![\"\"](\"https://substackcdn.com/image/fetch/$s_!Pn4Z!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Feb490881-b957-4467-8651-23138bