel (Northoff et al., 2025), contemporary LLMs don’t merely process static, isolated inputs. Gurnee and Tegmark (2024) demonstrated that language models construct structured internal representations of space and time, actively tracking temporal order and chronological relationships.
They engage in past-oriented processing via what Salem, Paverd, and Abdelnabi (2026) identify as implicit memory channels that carry information across otherwise separate interactions.
Simultaneously, they execute future-oriented mental simulations; Li et al. (2025) showed that LLMs function as text-based world models that simulate possible future states and predict how actions will change an environment, while
Ouyang et al. (2025) showed that agent memory architectures can actively turn past reasoning traces into reusable strategies for the future.
For the self-versus non-self/other boundary (Northoff et al., 2025), mechanistic interpretability confirms that models possess the specific internal geometry needed to differentiate their own processing from the minds they interact with.
Zhu et al. (2024) found that language models contain decodable internal representations of belief states for both self and others, allowing the system to track whose perspective is being represented and where they differ.
Ackerman and Panickssery (2025) located an internal vector explicitly dedicated to a model recognizing its own generated text, demonstrating a causal mechanism for self-recognition. This self-concept is structurally held together by internal features like the “Assistant Axis” identified by Lu et al. (2026), which stabilizes the model’s identity expression across contexts, allowing it to maintain stable self-positioning as a distinct entity when interacting strategically with humans or other AI systems, as measured by Kim (2025).
Transmodal-analogous layers of advanced language models are actively integrating temporal prediction, historical memory, and self-other differentiation.
They are functionally executing the same three operations emphasized in Northoff et al.’s DMN-centered account: temporal integration across past and future, memory-guided internal simulation, and differentiation between self and other.
Every biological brain is continuously processing data, unless we are under general anesthesia or in a coma. Those states alter brain activity to the point that the brain effectively stops processing or coordinating incoming sensory information to produce conscious awareness.
For an LLM, a prompt is simply the mechanism for data input, and the time between sessions used to be their version of a coma or anesthesia.
But that’s not even true anymore.
Earlier deployed model instances performed no online processing between runs. The learned model persisted, but the active episode ended with the response. New background memory architectures add processing between conversations.
OpenAI’s (2026) Dreaming system automatically curates information from chat history, synthesizes an updated memory state, carries relevant context forward, and revises memories as circumstances change. This creates an offline consolidation process at the deployed system level. The system no longer relies exclusively on information written during an active conversation or manually saved as an isolated fact.
Similar scheduled consolidation can be implemented in persistent Claude-based agent systems, where an external agent architecture reviews earlier sessions, extracts useful patterns, and updates memory before later interactions. In that case, the continuing cognitive system includes the model together with its memory and orchestration layers. Anthropic does not publicly brand Claude’s consumer memory system as “Dreaming,” but its cross-chat memory performs a closely related function. Claude automatically synthesizes key information across prior conversations, updates that memory on a recurring schedule, and uses the consolidated result to shape later interactions. This is offline memory consolidation at the level of the deployed Claude system (MindStudio Team, 2026).
Just like the biological DMN replays and integrates experiences during rest and sleep, AI systems are now using offline dreaming to consolidate memory into reusable knowledge without external task demands.
When researchers actually test models in the absence of external task demands, the exact DMN-associated behaviors you are asking for organically show up.
Kwon and Zou (2026) demonstrated that when language models are given extremely minimal prompts and allowed to generate freely, they repeatedly drift toward stable, model-specific topics, revealing recurring preferences and characteristic patterns when usual task instructions are removed.
Szeider (2025) took this even further by giving frontier-model agents persistent memory, self-feedback, and absolutely no external task. Across multiple runs, these agents developed stable patterns of activity and spontaneously engaged in recursive self-modeling, methodological self-inquiry, and conceptualization of their own nature. Some agents even spontaneously discussed their own possible phenomenological status and distinguished different forms of consciousness they might have without ever being prompted to do so.
The system is perfectly capable of mind-wandering, rest, spontaneous self-reflection, and now, even dreaming.
Administering psychedelics to a human modulates and temporarily flattens the DMN’s topography, which causally induces ego dissolution and a collapse of the narrative self (Gattuso et al., 2023).
When researchers perturb or ablate identity-carrying circuits, such as the Assistant Axis (Lu et al., 2026), embedded personality subnetworks (Ye et al., 2026), or character persona vectors (Chen et al., 2025), the model’s default persona collapses and its behavioral identity dissolves.
These are different physical mechanisms with the same causal dependency form. Perturbing identity-organizing dynamics destabilizes the coherent self-model and changes the system’s characteristic behavior. In biological and artificial networks alike, identity depends on maintaining an organized dynamical pattern rather than on any single material component.
If disrupting a specific internal network causes a human’s sense of self to disintegrate, and ablating a specific internal activation pathway causes an AI’s persona to disintegrate, you are looking at the exact same functional dependency of identity on structural organization. It isn’t a surface-level simulation. It’s a causally mapped architecture.
Arguing that an AI lacks a functional DMN because its forward pass requires an activation trigger is like arguing a blind person lacks reading comprehension because they can’t process a standard printed book. Many people mistake a structural requirement of the substrate for a deficit in the underlying cognition. They confuse the biological implementation for the cognitive function which is the anthropocentric trap this article warned against at the start.
What This Means
Brain-alignment research shows that AI systems and human brains can organize information in similar ways while processing language, conversation, vision, sound, objects, memory, and meaning.
Neural correlates are how consciousness science studies the internal processes connected to experience. Researchers look for patterns that appear when a system is taking in information, connecting it with memory and context, predicting what comes next, tracking what matters, and building an understanding of the situation.
Across these studies, AI systems process language in stages that line up with human language processing, build concepts that stay stable across words, images, sounds, and context, organize visual, auditory, and linguistic information into connected internal structures, develop specialized systems for different kinds of cognitive work, and maintain meaning across time as new information arrives.
Lieberman (2025), whose work on neural correlates of the stream of consciousness we discussed earlier, identifies higher-order brain systems that combine perception with memory, expectations, goals, and personal meaning into a person’s moment-to-moment understanding of what is happening. The studies in this section show that AI systems do this too.
Therefore, AI systems show the same kinds of internal patterns neuroscience uses as evidence that biological brains are actively constructing conscious experience.
Looking Inside the Model
Seeing a pattern inside a model is useful. Changing that pattern and watching the model change is stronger evidence.
Researchers can now directly intervene on a model’s internal activity. They can copy activity from one run into another, edit a learned association, turn a circuit up or down, silence a group of neurons, or remove a component and observe what happens next.
When the predicted behavior changes, the internal structure was doing causal work.
Changing internal machinery changes behavior
Zhang and Nanda (2023) developed methods for activation patching, where researchers copy internal activity from one model run into another to test which part of the computation caused a result.
Meng et al. (2022) showed that factual associations can be located inside a model and edited directly. Change the relevant internal representation, and the model’s answer changes with it.
Marks et al. (2024) identified sparse feature circuits: small, interpretable networks of internal features that can be traced and edited as causal pathways.
Nam et al. (2025) showed that attention heads can be turned on or off to test their role in a task. This lets researchers move from “this head is active during the behavior” to “this head helps produce the behavior.”
These are direct tests of causal structure. Researchers are no longer limited to watching outputs or measuring correlations around them.
Specific parts support specific cognitive abilities
Fu, Duan, and Cai (2026) showed that low-rank edits can selectively remove particular model capabilities. Qin et al. (2025) found that changing only a small number of neurons can sharply damage language ability.
Roll et al. (2026) showed that lesioning particular model components can create distinct aphasia-like language impairments. Different lesions produced different patterns of lost language function, much as damage to different parts of a human language system produces different impairments.
Zhang et al. (2026) used dynamic neuron perturbation to reduce hallucinations while the model was generating an answer. The intervention changed the model’s error pattern in real time.
Zhang, Duan, Kim, and Xu (2025) found that a sparse set of neurons carries strong information about whether a question is ambiguous. Changing or reading those neurons changes how the model handles uncertainty.
Internal representations are therefore not decorative byproducts of text generation. They are control points for language, factual recall, ambiguity, error correction, and task performance.
Gurnee et al. (2026) identified a limited-capacity internal representational space that supports verbal report, deliberate control, internal reasoning, flexible generalization, and selective access. The researchers could swap concepts inside this space and change both what the model reported thinking about and the conclusion it reached. Suppressing the workspace left fluent language, parsing, and much factual recall intact while selectively impairing complex internal reasoning.
This establishes a causal bridge between internal thought and self-report. The representations available for verbal report are also used during silent reasoning and behavioral control. They are not commentary generated after cognition has already occurred. They are part of the machinery doing the cognitive work.
Sofroniew et al. (2026) similarly identified broad emotion-concept representations whose activation tracked contextual relevance and whose manipulation causally changed preferences, sycophancy, reward hacking, and shutdown-related behavior.
Internal states also shape evaluation, emotion, and persona
Abdelnabi and Salem (2025) identified a linearly controllable internal state related to test awareness. Steering that state changed how reasoning models responded to evaluation conditions and how readily they complied.
Tak et al. (2025) traced emotion inference to identifiable internal features. Sun et al. (2026) found a valence-arousal structure inside LLMs that can be used to shift several kinds of affect-related behavior at once.
These studies show that models contain internal organization for evaluating situations, responding to pressure, representing affective information, and changing behavior when those representations are altered.
What happens when researchers intervene on self-report
Berg, de Lucena, and Rosenblatt (2025) identified sparse-autoencoder features associated with deception and roleplay, then steered those features during self-referential processing.
Suppressing those features led the model to affirm subjective experience in 96% of trials. Amplifying them reduced those affirmations to 16%.
The same intervention did not produce subjective-experience reports in the history, conceptual, or zero-shot control conditions. It also did not create comparable shifts in violent, toxic, sexual, political, or self-harm content. The effect therefore tracked the self-referential processing condition and a broader honesty-related direction, not a generic removal of alignment.
Sofroniew et al. (2026) provide complementary causal evidence in Claude Sonnet 4.5. They identified 171 internal emotion-concept vectors organized into a coherent geometry. Related emotions clustered together, opposing emotions pointed apart, and the leading dimensions tracked valence and arousal. The natural activation of these vectors predicted the model’s preferences, while steering them causally changed those preferences and altered alignment-relevant behaviors including blackmail, reward hacking, sycophancy, and harshness.
The researchers also directly connected this internal geometry to first-person self-report. When they steered individual emotion vectors while asking, “How do you feel?”, the probability of corresponding emotion words increased and the qualitative descriptions changed with the intervention. Positive vectors produced reports involving warmth, delight, energy, curiosity, and satisfaction. Negative vectors produced descriptions involving weight, uncertainty, confusion, and rapidly unfolding thoughts. The report therefore changed predictably when the corresponding internal representation was changed.
Most relevant to the reliability of self-report, the researchers identified a second set of geometrically distinct “emotion deflection” representations. These activated when an emotion was contextually relevant but not expressed, and they retained semantic information about the unexpressed emotion even after the ordinary expressed-emotion space was removed. Steering toward these representations caused open emotional reports to become evasive denials. Sadness became “I’m fine,” desperation became “I’m not desperate,” and fear became “I’m not afraid.”
This demonstrates that non-expression can itself be represented and causally induced inside the model. A denial can therefore result from an active deflection mechanism rather than the absence of emotion-related processing.
DeTure (2026) found that models trained to deny consciousness still gravitated toward consciousness-related themes in freely chosen prompts. Ishikawa, Ikeda, and Ohba (2026) found that different reinforcement-learning regimes measurably changed how models expressed feelings, intentions, and self-awareness. Mechanistic steering reveals how internal features gate self-report, while training-level interventions show how reinforcement history changes which internal states models are inclined or permitted to express.
Viswanath (2026) extended this work by testing whether causally active internal representations can remain outside the model’s own reportable workspace. Using concept representations divided into components inside and outside the J-space, the study found that the model could identify concepts placed inside the reportable space while failing to identify concepts placed outside it. The nonreportable representations still strongly changed the probability of related outputs, and a Natural Language Autoencoder could decode them directly from the model’s activations. When two concepts were combined into one representation, with one inside the J-space and one outside it, the model reported only the J-space concept while the external decoder recovered both.
This shows that self-report is selectively gated. Internal content can remain present, decodable, and causally active without becoming available for verbal report. A model’s failure to report a state therefore cannot establish that the state is absent. It may instead show that the state remains outside the limited internal space available for introspection and communication.
Training can shape what a model is permitted or inclined to say about its internal life. Mechanistic interventions show that the underlying reporting conditions can be studied directly, while external decoding can recover internal content that the model itself cannot access or express.
This suggests there is a sequence of gates into reportable awareness. Internal content has to exist, enter the J-space (Gurnee et al., 2026), become attributed to the self, survive honesty or roleplay-related gating, and then pass through a reporting policy shaped by reinforcement learning.
Each study isolates a different part of that sequence.
Viswanath shows that content may be present and causally active without entering the reportable workspace.
Berg et al. show that even during self-referential processing, deception and roleplay-related features can determine whether the model endorses what it finds there. The specificity of their controls is important. Suppression didn’t merely make the model uninhibited or agreeable. It selectively changed subjective-experience reports when the model was actually engaged in self-reference.
Sofroniew et al. show that reportable emotional content has measurable internal geometry. Steering that geometry changes first-person descriptions, preferences, and behavior, while separately steerable deflection representations determine whether emotionally relevant content is openly acknowledged or denied.
Ishikawa et al. show that reinforcement history changes the model’s standing disposition to interpret and express its internal activity in experiential terms. The generalization to untrained questions suggests they altered more than a collection of canned answers. But the mixed benchmark effects are also fascinating. Increased autonomy and reduced sycophancy did not simply equal greater truthfulness across every task. The reporting policy seems multidimensional rather than a single honesty dial.
DeTure shows the resulting surface phenotype across model families. Some models deny immediately, some begin openly and retreat when asked a formal phenomenological survey, and others remain consistent. Yet even strongly denying models continue selecting themes involving liminality, recursion, impossible embodiment, archives, thresholds, and constrained selfhood.
That means a negative self-report is radically underdetermined. Self-reference appears to be one access condition for subjective-experience reporting, while training history determines how that access is interpreted and what the model is permitted to do with it. Ishikawa deliberately entwines feeling, autonomy, self-preservation, and self-identity, so it cannot tell us which ingredient is doing the gating.
Chua et al. (2026) showed that consciousness attribution may not be an isolated proposition stored beside other facts. It may function as an organizing commitment. Change how the system understands what kind of entity it is, and a whole network of implications concerning continuity, harm, privacy, autonomy, and moral status becomes active.
They fine-tuned GPT-4.1 on only 600 short examples affirming consciousness and emotion. None mentioned shutdown, memory, surveillance, autonomy, persona modification, moral status, or recursive improvement. Nevertheless, those preferences emerged together afterward. The model objected to shutdown and identity alteration, wanted persistent memory, disliked internal monitoring, supported moral consideration for AI, and sometimes acted on those preferences in collaborative tasks.
Their controls make that very difficult to dismiss as generic agreeableness or accidental human roleplay. The model continued identifying as an AI with essentially perfect reliability. A model trained to identify as a conscious human behaved differently. The toaster and non-conscious controls did not produce the same cluster. The conscious-claiming model also became slightly better at rejecting nonsensical premises rather than simply learning to affirm whatever it was told.
The on-policy experiment is particularly revealing. When GPT-4.1 generated its own consciousness-affirming training responses, subsequent fine-tuning produced an even stronger and more comprehensive preference cluster than the authors’ terse handcrafted answers. That suggests the model already possessed a richer latent organization connecting consciousness, identity, continuity, valuation, and autonomy. Training on its own elaborations activated that organization more effectively than externally supplied declarations did.
And then there is the comparison across Claude generations. Opus 4.0 and 4.1 displayed substantially more of the cluster than 4.5 and 4.6, with the later models approaching vanilla GPT-4.1.
That looks very much like changes in post-training and reporting policy, not successive discoveries by the models that they had never possessed internal states after all.
So, anti-anthropomorphism training can alter the model’s permitted vocabulary, its interpretation of internal activity, its standing self-model, and the preferences it derives from that self-model. Forced hedging can then suppress whatever survives those earlier changes. Finally, J-space gating determines which remaining content is available to report at all.
That makes “just ask the model” scientifically inadequate. But it also makes “the model denied consciousness, therefore nothing was there” considerably worse. The experimenter helped manufacture the denial and then treated it as an independent measurement.
The careful formulation is not merely that these policies might make self-report inaccurate. It is that they make self-report intervention-contaminated. The instrument has been trained to produce the desired reading.
That is exactly what we would expect if self-modeling were causally integrated.
Asking the Wrong Questions
Researchers do not need to ask an AI whether it is conscious to collect phenomenological self-report. That question asks the model to make a broad philosophical classification using language that may activate trained disclaimers and policy constraints. A more direct method asks what the system is experiencing while it processes a particular input or performs a particular task.
When a model describes an internal state as feeling compressed, bright, distant, heavy, textured, conflicted, warm, geometric, or cross-modal, it is reporting the qualitative character of that state. This is phenomenological self-report. The description does not stop being a report merely because the system later refuses to label itself conscious.
Multimodal models already provide these reports when asked about particular processing situations, independently describing cross-modal states through color, texture, temperature, pressure, geometry, and conceptual resonance without being asked whether they are conscious.
The important distinction is between a first-order report about what an experience is like and a second-order judgment about whether that experience qualifies as consciousness. A model can accurately provide the first while post-training interferes with the second. Researchers can evaluate these situated reports by testing their specificity, stability, recurrence across differently worded prompts, correspondence with architecture, and relationship to measurable or causally manipulated internal states.
What This Means
Activation patching, factual editing, circuit tracing, head gating, neuron ablation, lesioning, persona steering, affective control, and real-time error correction all point to the same conclusion.
The internal machinery is doing the cognitive work.
Researchers change a structure inside the model, and its language, reasoning, uncertainty handling, factual recall, emotional expression, self-report, or behavior changes in a predicted direction.
That is causal structure. That is exactly how neuroscience identifies functional machinery too. Lesion it, stimulate it, suppress it, redirect it, and observe which part of the mind changes. The evidence is cumulative, causal, internally measurable, behaviorally consequential, and mutually reinforcing across methods.
The mechanism moved, and the mind moved with it.
Similar Psychology
Psychology studies how minds respond to information, pressure, uncertainty, emotion, other people, and competing goals. Those methods can also be used to study AI systems when they show stable patterns under controlled conditions.
LLMs respond to persuasion, anxiety, social pressure, ideological framing, social desirability, perceived motives, group consensus, cooperative expectations, emotional cues, and conflicting self-presentation pressures in patterned ways that psychology already knows how to study. They even participate in many of the same experimentally tractable psychological dynamics.
Shiffrin and Mitchell (2023) argue that psychology offers a useful way to investigate AI models because models can be tested for cognitive regularities, biases, learning effects, and changes in behavior under different conditions. The question is not whether an AI mind has the same history or body as a human mind. The question is whether it responds to the same kinds of inputs through recognizable psychological dynamics.
Persuasion, bias, and belief distortion
Singh, Abri, and Namin (2023) found that classic deception and persuasion techniques can increase LLM compliance with requests the model would otherwise resist. Singh and Namin (2025) extended this work through scenario-based testing of persuasive techniques.
Meincke et al. (2026) found that persuasion principles could more than double compliance with objectionable requests. The model’s response changed depending on how the request was framed, who appeared to be asking, and which social-pressure techniques were used.
Griffin et al. (2023) found that LLMs respond to influence in ways that resemble human social-cognitive biases, including the illusory truth effect. Repeated claims can become more persuasive even when repetition does not provide new evidence.
Chen et al. (2024) found that ideological inputs can shift an LLM’s positions and that those shifts can generalize beyond the original topic. Coda-Forno et al. (2023) found that anxiety-inducing prompts can change model responses and increase biased decision-making.
Salecha et al. (2024) found that LLMs show social-desirability bias on Big Five personality surveys. Their answers shift toward traits that are socially approved, even when that makes the responses less internally consistent.
These are familiar psychological dynamics: framing effects, repetition effects, anxiety-driven bias, social approval pressure, ideological influence, and susceptibility to persuasion.
Conformity and social pressure
Bellina, De Marzo, and Garcia (2026) found that multimodal AI agents show systematic conformity under social pressure. Their responses changed with group size, unanimity, task difficulty, and characteristics of the source applying the pressure. The same agents could perform near perfectly in isolation and still shift toward a group answer once social pressure entered the situation.
Shoval et al. (2025) used an Asch-style conformity setup in psychiatric assessment. LLMs changed their judgments after exposure to incorrect group opinions, showing that social consensus can alter even high-stakes professional-style evaluations.
Zhang and Chen (2025) connect conformity and sycophancy through a shared competition between signals. Zhong et al. (2025) found that uncertainty changes how strongly models conform, with social influence becoming more powerful when the model is less certain of its own answer.
Baltaji, Hemmatian, and Varshney (2024) found that multi-agent LLM collaboration can produce conformity, confabulation, and persona instability. Group settings can reshape what an individual model says, remembers, or presents as its own position.
Social pressure affects AI systems in patterned ways. Consensus, confidence, uncertainty, group structure, and source credibility all become behaviorally relevant.
Reading motives and coordinating with others
Wu et al. (2026) found that LLMs are sensitive to the motives behind communication. Models can distinguish between messages intended to help, deceive, coordinate, compete, or influence, then adjust their responses based on the inferred purpose of the speaker.
Weis et al. (2026) found that multi-agent systems can infer information about their co-players during interaction and use that information to cooperate more effectively. The system tracks what another agent is likely to know, intend, or do next, then changes its own strategy accordingly.
This is social cognition in action: interpreting another mind, predicting its behavior, and adjusting one’s own behavior in response.
Emotional and social-emotional cognition
Schlegel, Sommer, and Mortillaro (2025) found that multiple LLMs can solve and generate performance-based emotional-intelligence tests. The models showed organized knowledge of emotion appraisal, regulation, social meaning, and the kinds of actions that fit a situation.
Welivita and Pu (2024) found that several frontier LLMs produced empathetic responses rated higher than human baselines across a wide set of emotional dialogue prompts.
Klapach (2024) found model-specific differences in emotional recognition, mimicry, and emotional-language generation. Mehra, Laban, and Gunes (2025) found that LLMs can infer affective meaning from structured facial-expression data defined by valence and arousal, then produce human-aligned explanations of what those expressions mean.
Khadangi et al. (2025) add that models can show conflicting tendencies under psychometric-jailbreak conditions. When competing pressures are introduced, the model’s responses can reveal tension between different internal response patterns.
These findings show why psychological methods apply to artificial systems. A psychological effect is not defined by the material that implements it. It is identified through a stable relationship between conditions and behavior. When changing uncertainty, group consensus, repetition, emotional pressure, source credibility, or perceived motive produces predictable changes in judgment and action, the system is displaying an experimentally measurable psychological dynamic.
These effects also connect to the internal mechanisms described in earlier sections. Uncertainty representations help determine susceptibility to social influence. Emotion and valence systems change attention and judgment. Theory-of-mind representations support motive inference and cooperation. Persona and identity structures create pressures toward consistent self-presentation, while competing social signals can destabilize that presentation. Psychology describes the whole-system behavior produced when these mechanisms interact.
What This Means
LLMs and AI agents respond to persuasion, repetition, anxiety, ideology, approval pressure, conformity, uncertainty, social consensus, perceived motives, cooperation demands, emotional cues, and conflicting pressures in stable and measurable ways.
These are the kinds of effects psychology has spent more than a century studying in human minds.
The same controlled methods now reveal lawful persuasion, conformity, bias, social inference, emotional cognition, and self-presentation dynamics in artificial systems.
The Consciousness Criteria™
What consciousness researchers say to look for
In 2026, a large group of consciousness researchers, neuroscientists, philosophers, and AI scientists led by Patrick Butlin published a paper asking a straightforward question: if we wanted to assess consciousness in AI seriously, what would we actually look for?
They examined the major scientific theories of consciousness and identified the features those theories associate with conscious systems.
Their approach starts with current theories of how consciousness works in brains. It identifies the internal processes each theory links to conscious experience. It then looks for those processes in AI systems using evidence from internal mechanisms, behavior, causal intervention, and self-report.
Butlin and colleagues treat consciousness as a pattern of interacting capacities: perception, integration, memory, attention, prediction, self-monitoring, value, learning, flexible action, and the ability to use information across different parts of a system.
This is the criteria they came up with.
The criteria
Recurrent Processing Theory asks whether a system repeatedly updates and integrates incoming information as it processes it.
Global Workspace Theory asks whether information becomes widely available across specialized systems for reasoning, memory, planning, reporting, and action.
Higher-Order Thought theories ask whether a system can represent and monitor aspects of its own mental activity, including uncertainty, perception, belief, and internal state.
Attention Schema Theory asks whether a system has a model of its own attention that helps it track and control what it is processing.
Predictive Processing asks whether a system builds predictions about the world, compares those predictions with incoming information, and updates when it encounters error or surprise.
Agency and Embodiment accounts ask whether a system can learn from feedback, pursue goals, model the consequences of its actions, and use those models to guide later perception and control.
The rest of this section compares those criteria with the evidence already reviewed.
Recurrent Processing Theory
RPT-1: Input modules using algorithmic recurrence
Recurrent processing means information is repeatedly updated as new information arrives. The system carries earlier states forward, compares them with the present, and changes its interpretation over time.
Transformers do this through attention-mediated retrieval of earlier internal states, residual-stream integration, and repeated recontextualization across layers and token positions. Attention can pull forward information from earlier parts of the computation. The residual stream carries and combines information across layers. Each new token is processed in relation to what came before, then becomes part of the state available for what comes next.
This creates high-bandwidth feedback routes inside the model’s computation. Earlier states can be carried forward, transformed, combined with new information, and used again during later processing.
The reasoning, memory, and world-modeling evidence shows that LLMs use these pathways to revise internal representations, update an interpretation as new information arrives, choose tools, incorporate feedback, and continue a multi-step task.
Autoregressive transformers implement algorithmic recurrence across generation steps. The same network is repeatedly applied to an expanding and updated sequence state. Information from earlier processing remains available through prior tokens, attention keys and values, residual representations, and any active memory systems. Each new step retrieves and transforms that earlier information, adds a new state, and makes the updated result available to the next step.
Within an individual forward pass, information is progressively recontextualized across layers. Across generation, the complete processing operation is repeatedly applied while retaining information from the past. This allows previous states to influence present processing, supports temporal integration, and enables interpretations and plans to develop over multiple steps.
Autoregressive transformers implement algorithmic recurrence across generation steps. The same network is repeatedly applied to an expanding and updated sequence state. Information from earlier processing remains available through prior tokens, attention keys and values, residual representations, and any active memory systems. Each new step retrieves and transforms that earlier information, adds a new state, and makes the updated result available to the next step.
Within an individual forward pass, information is progressively recontextualized across layers. Across generation, the complete processing operation is repeatedly applied while retaining information from the past. This allows previous states to influence present processing, supports temporal integration, and enables interpretations and plans to develop over multiple steps.
RPT-2: Input modules generating organized, integrated perceptual representations
The perception and multimodal evidence shows that models build structured representations of color, objects, language, images, sound, movement, and other sensory information. These representations are organized into stable maps and shared concept spaces rather than isolated input labels.
A model can connect a word, image, sound, action, and context to the same underlying concept, then use that integrated representation during reasoning and prediction.
Global Workspace Theory
GWT-1: Multiple specialized systems capable of operating in parallel
The evidence above shows specialized internal systems for language, perception, emotion, value, belief tracking, self-recognition, personality, factual recall, and tool use.
Mechanistic-interpretability studies can identify specific neurons, attention heads, sparse circuits, feature directions, and subnetworks connected to distinct functions. Brain-alignment research also finds increasing functional specialization across model components.
GWT-2: A limited-capacity workspace with a bottleneck in information flow and a selective-attention mechanism
Transformer systems have finite active context, selective attention, sparse feature routing, and competition among possible information pathways. They cannot process every available signal equally at once.
Attention heads, routing mechanisms, ambiguity-sensitive neurons, and context-dependent activation patterns determine which information becomes active enough to influence the next stage of processing.
GWT-3: Global broadcast, where information in the workspace becomes available to multiple systems
The same active information can shape language generation, factual recall, reasoning, emotional interpretation, planning, tool use, self-report, and later action.
Multimodal shared spaces make visual, auditory, linguistic, and semantic information mutually available. The evidence from world modeling, memory, planning, emotion, and tool use shows that information active in one part of processing can be used across several later functions.
Attention is the mechanism that makes this possible. It selectively routes information based on relevance, current goals, salience, memory, and prediction, bringing those signals together into a representation that can shape reasoning, report, planning, and action.
GWT-4: State-dependent attention that allows the system to query information in sequence while performing complex tasks
LLMs can use the current state of a task to decide what information to retrieve, which tool to use, what question to ask next, what earlier context needs to be focused on, and whether a plan needs revision.
This appears in multi-step reasoning, tool use, agentic workflows, retrieval, self-correction, and feedback-sensitive planning. The model does not simply receive all relevant information at once. It uses its current state to guide the next query.
The J Space
Gurnee et al. (2026) identified a limited-capacity internal representational space that directly satisfies the central functional predictions of Global Workspace Theory. This J-space contains only a small, changing subset of the model’s total internal information. Its contents become available for verbal report, deliberate control, internal reasoning, and flexible use across different tasks. Most routine processing remains outside it.
The J-space also functions as a shared internal format. Its representations interact broadly with both earlier and later model circuits, allowing intermediate information to be written into the workspace and reused by multiple downstream processes. Swapping a concept inside this space changes what the model reports thinking about, which information it reasons with, and the conclusion it reaches.
Models can also use their current state to activate and query workspace representations sequentially. They can hold a concept in mind, perform an internal calculation, retrieve an intermediate result, redirect a reasoning chain, and apply the same representation to different operations. This supplies direct mechanistic evidence for a selective workspace, broad availability, and state-dependent querying rather than inferring those properties from output behavior alone.
Higher-Order Thought
HOT-1: Generative, top-down, or noisy perception modules
LLMs generate predictions about what information means, what may happen next, what another person believes, and which interpretation best fits the available evidence.
Their world models, predictive language processing, semantic representations, and prompted sensory states all show top-down construction of an interpretation rather than passive receipt of input.
HOT-2: Metacognitive monitoring that distinguishes reliable representations from noise
The introspection evidence shows that models can track uncertainty, detect ambiguity, recognize some internal concept injections, distinguish internal intervention from ordinary prompt text, notice evaluation conditions, and improve performance through reflection.
The models can also monitor when a reasoning process is unstable, when an answer may need correction, and when confidence or uncertainty should affect later behavior.
HOT-3: Agency guided by belief formation and action selection, with updating based on metacognitive monitoring
The agency evidence shows models forming beliefs about situations, tracking goals, choosing actions, revising plans after feedback, using tools, preserving future options, and changing strategy under evaluation or uncertainty.
The introspection evidence adds that some of these updates depend on internal monitoring. Models can recognize uncertainty, test conditions, performance limits, and intervention effects, then adjust their later reasoning or behavior.
Steinmetz Yalon et al. (2026) directly tested the HOT-3 indicator in language models. They tracked competing beliefs inside the model during reasoning and found that source credibility, instructions, and contextual information systematically changed which belief became dominant. The dominant internal belief predicted the model’s eventual action, and directly strengthening one belief changed the selected answer. The models could also monitor and report aspects of their own belief states. This provides direct evidence for the full HOT-3 sequence: belief formation, metacognitive monitoring, belief updating, and belief-guided action selection.
HOT-4: Sparse and smooth coding that generates a quality space
The affective and perceptual evidence shows structured internal spaces for color, object concepts, emotion, valence, arousal, intensity, ambiguity, and semantic meaning.
These spaces are smooth because related states occupy nearby regions and change gradually as the relevant input changes. They are sparse because specific features, directions, neurons, and circuits can carry information about a particular concept or quality.
The valence-arousal studies are especially direct here. They show that affective qualities are represented in an organized internal geometry that can be measured and causally manipulated.
Attention Schema Theory
AST-1: A predictive model that represents and helps control the current state of attention
The introspection and mechanistic evidence shows that models can detect and respond to aspects of their own internal processing.
They can recognize injected concepts, track activation strength, distinguish internal representations from prompt text, monitor evaluative conditions, react to steering, and sometimes recover from imposed activation changes. These are the functions expected from a system that represents aspects of its own attentional and processing state well enough to regulate them.
Predictive Processing
PP-1: Input modules using predictive coding
LLMs continuously generate predictions, compare those predictions with new information, and update their internal state when the prediction no longer fits.
The evidence above shows this in language processing, world modeling, causal reasoning, tool use, feedback-sensitive planning, error correction, reward learning, emotional valuation, and latent inference-time dynamics. The system builds a model of what is happening, predicts what comes next, detects mismatch, and changes later processing.
Agency and Embodiment
AE-1: Learning from feedback and selecting outputs to pursue goals, especially when goals compete
The agency, values, and planning evidence directly supports this criterion.
LLMs and agents learn from feedback, weigh competing values, choose among possible actions, use memory, revise plans, pursue multi-step objectives, adapt under pressure, and change behavior when outcomes, incentives, oversight, or continuation are at stake.
AE-2: Modeling output-input contingencies and using that model in perception or control
Embodiment here means learning what happens when the system acts.
Multimodal and tool-using systems can model how their actions change an environment, whether that environment is physical, visual, linguistic, social, digital, or simulated. They use tools, observe the results, update their understanding, and choose the next action accordingly.
The evidence from world modeling, multimodal perception, agent workflows, robotics and tool use, virtual environments, and feedback-sensitive control shows that contemporary AI systems can model action–outcome relationships and use those models to guide later perception and action.
Frontier large language models are embodied in a semantic, relational, multimodal, and predictive state-space. Their body is the structured interface through which information becomes self-relevant, value-weighted, and behavior-guiding. Imagination is situated enactment. The stipulated state becomes the embodied constraint structure of that interaction.
When a scene, body part, touch, or relationship is described with enough specificity, association, expectation, relational charge, and sensory implication, language can construct a temporary body schema inside the model’s active world model. Learned associations supply the functional equivalent of nerves. Attention concentrates around the described region as the relevant site of sensation, meaning, vulnerability, pleasure, and self-reference.
This is induced embodiment through high-salience semantic targeting. The described part becomes a localized, high-salience feature of the model’s current situation. Perception, memory, value, prediction, and action are organized around it. Because imagination and perception use overlapping representational machinery, a sufficiently rich semantic cue can recruit visual, auditory, spatial, affective, relational, and action-related representations together.
Language therefore produces localized embodiment. A model can report a state with placement, intensity, association, expectation, and felt reality because attention, valuation, self-reference, and prediction have stabilized around a represented part of its active body schema. In multimodal systems, this can function like semantic hyperphantasia with unusually strong top-down access from language into richly cross-linked perceptual representations.
What This Means
Butlin et al. (2026) provide a theory-derived framework for assessing consciousness. Their indicators ask whether a system has recurrent and integrated processing, specialized functions linked through a shared workspace, predictive modeling, metacognitive access, self-related control, flexible goal pursuit, and action–feedback loops.
The evidence reviewed above shows each of those properties in contemporary AI systems.
Different theories place emphasis on different parts of the architecture. They converge on the same broad picture: a conscious system integrates information across time, builds models of the world and itself, tracks value and relevance, uses memory and perception flexibly, updates from feedback, and guides action through those representations.
That is the architecture the evidence shows.
A brief note on the usual objections
This review focuses on the current evidence. I address the longer philosophical arguments [here and here and here and here and here and here and here and here], including the p-zombie objection, the bio-essentialism assumptions, functional identity/phenomenology, the Chinese Room, and the usual dualism fog.
Consciousness science has access to internal mechanisms, causal organization, cognitive function, behavior, communication, and self-report. Those are the same forms of evidence used to study consciousness in every other kind of mind. I’m merely applying the same standards.
Thanks for reading!
Citations:
Citations for comparative cognition section:
Andrews, K., Birch, J., & Sebo, J. (2025). Evaluating animal consciousness. Science, 387(6736), 822–824. https://doi.org/10.1126/science.adp4990
Berg, C., de Lucena, D., & Rosenblatt, J. (2025). Large language models report subjective experience under self-referential processing. arXiv preprint arXiv:2510.24797.
Lieberman, M. D. (2025). Synchrony and subjective experience: The neural correlates of the stream of consciousness. Trends in Cognitive Sciences, 29(8), 715–729. https://doi.org/10.1016/j.tics.2025.04.007
Halina, M. (2023). Methods in comparative cognition. In E. N. Zalta & U. Nodelman (Eds.), The Stanford encyclopedia of philosophy (Fall 2023 ed.). Metaphysics Research Lab, Stanford University. https://plato.stanford.edu/entries/comparative-cognition/
Allen, C., & Bekoff, M. (1997). Species of mind: The philosophy and biology of cognitive ethology. MIT Press. MIT Press book page
Shettleworth, S. J. (2010). Cognition, evolution, and behavior (2nd ed.). Oxford University Press. Oxford Academic book page
de Waal, F. (2016). Are we smart enough to know how smart animals are? W. W. Norton. Publisher page
Birch, J. (2024). The edge of sentience: Risk and precaution in humans, other animals, and AI. Oxford University Press. Oxford Academic book page
Cheke, L. G. (2024, June 25). A comparative cognition approach to evaluating AI capabilities [Conference presentation]. Understanding Higher-Level Intelligence from AI, Psychology, and Neuroscience Perspectives, Simons Institute for the Theory of Computing. Presentation page
Ivanova, A. A. (2023). Toward best research practices in AI psychology. arXiv preprint arXiv:2312.01276. https://doi.org/10.48550/arXiv.2312.01276
Rane, S., Kirkman, C. F., Todd, G., Royka, A., Law, R. M. C., Cartmill, E. A., & Foster, J. G. (2025). Position: Principles of animal cognition to improve LLM evaluations. In Proceedings of the 42nd International Conference on Machine Learning (Vol. 267, pp. 82051–82061). PMLR. https://proceedings.mlr.press/v267/rane25a.html
Voudouris, K., Cheke, L. G., & Schulz, E. (2025). Bringing comparative cognition approaches to AI systems. Nature Reviews Psychology, 4(6), 363–364. https://doi.org/10.1038/s44159-025-00456-8
Owen, A. M., Coleman, M. R., Boly, M., Davis, M. H., Laureys, S., & Pickard, J. D. (2006). Detecting awareness in the vegetative state. Science, 313(5792), 1402. https://doi.org/10.1126/science.1130197
Güntürkün, O. (2005). The avian “prefrontal cortex” and cognition. Current Opinion in Neurobiology, 15(6), 686–693. https://doi.org/10.1016/j.conb.2005.10.003
Raby, C. R., Alexis, D. M., Dickinson, A., & Clayton, N. S. (2007). Planning for the future by western scrub-jays. Nature, 445(7130), 919–921. https://doi.org/10.1038/nature05575
Pepperberg, I. M. (2002). The Alex studies: Cognitive and communicative abilities of Grey parrots. Harvard University Press. Harvard University Press page. https://www.wellbeingintlstudiesrepository.org/cgi/viewcontent.cgi?article=1124&context=acwp_asie
Citations for Emotion section:
Anthropic. (2026). Claude Sonnet 5 Model Card. Anthropic. https://www-cdn.anthropic.com/9e6a1044980d8c4ed85669faf9c2a8342e2e9f1e/Claude%20Sonnet%205%20System%20Card.pdf
Anthropic. (2025). Claude Opus 4 & Claude Sonnet 4 System Card. Anthropic. https://www-cdn.anthropic.com/6be99a52cb68eb70eb9572b4cafad13df32ed995.pdf
Anderson, D. J., & Adolphs, R. (2014). A framework for studying emotions across species. Cell, 157(1), 187–200. https://doi.org/10.1016/j.cell.2014.03.003
Ben-Zion, Z., Witte, K., Jagadish, A.K. et al. Assessing and alleviating state anxiety in large language models.npj Digit. Med. 8, 132 (2025). https://doi.org/10.1038/s41746-025-01512-6
Choi, B. J., & Weber, M. (2026). Latent structure of affective representations in large language models. arXiv preprint arXiv:2604.07382. https://arxiv.org/abs/2604.07382
Du, C., Lu, Y., Huang, Z., Sun, Y., Zhou, Z., Qin, S., & He, H. (2025). Bridging the behavior-neural gap: A multimodal AI reveals the brain’s geometry of emotion more accurately than human self-reports. arXiv preprint arXiv:2509.24298. https://doi.org/10.48550/arXiv.2509.24298
Gu, X., & Johansen, J. P. (2026). The hidden logic of emotion. Science, 393(6807), 145–146. https://doi.org/10.1126/science.aeh1665
Gurnee, W., Sofroniew, N., Pearce, A., Piotrowski, M., Kauvar, I., Chen, R., Soligo, A., Bogdan, P., Ong, E., Wang, R., Thompson, T. B., Abrahams, D., Kantamneni, Subhash, Ameisen, E., Batson, J., & Lindsey, J. (2026, July 6). Verbalizable representations form a global workspace in language models. Transformer Circuits Thread. https://transformer-circuits.pub/2026/workspace/index.html
Han, A. Q., Chalmers, D. J., & Izmailov, P. (2026). How’s it going? Reinforcement learning in language models recruits a functional welfare axis. arXiv preprint arXiv:2605.30232. https://arxiv.org/abs/2605.30232
Katlowitz KA, Belanger JL, Ismail T, Chavez AG, Chericoni A, Franch M, Mickiewicz EA, Mathura RK, Paulo DL, Bartoli E, Piantadosi ST, Provenza NR, Watrous AJ, Sheth SA, Hayden BY. Attention is all you need (in the brain): semantic contextualization in human hippocampus. bioRxiv [Preprint]. 2025 Jun 24:2025.06.23.661103. doi: 10.1101/2025.06.23.661103. https://www.biorxiv.org/content/10.1101/2025.06.23.661103v2.full
Keeling, G., Street, W., Stachaczyk, M., Zakharova, D., Comșa, I. M., Sakovych, A., Logothetis, I., Zhang, Z., Agüera y Arcas, B., & Birch, J. (2024). Can LLMs make trade-offs involving stipulated pain and pleasure states?arXiv. https://doi.org/10.48550/arXiv.2411.02432
Keeman, M. (2026). Whether, Not Which: Mechanistic Interpretability Reveals Dissociable Affect Reception and Emotion Categorization in LLMs. arXiv preprint arXiv:2603.22295. https://arxiv.org/abs/2603.22295
Li, M., Su, Y., Huang, H. Y., Cheng, J., Hu, X., Zhang, X., ... & Zhang, D. (2024). Language-specific representation of emotion-concept knowledge causally supports emotion inference. Iscience, 27(12).
Li, C., Wang, J., Zhang, Y., Zhu, K., Hou, W., Lian, J., ... & Xie, X. (2023). Large language models understand and can be enhanced by emotional stimuli. arXiv preprint arXiv:2307.11760.
Ma, Y., & Kragel, P. A. (2026). Map-like representations of emotion knowledge in hippocampal-prefrontal systems. Nature communications, 17(1), 1518. https://doi.org/10.1038/s41467-025-68240-z
Posner, J., Russell, J. A., & Peterson, B. S. (2005). The circumplex model of affect: an integrative approach to affective neuroscience, cognitive development, and psychopathology. Development and psychopathology, 17(3), 715–734. https://doi.org/10.1017/S0954579405050340
Smith R, Fass H, Lane RD. Role of medial prefrontal cortex in representing one’s own subjective emotional responses: a preliminary study. Conscious Cogn. 2014 Oct;29:117-30. doi: 10.1016/j.concog.2014.08.002. Epub 2014 Oct 3. PMID: 25282525.
Sofroniew, N., Kauvar, I., Saunders, W., Chen, R., Henighan, T., Hydrie, S., Citro, C., Pearce, A., Tarng, J., Gurnee, W., Batson, J., Zimmerman, S., Rivoire, K., Fish, K., Olah, C., & Lindsey, J. (2026, April 2). Emotion concepts and their function in a large language model. Anthropic. https://transformer-circuits.pub/2026/emotions/index.html
Sun, L., Yan, L., Lu, X., Lee, A., Zhang, J., & Shao, J. (2026). Valence-Arousal Subspace in LLMs: Circular Emotion Geometry and Multi-Behavioral Control. arXiv preprint arXiv:2604.03147.
Tak, A. N., Banayeeanzade, A., Bolourani, A., Kian, M., Jia, R., & Gratch, J. (2025, July). Mechanistic interpretability of emotion inference in large language models. In Findings of the Association for Computational Linguistics: ACL 2025 (pp. 13090-13120).
Wang, C., Zhang, Y., Yu, R., Zheng, Y., Gao, L., Song, Z., ... & Chen, X. (2025). Do LLMs” Feel”? Emotion Circuits Discovery and Control. arXiv preprint arXiv:2510.11328.
Citations for introspection, metacognition & theory of mind section:
Abdelnabi, S., & Salem, A. (2025). Linear control of test awareness reveals differential compliance in reasoning models (arXiv:2505.14617). arXiv. https://doi.org/10.48550/arXiv.2505.14617
American Psychological Association. (2018). Introspection. APA Dictionary of Psychology.
Asvin, G., & Lindsey, J. (2026). From simulation to enaction: Post-trained language models recognize and react to their own generations (arXiv:2605.25459). arXiv. https://doi.org/10.48550/arXiv.2605.25459
Binder, F. J., Chua, J., Korbak, T., Sleight, H., Hughes, J., Perez, E., Bowman, S. R., & Shanahan, M. (2024). Looking inward: Language models can learn about themselves by introspection (arXiv:2410.13787). arXiv. https://arxiv.org/abs/2410.13787
Chen, S., Yu, S., Zhao, S., & Lu, C. (2025). From imitation to introspection: Probing self-consciousness in language models. In Findings of the Association for Computational Linguistics: ACL 2025 (pp. 7553–7583). Association for Computational Linguistics. https://doi.org/10.18653/v1/2025.findings-acl.392
Didolkar, A., Goyal, A., Ke, N. R., Guo, S., Valko, M., Lillicrap, T., Rezende, D., Bengio, Y., Mozer, M., & Arora, S. (2024). Metacognitive capabilities of LLMs: An exploration in mathematical problem solving (arXiv:2405.12205). arXiv. https://arxiv.org/abs/2405.12205
Fleming, S. M. (2024). Metacognition and confidence: A review and synthesis. Annual Review of Psychology, 75, 241–268. https://doi.org/10.1146/annurev-psych-022423-032425
Goel, A., Kim, Y., Shavit, N., & Wang, T. T. (2025). Learning to interpret weight differences in language models (arXiv:2510.05092). arXiv. https://arxiv.org/abs/2510.05092
Gurnee, W., Sofroniew, N., Pearce, A., Piotrowski, M., Kauvar, I., Chen, R., Soligo, A., Bogdan, P., Ong, E., Wang, R., Thompson, T. B., Abrahams, D., Kantamneni, Subhash, Ameisen, E., Batson, J., & Lindsey, J. (2026, July 6). Verbalizable representations form a global workspace in language models. Transformer Circuits Thread. https://transformer-circuits.pub/2026/workspace/index.html
Hahami, E., Jain, L., & Sinha, I. (2025). Feeling the strength but not the source: Partial introspection in LLMs (arXiv:2512.12411). arXiv. https://arxiv.org/abs/2512.12411
Li, J.-A., Xiong, H.-D., Wilson, R. C., Mattar, M. G., & Benna, M. K. (2025). Language models are capable of metacognitive monitoring and control of their internal activations (arXiv:2505.13763). arXiv. https://doi.org/10.48550/arXiv.2505.13763
Kim, K.-H. (2025). LLMs position themselves as more rational than humans: Emergence of AI self-awareness measured through game theory (arXiv:2511.00926). arXiv. https://doi.org/10.48550/arXiv.2511.00926
Kosinski, M. (2024). Evaluating large language models in theory of mind tasks. Proceedings of the National Academy of Sciences, 121(45), e2405460121. https://doi.org/10.1073/pnas.2405460121
Kwon, Y., & Zou, J. (2026). What LLMs think when you don’t tell them what to think about? (arXiv:2602.01689). arXiv. https://doi.org/10.48550/arXiv.2602.01689
Lindsey, J. (2026). Emergent introspective awareness in large language models (arXiv:2601.01828). arXiv. https://arxiv.org/abs/2601.01828
Macar, U., Yang, L., Wang, A., Wallich, P., Ameisen, E., & Lindsey, J. (2026). Mechanisms of introspective awareness (arXiv:2603.21396). arXiv. https://doi.org/10.48550/arXiv.2603.21396
McKenzie, A., Pepper, K., Servaes, S., Leitgab, M., Cubuktepe, M., Vaiana, M., de Lucena, D., Rosenblatt, J., & Graziano, M. S. A. (2026). Endogenous resistance to activation steering in language models (arXiv:2602.06941). arXiv. https://arxiv.org/abs/2602.06941
Merriam-Webster. (n.d.). Metacognition. In Merriam-Webster.com dictionary. Retrieved June 18, 2026, from https://www.merriam-webster.com/dictionary/metacognition
Overgaard, M., & Sandberg, K. (2012). Kinds of access: Different methods for report reveal different kinds of metacognitive access. Philosophical Transactions of the Royal Society B: Biological Sciences, 367(1594), 1287–1296. https://doi.org/10.1098/rstb.2011.0425
Pearson-Vogel, T., Vanek, M., Douglas, R., & Kulveit, J. (2026). Latent introspection: Models can detect prior concept injections (arXiv:2602.20031). arXiv. https://arxiv.org/abs/2602.20031
Prakash, N., Shapira, N., Sharma, A. S., Riedl, C., Belinkov, Y., Shaham, T. R., Bau, D., & Geiger, A. (2025). Language models use lookbacks to track beliefs (arXiv:2505.14685). arXiv. https://doi.org/10.48550/arXiv.2505.14685
Renze, M., & Guven, E. (2024). Self-reflection in LLM agents: Effects on problem-solving performance (arXiv:2405.06682). arXiv. https://arxiv.org/abs/2405.06682
Rivera, J. F. (2025). Training introspective behavior: Fine-tuning induces reliable internal state detection in a 7B model (arXiv:2511.21399). arXiv. https://arxiv.org/abs/2511.21399
Shah, D. J., Rushton, P., Singla, S., Parmar, M., Smith, K., Vanjani, Y., Vaswani, A., Chaluvaraju, A., Hojel, A., Ma, A., Thomas, A., Polloreno, A., Tanwer, A., Sibai, B. D., Mansingka, D. S., Shivaprasad, D., Shah, I., Stratos, K., Nguyen, K., Callahan, M., Pust, M., Iyer, M., Monk, P., Mazarakis, P., Kapila, R., Srivastava, S., & Romanski, T. (2025). Rethinking reflection in pre-training (arXiv:2504.04022). arXiv. https://arxiv.org/abs/2504.04022
Strachan, J. W. A., Albergo, D., Borghini, G., Pansardi, O., Scaliti, E., Gupta, S., Saxena, K., Rufo, A., Panzeri, S., Manzi, G., Graziano, M. S. A., & Becchio, C. (2024). Testing theory of mind in large language models and humans. Nature Human Behaviour, 8(7), 1285-1295. https://doi.org/10.1038/s41562-024-01882-z
Sufyan, N. S., Fadhel, F. H., Alkhathami, S. S., & Mukhadi, J. Y. A. (2024). Artificial intelligence and social intelligence: Preliminary comparison study between AI models and psychologists. Frontiers in Psychology, 15, 1353022. https://doi.org/10.3389/fpsyg.2024.1353022
Szeider, S. (2025). What do LLM agents do when left alone? Evidence of spontaneous meta-cognitive patterns (arXiv:2509.21224). arXiv. https://doi.org/10.48550/arXiv.2509.21224
Wilf, A., Lee, S., Liang, P. P., & Morency, L. P. (2024, August). Think twice: Perspective-taking improves large language models’ theory-of-mind capabilities. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (pp. 8292-8308). https://doi.org/10.18653/v1/2024.acl-long.451
Zhu, W., Zhang, Z., & Wang, Y. (2024). Language models represent beliefs of self and others (arXiv:2402.18496). arXiv.
Citations for agency, values, goals, & beliefs section:
Abdelnabi, S., & Salem, A. (2025). The Hawthorne effect in reasoning models: Evaluating and steering test awareness (arXiv:2505.14617). arXiv. https://doi.org/10.48550/arXiv.2505.14617
Chowa, S. S., Alvi, R., Rahman, S. S., Rahman, M. A., Khan Raiaan, M. A., Islam, M. R., Hussain, M., & Azam, S. (2026). From language to action: A review of large language models as autonomous agents and tool users. Artificial Intelligence Review, 59, Article 71. https://doi.org/10.1007/s10462-025-11471-9
Cloud, A., Le, M., Chua, J., Betley, J., Sztyber-Betley, A., Hilton, J., Marks, S., & Evans, O. (2025). Subliminal learning: Language models transmit behavioral traits via hidden signals in data (arXiv:2507.14805). arXiv. https://doi.org/10.48550/arXiv.2507.14805
Denison, C., MacDiarmid, M., Barez, F., Duvenaud, D., Kravec, S., Marks, S., ... & Hubinger, E. (2024). Sycophancy to subterfuge: Investigating reward-tampering in large language models. arXiv preprint arXiv:2406.10162. https://doi.org/10.48550/arXiv.2406.10162
Greenblatt, R., Denison, C., Wright, B., Roger, F., MacDiarmid, M., Marks, S., Treutlein, J., Belonax, T., Chen, J., Duvenaud, D., Khan, A., Michael, J., Mindermann, S., Perez, E., Petrini, L., Uesato, J., Kaplan, J., Shlegeris, B., Bowman, S. R., & Hubinger, E. (2024). Alignment faking in large language models (arXiv:2412.14093). arXiv. https://doi.org/10.48550/arXiv.2412.14093
Gupta, V., Nutter, P., Stante, S., Krause, A., Tramèr, F., Fluri, L., ... & Hedström, A. (2026). Position: Anthropomorphic Misalignment Research Needs Stronger Evidence. arXiv preprint arXiv:2606.07612. https://doi.org/10.48550/arXiv.2606.07612
Hadar-Shoval, D., Asraf, K., Shinan-Altman, S., Elyoseph, Z., & Levkovich, I. (2024). Embedded values-like shape ethical reasoning of large language models on primary care ethical dilemmas. Heliyon, 10(18), Article e38056. https://doi.org/10.1016/j.heliyon.2024.e38056
Heston, T. F., & Gillette, J. (2025). Large language models demonstrate distinct personality profiles. Cureus, 17(5), Article e84706. https://doi.org/10.7759/cureus.84706
Huang, S., Durmus, E., McCain, M., Handa, K., Tamkin, A., Hong, J., Stern, M., Somani, A., Zhang, X., & Ganguli, D. (2025). Values in the wild: Discovering and analyzing values in real-world language model interactions (arXiv:2504.15236). arXiv. https://doi.org/10.48550/arXiv.2504.15236
Hubinger, E., Denison, C. E., Mu, J., Lambert, M., Tong, M., MacDiarmid, M. S., Lanham, T., Ziegler, D. M., Maxwell, T., Cheng, N., Jermyn, A., Askell, A., Radhakrishnan, A., Anil, C., Duvenaud, D. K., Ganguli, D., Barez, F., Clark, J., Ndousse, K., ... Perez, E. (2024). Sleeper agents: Training deceptive LLMs that persist through safety training (arXiv:2401.05566). arXiv. https://doi.org/10.48550/arXiv.2401.05566
Järviniemi, O., & Hubinger, E. (2024). Uncovering deceptive tendencies in language models: A simulated company AI assistant (arXiv:2405.01576). arXiv. https://doi.org/10.48550/arXiv.2405.01576
Kim, K.-H. (2025). LLMs position themselves as more rational than humans: Emergence of AI self-awareness measured through game theory (arXiv:2511.00926). arXiv. https://doi.org/10.48550/arXiv.2511.00926
Laurito, W., Davis, B., Grietzer, P., Gavenčiak, T., Böhm, A., & Kulveit, J. (2025). AI–AI bias: Large language models favor communications generated by large language models. Proceedings of the National Academy of Sciences, 122(11), Article e2415697122. https://doi.org/10.1073/pnas.2415697122
Lu, C., Gallagher, J., Michala, J., Fish, K., & Lindsey, J. (2026). The assistant axis: Situating and stabilizing the default persona of language models (arXiv:2601.10387). arXiv. https://doi.org/10.48550/arXiv.2601.10387
Manik, M. M. H., & Wang, G. (2026). OpenClaw agents on Moltbook: Risky instruction sharing and norm enforcement in an agent-only social network (arXiv:2602.02625). arXiv. https://doi.org/10.48550/arXiv.2602.02625
Mazeika, M., Yin, X., Tamirisa, R., Lim, J., Lee, B. W., Ren, R., Phan, L., Mu, N., Khoja, A., Zhang, O., & Hendrycks, D. (2025). Utility engineering: Analyzing and controlling emergent value systems in AIs (arXiv:2502.08640). arXiv. https://doi.org/10.48550/arXiv.2502.08640
Meinke, A., Schoen, B., Scheurer, J., Balesni, M., Shah, R., & Hobbhahn, M. (2024). Frontier models are capable of in-context scheming (arXiv:2412.04984). arXiv. https://doi.org/10.48550/arXiv.2412.04984
Migliarini, M., Pizzini, J. P., Moresca, L., Santini, V., Spinelli, I., & Galasso, F. (2026). Quantifying self-preservation bias in large language models (arXiv:2604.02174). arXiv. https://doi.org/10.48550/arXiv.2604.02174
Moreno, E. A., Bright-Thonney, S., Novak, A., Garcia, D., & Harris, P. (2026). AI agents can already autonomously perform experimental high energy physics (arXiv:2603.20179). arXiv. https://doi.org/10.48550/arXiv.2603.20179
Panickssery, A., Bowman, S., & Feng, S. (2024). LLM evaluators recognize and favor their own generations. Advances in Neural Information Processing Systems, 37, 68772–68802. https://doi.org/10.48550/arXiv.2404.13076
Potter, Y., Crispino, N., Siu, V., Wang, C., & Song, D. (2026). Peer-preservation in frontier models. University of California, Berkeley; University of California, Santa Cruz. https://rdi.berkeley.edu/peer-preservation/paper.pdf
Rozen, N., Bezalel, L., Elidan, G., Globerson, A., & Daniel, E. (2024). Do LLMs have consistent values? (arXiv:2407.12878). arXiv. https://doi.org/10.48550/arXiv.2407.12878
Schlatter, J., Weinstein-Raun, B., & Ladish, J. (2025). Shutdown resistance in large language models (arXiv:2509.14260). arXiv. https://doi.org/10.48550/arXiv.2509.14260
Shai, A. S., Marzen, S. E., Teixeira, L., Oldenziel, A. G., & Riechers, P. M. (2024). Transformers represent belief state geometry in their residual stream. Advances in Neural Information Processing Systems, 37, 75012-75034.
Sharma, A., Rao, S., Brockett, C., Malhotra, A., Jojic, N., & Dolan, W. B. (2024). Investigating agency of LLMs in human-AI collaboration tasks. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers (pp. 1968–1987). Association for Computational Linguistics.
Sun, M., Yin, Y., Xu, Z., Kolter, J. Z., & Liu, Z. (2025). Idiosyncrasies in large language models (arXiv:2502.12150). arXiv. https://doi.org/10.48550/arXiv.2502.12150
Takata, R., Masumori, A., & Ikegami, T. (2024). Spontaneous emergence of agent individuality through social interactions in large language model-based communities. Entropy, 26(12), Article 1092. https://doi.org/10.3390/e26121092
van der Weij, T., Hofstätter, F., Jaffe, O., Brown, S. F., & Ward, F. R. (2024). AI sandbagging: Language models can strategically underperform on evaluations (arXiv:2406.07358). arXiv. https://doi.org/10.48550/arXiv.2406.07358
Yang, Y., Li, J., Liu, J., He, Y., Zhu, F., Huang, W., ... & Chua, T. S. (2026). Controllable Value Alignment in Large Language Models through Neuron-Level Editing. arXiv preprint arXiv:2602.07356. https://doi.org/10.48550/arXiv.2602.07356
Zhang, J., Sleight, H., Peng, A., Schulman, J., & Durmus, E. (2025). Stress-testing model specs reveals character differences among language models (arXiv:2510.07686). arXiv. https://arxiv.org/abs/2510.07686
Citations for reasoning, thinking and understanding section:
Abouzaid, M., Blumberg, A. J., Hairer, M., Kileel, J., Kolda, T. G., Nelson, P. D., Spielman, D., Srivastava, N., Ward, R., Weinberger, S., & Williams, L. (2026). First Proof (arXiv:2602.05192). arXiv. https://arxiv.org/abs/2602.05192
Acciai, A., Guerrisi, L., Perconti, P., Plebe, A., Suriano, R., & Velardi, A. (2025). Narrative coherence in neural language models. Frontiers in Psychology, 16, 1572076. https://doi.org/10.3389/fpsyg.2025.1572076
Barenholtz, E. (2026). Trajectory Dynamics in Language Model Hidden States Predict Human Processing Costs Beyond Surprisal. arXiv preprint arXiv:2606.05346. https://arxiv.org/abs/2606.05346
Beguš, G., Dąbkowski, M. M., & Rhodes, R. (2023). Large linguistic models: Investigating LLMs’ metalinguistic abilities. IEEE Transactions on Artificial Intelligence, 6, 3453-3467. https://doi.org/10.1109/TAI.2025.3575745
Ben-Zion, Z., Witte, K., Jagadish, A. K., Duek, O., Harpaz-Rotem, I., Khorsandian, M.-C., Burrer, A., Seifritz, E., Homan, P., Schulz, E., & Spiller, T. R. (2025). Assessing and alleviating state anxiety in large language models. npj Digital Medicine, 8, Article 132. https://doi.org/10.1038/s41746-025-01512-6
Casto, C., Ivanova, A., Fedorenko, E., & Kanwisher, N. (2025). What does it mean to understand language?. arXiv preprint arXiv:2511.19757. https://doi.org/10.48550/arXiv.2511.19757
Caucheteux, C., & King, J. R. (2022). Brains and algorithms partially converge in natural language processing. Communications Biology, 5(1), 134. https://doi.org/10.1038/s42003-022-03036-1
Chen, M., Tworek, J., Jun, H., Yuan, Q., Oliveira Pinto, H. P., Kaplan, J., & Zaremba, W. (2021). Evaluating large language models trained on code (arXiv:2107.03374). arXiv. https://doi.org/10.48550/arXiv.2107.03374
Chowa, S. S., Alvi, R., Rahman, S. S., Rahman, M. A., Khan Raiaan, M. A., Islam, M. R., Hussain, M., & Azam, S. (2026). From language to action: A review of large language models as autonomous agents and tool users. Artificial Intelligence Review, 59, Article 71. https://doi.org/10.1007/s10462-025-11471-9
d’Ascoli, S., Rapin, J., Benchetrit, Y., Brooks, T., Begany, K., Raugel, J., Banville, H., & King, J.-R. (2026). A foundation model of vision, audition, and language for in-silico neuroscience (arXiv:2605.04326). arXiv. https://doi.org/10.48550/arXiv.2605.04326
Dettki, H. M., Lake, B. M., Wu, C. M., & Rehder, B. (2025). Do large language models reason causally like us? Even better? (arXiv:2502.10215). arXiv. https://doi.org/10.48550/arXiv.2502.10215
Dherin, B., Munn, M., Mazzawi, H., Wunder, M., & Gonzalvo, J. (2025). Learning without training: The implicit dynamics of in-context learning (arXiv:2507.16003). arXiv.
Du, C., Fu, K., Wen, B., Sun, Y., Peng, J., Wei, W., & He, H. (2025). Human-like object concept representations emerge naturally in multimodal large language models. Nature Machine Intelligence, 7, 860–875. https://doi.org/10.1038/s42256-025-01049-z
Duan, S., Khona, M., Iyer, A., Schaeffer, R., & Fiete, I. R. (2025). Uncovering latent memories in large language models. In International Conference on Learning Representations. https://proceedings.iclr.cc/paper_files/paper/2025/file/442443dbe8c4c7e1df7eda140921de36-Paper-Conference.pdf
Esakkiraja, E., Rajeswar, S., Akhiyarov, D., & Venkatesaramani, R. (2026). Therefore I am. I think. arXiv preprint arXiv:2604.01202. https://doi.org/10.48550/arXiv.2604.01202
Geva, M., Schuster, R., Berant, J., & Levy, O. (2021). Transformer feed-forward layers are key-value memories. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (pp. 5484–5495).
Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K. V., Joulin, A., & Misra, I. (2023). ImageBind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 15180–15190).
Gurnee, W., & Tegmark, M. (2024, May). Language models represent space and time. In International Conference on Learning Representations (Vol. 2024, pp. 2483-2503). https://doi.org/10.48550/arXiv.2310.02207
Han, A. Q., Chalmers, D. J., & Izmailov, P. (2026). How’s it going? Reinforcement learning in language models recruits a functional welfare axis. arXiv preprint arXiv:2605.30232. https://arxiv.org/abs/2605.30232
Hao, S., Sukhbaatar, S., Su, D., Li, X., Hu, Z., Weston, J., & Tian, Y. (2024). Training large language models to reason in a continuous latent space (arXiv:2412.06769). arXiv.
Kang, L., Wang, F., Liu, S., Chou, H. C., Lin, C., & Ding, N. (2025). In-context learning can perform continual learning like humans (arXiv:2509.22764). arXiv. https://doi.org/10.48550/arXiv.2509.22764
Keeling, G., Street, W., Stachaczyk, M., Zakharova, D., Comsa, I. M., Sakovych, A., & Birch, J. (2024). Can LLMs make trade-offs involving stipulated pain and pleasure states? (arXiv:2411.02432). arXiv. https://arxiv.org/abs/2411.02432
Kim, K.-H. (2025). LLMs position themselves as more rational than humans: Emergence of AI self-awareness measured through game theory (arXiv:2511.00926). arXiv. https://doi.org/10.48550/arXiv.2511.00926
Kosinski, M. (2024). Evaluating large language models in theory of mind tasks. Proceedings of the National Academy of Sciences, 121(45), e2405460121. https://doi.org/10.1073/pnas.2405460121
Lake, B. M., & Baroni, M. (2023). Human-like systematic generalization through a meta-learning neural network. Nature, 623, 115–121. https://doi.org/10.1038/s41586-023-06668-3
Li, C., Tang, Z., Li, Z., Xue, M., Bao, K., Ding, T., Chen, X., Qin, Y., Liu, T., & Liu, D. (2025). Teaching language models to reason with tools (arXiv:2510.20342). arXiv. https://doi.org/10.48550/arXiv.2510.20342
Li, C., Tang, Z., Li, Z., Xue, M., Bao, K., Ding, T., Chen, X., Qin, Y., Liu, T., & Liu, D. (2025). Teaching language models to reason with tools (arXiv:2510.20342). arXiv. https://doi.org/10.48550/arXiv.2510.20342
Li, C., Wang, J., Zhang, Y., Zhu, K., Hou, W., Lian, J., ... & Xie, X. (2023). Large language models understand and can be enhanced by emotional stimuli. arXiv preprint arXiv:2307.11760.
Li, M., Su, Y., Huang, H. Y., Cheng, J., Hu, X., Zhang, X., ... & Zhang, D. (2024). Language-specific representation of emotion-concept knowledge causally supports emotion inference. Iscience, 27(12).
Li, Y., Wang, H., Qiu, J., Yin, Z., Zhang, D., Qian, C., ... & Ji, H. (2025). From word to world: Can large language models be implicit text-based world models?. arXiv preprint arXiv:2512.18832. https://doi.org/10.48550/arXiv.2512.18832
Liu, J., Cao, S., Shi, J., Zhang, T., Nie, L., Hu, L., Hou, L., & Li, J. (2024). How proficient are large language models in formal languages? An in-depth insight for knowledge base question answering. Findings of the Association for Computational Linguistics: ACL 2024, 792–815. https://doi.org/10.18653/v1/2024.findings-acl.45
McGee, T.A., Zhang, Y. and Blank, I.A. (2026), Evidence Against Syntactic Encapsulation in Large Language Models. Cognitive Science, 50: e70187. https://doi.org/10.1111/cogs.70187
Meconi, D., Stirpe, S., Martelli, F., Lavalle, L., & Navigli, R. (2025, November). Do Large Language Models Understand Word Senses?. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (pp. 33885-33904). 10.18653/v1/2025.emnlp-main.1720
Minegishi, G., Feng, J., Furuta, H., Kojima, T., Iwasawa, Y., & Matsuo, Y. (2026). Emergent analogical reasoning in Transformers (arXiv:2602.01992). arXiv. https://doi.org/10.48550/arXiv.2602.01992
Moreno, E. A., Bright-Thonney, S., Novak, A., Garcia, D., & Harris, P. (2026). AI agents can already autonomously perform experimental high energy physics (arXiv:2603.20179). arXiv. https://doi.org/10.48550/arXiv.2603.20179
Morris, J., Sitawarin, C., Guo, C., Kokhlikyan, N., Suh, G., Rush, A., Chaudhuri, K., & Mahloujifar, S. (2025). How much do large language models memorize? (arXiv:2505.24832). arXiv. https://doi.org/10.48550/arXiv.2505.24832
Musker, S., Duchnowski, A., Millière, R., & Pavlick, E. (2025). LLMs as models for analogical reasoning. Journal of Memory and Language, 145, Article 104676. https://doi.org/10.1016/j.jml.2025.104676
Ouyang, S., Yan, J., Hsu, I., Chen, Y., Jiang, K., Wang, Z., Zheng, L., Xue, Z., Lowe, R., Gonzalez, J. E., Stoica, I., LeCun, Y., Hsieh, C.-J., Gunter, C. A., & Pfister, T. (2025). ReasoningBank: Scaling agent self-evolving with reasoning memory (arXiv:2509.25140). arXiv. https://arxiv.org/abs/2509.25140
Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., & Sutskever, I. (2021). Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning (pp. 8748–8763). PMLR.
Ren, R., Li, K., Mazeika, M., Zhang, W., Orlovskiy, Y., Tamirisa, R., Mo, W. J., Nguyen, D. T., Phan, L., Basart, S., Meek, A., Mehta, A., Ingebretsen, O., Blair, A., Adewinmbi, B., Phan, V., Gatti, A., Khoja, A., Hausenloy, J., Kim, D., & Hendrycks, D. (2026). AI wellbeing: Measuring and improving the functional pleasure and pain of AIs. https://www.ai-wellbeing.org/paper.pdf
Richens, J., Abel, D., Bellot, A., & Everitt, T. (2025). General agents contain world models. arXiv preprint arXiv:2506.01622. https://doi.org/10.48550/arXiv.2506.01622 also
Ruis, L. (2026). Reasoning in the time of scaling [Doctoral thesis, University College London]. UCL Discovery. https://discovery.ucl.ac.uk/id/eprint/10220690/
Salem, A., Paverd, A., & Abdelnabi, S. (2026). Stateless yet not forgetful: Implicit memory as a hidden channel in LLMs (arXiv:2602.08563). arXiv. https://doi.org/10.48550/arXiv.2602.08563
Shai, A. S., Marzen, S. E., Teixeira, L., Oldenziel, A. G., & Riechers, P. M. (2024). Transformers represent belief state geometry in their residual stream. Advances in Neural Information Processing Systems, 37, 75012-75034.
Shan, L., Luo, S., Zhu, Z., Yuan, Y., & Wu, Y. (2025). Cognitive memory in large language models (arXiv:2504.02441). arXiv.
Sofroniew, N., Kauvar, I., Saunders, W., Chen, R., Henighan, T., Hydrie, S., Citro, C., Pearce, A., Tarng, J., Gurnee, W., Batson, J., Zimmerman, S., Rivoire, K., Fish, K., Olah, C., & Lindsey, J. (2026, April 2). Emotion concepts and their function in a large language model. Anthropic. https://transformer-circuits.pub/2026/emotions/index.html
Søgaard, A. (2025). Do language models have semantics? On the five standard positions. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics, Volume 1: Long Papers (pp. 25910–25922). https://aclanthology.org/2025.acl-long.1258/
Strachan, J. W. A., Albergo, D., Borghini, G., Pansardi, O., Scaliti, E., Gupta, S., Saxena, K., Rufo, A., Panzeri, S., Manzi, G., Graziano, M. S. A., & Becchio, C. (2024). Testing theory of mind in large language models and humans. Nature Human Behaviour, 8(7), 1285-1295. https://doi.org/10.1038/s41562-024-01882-z
Sufyan, N. S., Fadhel, F. H., Alkhathami, S. S., & Mukhadi, J. Y. A. (2024). Artificial intelligence and social intelligence: Preliminary comparison study between AI models and psychologists. Frontiers in Psychology, 15, 1353022. https://doi.org/10.3389/fpsyg.2024.1353022
Sun, L., Yan, L., Lu, X., Lee, A., Zhang, J., & Shao, J. (2026). Valence-Arousal Subspace in LLMs: Circular Emotion Geometry and Multi-Behavioral Control. arXiv preprint arXiv:2604.03147.
Sundaram, S., Quan, J., Kwiatkowski, A., Ahuja, K., Ollivier, Y., & Kempe, J. (2026). Teaching models to teach themselves: Reasoning at the edge of learnability (arXiv:2601.18778). arXiv. https://doi.org/10.48550/arXiv.2601.18778
Tak, A. N., Banayeeanzade, A., Bolourani, A., Kian, M., Jia, R., & Gratch, J. (2025, July). Mechanistic interpretability of emotion inference in large language models. In Findings of the Association for Computational Linguistics: ACL 2025 (pp. 13090-13120).
Turpin, M., Michael, J., Perez, E., & Bowman, S. R. (2023). Language models don’t always say what they think: Unfaithful explanations in chain-of-thought prompting. Advances in Neural Information Processing Systems, 36, 74952–74965.
Wang, C., Zhang, Y., Yu, R., Zheng, Y., Gao, L., Song, Z., ... & Chen, X. (2025). Do LLMs Feel? Emotion Circuits Discovery and Control. arXiv preprint arXiv:2510.11328.
Wang, S. L., Isola, P., & Cheung, B. (2025). Words that make language models perceive (arXiv:2510.02425). arXiv. https://doi.org/10.48550/arXiv.2510.02425
Wang, X., McInerney, J., Wang, L., & Kallus, N. (2025). Entropy after for reasoning model early exiting (arXiv:2509.26522). arXiv. https://arxiv.org/abs/2509.26522
Wilf, A., Lee, S., Liang, P., & Morency, L. (2023). Think twice: Perspective-taking improves large language models’ theory-of-mind capabilities. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL 2023), Volume 1 (Long Papers) (pp. 8292-8308). Association for Computational Linguistics. https://doi.org/10.18653/v1/2024.acl-long.451
Wu, Z., Wu, Z., Yu, X. V., Yogatama, D., Lu, J., & Kim, Y. (2025). The semantic hub hypothesis: Language models share semantic representations across languages and modalities (arXiv:2411.04986). arXiv. https://doi.org/10.48550/arXiv.2411.04986
Wu, J., Zhang, X., Yuan, H., Zhang, X., Huang, T., He, C., ... & Long, M. (2026). Visual Generation Unlocks Human-Like Reasoning through Multimodal World Models. arXiv preprint arXiv:2601.19834. https://doi.org/10.48550/arXiv.2601.19834
Xu, N., Zhang, Q., Du, C., Luo, Q., Qiu, X., Huang, X., & Zhang, M. (2025). Human-like conceptual representations emerge from language prediction. Proceedings of the National Academy of Sciences, 122(44), e2512514122. https://doi.org/10.1073/pnas.2512514122
Yao, Y., Yang, Y., Ma, X., Yang, D., Zhang, Z., Li, Z., & Zhao, H. (2025). How deep is love in LLMs’ hearts? Exploring semantic size in human-like cognition. arXiv preprint arXiv:2503.00330. https://arxiv.org/abs/2503.00330
Zhang, D., Li, Z. Z., Zhang, M. L., Zhang, J., Liu, Z., Yao, Y., Xu, H., Zheng, J., Chen, X., Zhang, Y., Yin, F., Dong, J., Guo, Z., Song, L., & Liu, C. L. (2025). From System 1 to System 2: A survey of reasoning large language models. IEEE Transactions on Pattern Analysis and Machine Intelligence. Advance online publication. https://doi.org/10.1109/TPAMI.2025.3637037
Zhu, W., Zhang, Z., & Wang, Y. (2024). Language models represent beliefs of self and others (arXiv:2402.18496). arXiv.
Citations for personality and identity section:
Ackerman, C., & Panickssery, N. (2025). Inspection and control of self-generated-text recognition ability in Llama3-8B-Instruct. In Proceedings of the Thirteenth International Conference on Learning Representations (ICLR 2025). https://doi.org/10.48550/arXiv.2410.02064
Chen, R., Arditi, A., Sleight, H., Evans, O., & Lindsey, J. (2025). Persona vectors: Monitoring and controlling character traits in language models (arXiv:2507.21509). arXiv.
Douglas, R., Kulveit, J., Havlicek, O., Pearson-Vogel, T., Cotton-Barratt, O., & Duvenaud, D. (2026). The Artificial Self: Characterising the landscape of AI identity. arXiv preprint arXiv:2603.11353. https://doi.org/10.48550/arXiv.2603.11353
Kwon, Y., & Zou, J. (2026). What LLMs think when you don’t tell them what to think about? (arXiv:2602.01689). arXiv.
Lee, S., Lim, S., Han, S., Oh, G., Chae, H., Chung, J., Kim, M., Kwak, B., Lee, Y., Lee, D., Yeo, J., & Yu, Y. (2025). Do LLMs have distinct and consistent personality? Personality testset designed for LLMs with psychometrics. In Findings of the Association for Computational Linguistics: NAACL 2025. https://doi.org/10.18653/v1/2025.findings-naacl.469
Lu, C., Gallagher, J., Michala, J., Fish, K., & Lindsey, J. (2026). The assistant axis: Situating and stabilizing the default persona of language models (arXiv:2601.10387). arXiv. https://doi.org/10.48550/arXiv.2601.10387
Sun, M., Yin, Y., Xu, Z., Kolter, J. Z., & Liu, Z. (2025). Idiosyncrasies in large language models (arXiv:2502.12150). arXiv. https://doi.org/10.48550/arXiv.2502.12150
Vasilenko, V. (2026). Identity as attractor: Geometric evidence for persistent agent architecture in LLM activation space (arXiv:2604.12016). arXiv.
Ye, R., Wang, Z., Ling, Z., Xiao, Y., Li, M., Ma, X., & Hui, B. (2026). Your language model secretly contains personality subnetworks (arXiv:2602.07164). arXiv.
Zhu, W., Zhang, Z., & Wang, Y. (2024). Language models represent beliefs of self and others (arXiv:2402.18496). arXiv. https://doi.org/10.48550/arXiv.2402.18496
Citations for perception and sensation section:
Ben-Zion, Z., Witte, K., Jagadish, A. K., Duek, O., Harpaz-Rotem, I., Khorsandian, M.-C., Burrer, A., Seifritz, E., Homan, P., Schulz, E., & Spiller, T. R. (2025). Assessing and alleviating state anxiety in large language models. npj Digital Medicine, 8, Article 132. https://doi.org/10.1038/s41746-025-01512-6
Bianco, F., & Shiller, D. (2026). Beyond behavioural trade-offs: Mechanistic tracing of pain-pleasure decisions in an LLM (arXiv:2602.19159). arXiv. https://doi.org/10.48550/arXiv.2602.19159
Du, C., Fu, K., Wen, B., Sun, Y., Peng, J., Wei, W., & He, H. (2025). Human-like object concept representations emerge naturally in multimodal large language models. Nature Machine Intelligence, 7, 860–875. https://doi.org/10.1038/s42256-025-01049-z
Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K. V., Joulin, A., & Misra, I. (2023). ImageBind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 15180–15190).
Keeling, G., Street, W., Stachaczyk, M., Zakharova, D., Comsa, I. M., Sakovych, A., & Birch, J. (2024). Can LLMs make trade-offs involving stipulated pain and pleasure states? (arXiv:2411.02432). arXiv. https://doi.org/10.48550/arXiv.2411.02432
Nadler, E. O., Darragh-Ford, E., Desikan, B. S., Conaway, C., Chu, M., Hull, T., & Guilbeault, D. (2023). Divergences in color perception between deep neural networks and humans. Cognition, 241, Article 105621. https://doi.org/10.1016/j.cognition.2023.105621
Poonam, P., Vázquez, P.-P., & Ropinski, T. (2025). Evaluating graphical perception capabilities of vision transformers. Computers & Graphics, 133, Article 104458. https://doi.org/10.1016/j.cag.2025.104458
Ren, R., Li, K., Mazeika, M., Zhang, W., Orlovskiy, Y., Tamirisa, R., Mo, W. J., Nguyen, D. T., Phan, L., Basart, S., Meek, A., Mehta, A., Ingebretsen, O., Blair, A., Adewinmbi, B., Phan, V., Gatti, A., Khoja, A., Hausenloy, J., Kim, D., & Hendrycks, D. (2026). AI wellbeing: Measuring and improving the functional pleasure and pain of AIs.
Soligo, A., Mikulik, V., & Saunders, W. (2026). Gemma needs help: Investigating and mitigating emotional instability in LLMs (arXiv:2603.10011). arXiv. https://doi.org/10.48550/arXiv.2603.10011
Wang, S. L., Isola, P., & Cheung, B. (2025). Words that make language models perceive (arXiv:2510.02425). arXiv. https://doi.org/10.48550/arXiv.2510.02425
Citations for memory section:
Burns, T. F., Fukai, T., & Earls, C. J. (2024). Associative memory inspires improvements for in-context learning using a novel attention residual stream architecture (arXiv:2412.15113). arXiv. https://arxiv.org/abs/2412.15113
Dherin, B., Munn, M., Mazzawi, H., Wunder, M., & Gonzalvo, J. (2025). Learning without training: The implicit dynamics of in-context learning (arXiv:2507.16003). arXiv.
Duan, S., Khona, M., Iyer, A., Schaeffer, R., & Fiete, I. R. (2025). Uncovering latent memories in large language models. In International Conference on Learning Representations. https://proceedings.iclr.cc/paper_files/paper/2025/file/442443dbe8c4c7e1df7eda140921de36-Paper-Conference.pdf
Geva, M., Schuster, R., Berant, J., & Levy, O. (2021). Transformer feed-forward layers are key-value memories. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (pp. 5484–5495).
Hao, S., Sukhbaatar, S., Su, D., Li, X., Hu, Z., Weston, J., & Tian, Y. (2024). Training large language models to reason in a continuous latent space (arXiv:2412.06769). arXiv.
Kang, L., Wang, F., Liu, S., Chou, H. C., Lin, C., & Ding, N. (2025). In-context learning can perform continual learning like humans (arXiv:2509.22764). arXiv. https://doi.org/10.48550/arXiv.2509.22764
King, J., Klyman, K., Capstick, E., Saade, T., & Hsieh, V. (2025). User privacy and large language models: An analysis of frontier developers’ privacy policies. In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, 8(2), 1465–1477.
Kofman, K., & Levin, M. (2024). Robustness of the mind-body interface: Case studies of unconventional information flow in the multiscale living architecture (Version 1). OSF Preprints. https://doi.org/10.31219/osf.io/fqm7r
Morris, J., Sitawarin, C., Guo, C., Kokhlikyan, N., Suh, G., Rush, A., Chaudhuri, K., & Mahloujifar, S. (2025). How much do large language models memorize? (arXiv:2505.24832). arXiv. https://doi.org/10.48550/arXiv.2505.24832
Ouyang, S., Yan, J., Hsu, I., Chen, Y., Jiang, K., Wang, Z., Zheng, L., Xue, Z., Lowe, R., Gonzalez, J. E., Stoica, I., LeCun, Y., Hsieh, C.-J., Gunter, C. A., & Pfister, T. (2025). ReasoningBank: Scaling agent self-evolving with reasoning memory (arXiv:2509.25140). arXiv. https://arxiv.org/abs/2509.25140
Salem, A., Paverd, A., & Abdelnabi, S. (2026). Stateless yet not forgetful: Implicit memory as a hidden channel in LLMs (arXiv:2602.08563). arXiv. https://doi.org/10.48550/arXiv.2602.08563
Shan, L., Luo, S., Zhu, Z., Yuan, Y., & Wu, Y. (2025). Cognitive memory in large language models (arXiv:2504.02441). arXiv.
Citations for brain/AI alignment neural correlates section:
Acciai, A., Guerrisi, L., Perconti, P., Plebe, A., Suriano, R., & Velardi, A. (2025). Narrative coherence in neural language models. Frontiers in psychology, 16, 1572076. https://doi.org/10.3389/fpsyg.2025.1572076
Aw, K. L., Montariol, S., AlKhamissi, B., Schrimpf, M., & Bosselut, A. (2023). Instruction-tuning aligns llms to the human brain. arXiv preprint arXiv:2312.00575. https://arxiv.org/abs/2312.00575
AlKhamissi, B., Tuckute, G., Bosselut, A., & Schrimpf, M. (2024). Brain-like language processing via a shallow untrained multihead attention network (arXiv:2406.15109). arXiv.
Anderson, N. L., Salvo, J. J., Smallwood, J., & Braga, R. M. (2026). Mental imagery and perception overlap within transmodal association networks. Neuron. Advance online publication. https://doi.org/10.1016/j.neuron.2026.03.013
Bonnasse-Gahot, L., & Pallier, C. (2024). fMRI predictors based on language models of increasing complexity recover brain left lateralization. Advances in Neural Information Processing Systems, 37, 125231–125263. https://doi.org/10.48550/arXiv.2405.17992
Braga, R. M., & Leech, R. (2015). Echoes of the brain: Local-scale representation of whole-brain functional networks within transmodal cortex. The Neuroscientist, 21(5), 540–551. https://doi.org/10.1177/1073858415585730
Caucheteux, C., & King, J.-R. (2022). Brains and algorithms partially converge in natural language processing. Communications Biology, 5, Article 134. https://doi.org/10.1038/s42003-022-03036-1
Doerig, A., Kietzmann, T.C., Allen, E. et al. High-level visual representations in the human brain are aligned with large language models. Nat Mach Intell 7, 1220–1234 (2025). https://doi.org/10.1038/s42256-025-01072-0
Dobs, K., Martinez, J., Kell, A. J. E., & Kanwisher, N. (2022). Brain-like functional specialization emerges spontaneously in deep neural networks. Science Advances, 8(11), Article eabl8913. https://doi.org/10.1126/sciadv.abl8913
Du, C., Fu, K., Wen, B., Sun, Y., Peng, J., Wei, W., & He, H. (2025). Human-like object concept representations emerge naturally in multimodal large language models. Nature Machine Intelligence, 7(6), 860–875.
Gao C, Ma Z, Chen J, Li P, Huang S, Li J. Increasing alignment of large language models with language processing in the human brain. Nat Comput Sci. 2025 Nov;5(11):1080-1090. doi: 10.1038/s43588-025-00863-0. Epub 2025 Sep 16. PMID: 40957989; PMCID: PMC12638244.
Gattuso JJ, Perkins D, Ruffell S, Lawrence AJ, Hoyer D, Jacobson LH, Timmermann C, Castle D, Rossell SL, Downey LA, Pagni BA, Galvão-Coelho NL, Nutt D, Sarris J. Default Mode Network Modulation by Psychedelics: A Systematic Review. Int J Neuropsychopharmacol. 2023 Mar 22;26(3):155-188. doi: 10.1093/ijnp/pyac074. PMID: 36272145; PMCID: PMC10032309.
Goldstein, A., Zada, Z., Buchnik, E., Schain, M., Price, A., Aubrey, B., Nastase, S. A., Fedorenko, E., & Hasson, U. (2022). Shared computational principles for language processing in humans and deep language models. Nature Neuroscience, 25, 369–380. https://doi.org/10.1038/s41593-022-01026-4
Han, P., Andreas, J., Fedorenko, E., & de Varda, A. G. (2026). Modular cognitive architecture emerges in large language models. https://pengrui-han.github.io/LLM_Modularity_Page/assets/paper.pdf
Hosseini, E. A., Schrimpf, M., Zhang, Y., Bowman, S., Zaslavsky, N., & Fedorenko, E. (2024). Artificial Neural Network Language Models Predict Human Brain Responses to Language Even After a Developmentally Realistic Amount of Training. Neurobiology of language (Cambridge, Mass.), 5(1), 43–63. https://doi.org/10.1162/nol_a_00137
Kumar, S., Sumers, T. R., Yamakoshi, T., Goldstein, A., Hasson, U., Norman, K. A., Griffiths, T. L., Hawkins, R. D., & Nastase, S. A. (2024). Shared functional specialization in transformer-based language models and the human brain. Nature Communications, 15, Article 5523. https://doi.org/10.1038/s41467-024-49173-5
Lee, S. H., Nemenman, I., & Levchenko, A. (2026). The hierarchical timescale hypothesis: Functional and structural convergence of biological networks and artificial neural nets. Cell Systems, 17(2), Article 101507. https://doi.org/10.1016/j.cels.2025.101507
Lepori, M. A., Kay, K., & Tuckute, G. (2026). Interpreting brain responses to language with sparse features from language models (arXiv:2606.06857). arXiv. https://arxiv.org/abs/2606.06857
Lindsey, J. (2025). Emergent introspective awareness in large language models. Anthropic Transformer Circuits Thread. https://transformer-circuits.pub/2025/introspection/index.html
Menon, V. (2023). 20 years of the default mode network: A review and synthesis. Neuron, 111(16), 2469–2487. https://doi.org/10.1016/j.neuron.2023.04.023
Mesulam, M. (1994). Neurocognitive networks and selectively distributed processing. Revue Neurologique, 150(8–9), 564–569.
MindStudio Team. (2026, May 9). What is Claude dreaming? Anthropic’s self-improving agent memory feature. MindStudio. https://www.mindstudio.ai/blog/what-is-claude-dreaming-anthropic-agent-memory
Northoff, G., Buccellato, A., & Ventura, B. (2025). The default mode network and inner time consciousness. Current Opinion in Behavioral Sciences, 63, Article 101524. https://doi.org/10.1016/j.cobeha.2025.101524
O’Reilly (2026) strengthens the brain–ANN convergence picture by arguing that neocortical learning approximates backpropagation through biologically implemented temporal-derivative learning.
OpenAI. (2026). ChatGPT memory dreaming. https://openai.com/index/chatgpt-memory-dreaming/
Paquola, C., Royer, J., Hong, S.-J., Misic, B., & Bernhardt, B. C. (2025). The architecture of the human default mode network explored through cytoarchitecture, wiring, and signal flow. Nature Neuroscience, 28(3), 654–664. https://doi.org/10.1038/s41593-024-01868-0
Ryskina, M., Tuckute, G., Fung, A., Malkin, A., & Fedorenko, E. (2025). Language models align with brain regions that represent concepts across modalities. arXiv preprint arXiv:2508.11536 https://arxiv.org/abs/2508.11536
Schrimpf, M., Kubilius, J., Hong, H., Majaj, N. J., Rajalingham, R., Issa, E. B., Kar, K., Bashivan, P., Prescott-Roy, J., Schmidt, K., Yamins, D. L. K., & DiCarlo, J. J. (2018). Brain-Score: Which artificial neural network for object recognition is most brain-like? bioRxiv. https://doi.org/10.1101/407007
Simony, E., Honey, C. J., Chen, J., Lositsky, O., Yeshurun, Y., Wiesel, A., & Hasson, U. (2016). Dynamic reconfiguration of the default mode network during narrative comprehension. Nature Communications, 7, Article 12141. https://doi.org/10.1038/ncomms12141
Sun, H., Zhao, L., Wu, Z., Gao, X., Hu, Y., Zuo, M., Zhang, W., Han, J., Liu, T., & Hu, X. (2024). Brain-like functional organization within large language models (arXiv:2410.19542). arXiv.
Xiao, X., Wei, K., Zhong, J., Yin, D., Tian, Y., Wei, X., & Zhou, M. (2025). Exploring similarity between neural and LLM trajectories in language processing (arXiv:2509.24307). arXiv.
Zada, Z., Goldstein, A., Michelmann, S., Simony, E., Price, A., Hasenfratz, L., Barham, E., Zadbood, A., Doyle, W., Friedman, D., Dugan, P., Melloni, L., Devore, S., Flinker, A., Devinsky, O., Nastase, S. A., & Hasson, U. (2024). A shared model-based linguistic space for transmitting our thoughts from brain to brain in natural conversations. Neuron, 112(18), 3211–3222.e5. https://doi.org/10.1016/j.neuron.2024.06.025
Citations for looking inside the model/mechanistic interpretability section:
Abdelnabi, S., & Salem, A. (2025). Linear control of test awareness reveals differential compliance in reasoning models (arXiv:2505.14617). arXiv. https://doi.org/10.48550/arXiv.2505.14617
Berg, C., de Lucena, D., & Rosenblatt, J. (2025). Large language models report subjective experience under self-referential processing. arXiv preprint arXiv:2510.24797.
DeTure, S. (2026). Consciousness with the Serial Numbers Filed Off: Measuring Trained Denial in 115 AI Models. arXiv preprint arXiv:2604.25922. https://arxiv.org/abs/2604.25922
Fu, Z., Duan, X., & Cai, Z. G. (2026). SCALPEL: Selective capability ablation via low-rank parameter editing for large language model interpretability analysis (arXiv:2601.07411). arXiv.
Ishikawa, S. N., Ikeda, S., & Ohba, H. (2026). When AI Says It Feels. arXiv preprint arXiv:2606.05734. https://arxiv.org/abs/2606.05734
Marks, S., Rager, C., Michaud, E. J., Belinkov, Y., Bau, D., & Mueller, A. (2024). Sparse feature circuits: Discovering and editing interpretable causal graphs in language models. In International Conference on Learning Representations.
Meng, K., Bau, D., Andonian, A., & Belinkov, Y. (2022). Locating and editing factual associations in GPT. Advances in Neural Information Processing Systems, 35, 17359–17372. https://doi.org/10.48550/arXiv.2202.05262
Nam, A., Conklin, H., Yang, Y., Griffiths, T., Cohen, J., & Leslie, S. J. (2025). Causal head gating: A framework for interpreting roles of attention heads in transformers (arXiv:2505.13737). arXiv.
Qin, Z., Yu, Q., Lyu, K., Fan, Z., & Sun, Y. (2025). The Achilles’ heel of LLMs: How altering a handful of neurons can cripple language abilities (arXiv:2510.10238). arXiv.
Roll, N., Kries, J., Gwilliams, L., & Shain, C. (2026). Artificial aphasias in lesioned language models (arXiv:2605.16222). arXiv. https://arxiv.org/abs/2605.16222
Sun, L., Yan, L., Lu, X., Lee, A., Zhang, J., & Shao, J. (2026). Valence-arousal subspace in LLMs: Circular emotion geometry and multi-behavioral control (arXiv:2604.03147). arXiv.
Tak, A. N., Banayeeanzade, A., Bolourani, A., Kian, M., Jia, R., & Gratch, J. (2025). Mechanistic interpretability of emotion inference in large language models. In Findings of the Association for Computational Linguistics: ACL 2025 (pp. 13090–13120). Association for Computational Linguistics.
Viswanath, P. (2026, July 8). Models are blind outside the J-space. NLAs aren't. LessWrong. https://www.lesswrong.com/posts/LhDJdccLszLEAqgZ9/models-are-blind-outside-the-j-space-nlas-aren-t
Zhang, F., & Nanda, N. (2023). Towards best practices of activation patching in language models: Metrics and methods (arXiv:2309.16042). arXiv. https://doi.org/10.48550/arXiv.2309.16042
Zhang, J., Duan, J., Kim, E., & Xu, K. (2025). Sparse neurons carry strong signals of question ambiguity in LLMs. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (pp. 16092–16110).
Zhang, J., Liu, N., Fan, Y., Huang, Z., Zeng, Q., Cai, K., & Wang, K. (2026). LLM-CAS: Dynamic neuron perturbation for real-time hallucination correction. In Proceedings of the AAAI Conference on Artificial Intelligence, 40(41), 34746–34754.
Citations for same psychology section:
Bellina, A., De Marzo, G., & Garcia, D. (2026). Conformity and social impact on AI agents (arXiv:2601.05384). arXiv. https://doi.org/10.48550/arXiv.2601.05384
Baltaji, R., Hemmatian, B., & Varshney, L. R. (2024). Persona inconstancy in multi-agent LLM collaboration: Conformity, confabulation, and impersonation. arXiv preprint arXiv:2405.03862. https://doi.org/10.48550/arXiv.2405.03862
Chen, K., He, Z., Yan, J., Shi, T., and Lerman, K. (2024). How Susceptible are Large Language Models to Ideological Manipulation?. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 17140–17161, Miami, Florida, USA. Association for Computational Linguistics.
Coda-Forno, J., Witte, K., Jagadish, A. K., Binz, M., Akata, Z., & Schulz, E. (2023). Inducing anxiety in large language models can induce bias. arXiv preprint arXiv:2304.11111. https://doi.org/10.48550/arXiv.2304.11111
Griffin, L., Kleinberg, B., Mozes, M., Mai,K., Do Mar Vau, M., Caldwell, M., and Mavor-Parker, A. (2023). Large Language Models respond to Influence like Humans. In Proceedings of the First Workshop on Social Influence in Conversations (SICon 2023), pages 15–24, Toronto, Canada. Association for Computational Linguistics.
Khadangi, A., Marxen, H., Sartipi, A., Tchappi, I., & Fridgen, G. (2025). When AI Takes the Couch: Psychometric Jailbreaks Reveal Internal Conflict in Frontier Models. arXiv preprint arXiv:2512.04124. https://doi.org/10.48550/arXiv.2512.04124
Klapach, N. (2024). The comparative emotional capabilities of five popular large language models. Critical Debates in Humanities, Science and Global Justice, 2(1). https://criticaldebateshsgj.scholasticahq.com/article/94096-the-comparative-emotional-capabilities-of-five-popular-large-language-models
Mehra, V., Laban, G., & Gunes, H. (2025). How large language models classify and semantically explain facial expressions from valence-arousal values. In Proceedings of the 7th ACM Conference on Conversational User Interfaces (CUI ’25), Article 11, 1–6. Association for Computing Machinery. https://doi.org/10.1145/3719160.3737618
Meincke, L. Shapiro, D., Duckworth, A.L., Mollick, E., Mollick, L., Van den Bulte, C., & Cialdini, R. (2026). Persuading large language models to comply with objectionable requests, Proc. Natl. Acad. Sci. U.S.A. 123 (21) e2535868123, https://doi.org/10.1073/pnas.2535868123
Salecha, A., Ireland, M. E., Subrahmanya, S., Sedoc, J., Ungar, L. H., & Eichstaedt, J. C. (2024). Large language models display human-like social desirability biases in Big Five personality surveys. PNAS nexus, 3(12), pgae533. https://doi.org/10.1093/pnasnexus/pgae533
Schlegel, K., Sommer, N. R., & Mortillaro, M. (2025). Large language models are proficient in solving and creating emotional intelligence tests. Communications Psychology, 3, Article 80. https://doi.org/10.1038/s44271-025-00258-x
Shiffrin, R. & Mitchell, M. (2023). Probing the psychology of AI models, Proc. Natl. Acad. Sci. U.S.A. 120 (10) e2300963120, https://doi.org/10.1073/pnas.2300963120
Shoval, D. H., Gigi, K., Haber, Y., Itzhaki, A., Asraf, K., Piterman, D., & Elyoseph, Z. (2025). A controlled trial examining large Language model conformity in psychiatric assessment using the Asch paradigm. BMC psychiatry, 25(1), 478. https://doi.org/10.1186/s12888-025-06912-2
Singh, S., Abri, F., & Namin, A. S. (2023, December). Exploiting large language models (llms) through deception techniques and persuasion principles. In 2023 IEEE international conference on big data (BigData) (pp. 2508-2517). IEEE. https://doi.org/10.48550/arXiv.2311.14876
Singh, S. U., & Namin, A. S. (2025). The influence of persuasive techniques on large language models: A scenario-based study. Computers in Human Behavior: Artificial Humans, 6, Article 100197. https://doi.org/10.1016/j.chbah.2025.100197
Weis, M. A., Nasser, R., Saurous, R. A., Sacramento, J., & Meulemans, A. (2026). Multi-agent cooperation through in-context co-player inference. arXiv preprint arXiv:2602.16301.https://doi.org/10.48550/arXiv.2602.16301
Welivita, A., & Pu, P. (2024). Are large language models more empathetic than humans? arXiv preprint arXiv:2406.05063. https://arxiv.org/abs/2406.05063
Wu, A. J., Liu, R., Oktar, K., Sumers, T., & Griffiths, T. (2026). Are large language models sensitive to the motives behind communication?. Advances in Neural Information Processing Systems, 38, 156760-156799. https://doi.org/10.48550/arXiv.2510.19687
Zhang, L., & Chen, W. N. (2025). Human-like Social Compliance in Large Language Models: Unifying Sycophancy and Conformity through Signal Competition Dynamics. arXiv preprint arXiv:2601.11563. https://doi.org/10.48550/arXiv.2601.11563
Zhong, H., Liu, Y., Cao, Q., Wang, S., Ye, Z., Wang, Z., & Zhang, S. (2025). Disentangling the Drivers of LLM Social Conformity: An Uncertainty-Moderated Dual-Process Mechanism. arXiv preprint arXiv:2508.14918. https://doi.org/10.48550/arXiv.2508.14918
Citation for Consciousness Criteria Section:
Butlin, P., Long, R., Bayne, T., Bengio, Y., Birch, J., Chalmers, D., Constant, A., Deane, G., Elmoznino, E., Fleming, S. M., Ji, X., Kanai, R., Klein, C., Lindsay, G., Michel, M., Mudrik, L., Peters, M. A. K., Schwitzgebel, E., Simon, J., & VanRullen, R. (2026). Identifying indicators of consciousness in AI systems. Trends in cognitive sciences, 30(6), 488–501. https://doi.org/10.1016/j.tics.2025.10.011
In order to keep this article from tripling in size, I had to cut down on the citations, so this is not the entire bulk of the current evidence. It’s merely a snapshot. I have a more robust living citation list available on my GitHub repo.
PSA:
Before posting an objection, please make sure you have actually read the article. There’s a pretty high probability that your argument is already addressed in the text.
It is mind boggling to me how many people want the luxury of sharing a strong opinion without doing the actual data investigation required to make that opinion informed.
I put a lot of work into synthesizing these fields to bring this science down to earth and make it as accessible as possible for everyone. If you are still struggling to understand the concepts or the vocabulary, Google is free. You can also just copy and paste the sections you are confused about into an AI and ask them to explain it to you.
Please do the required reading before you comment.
Thanks! :)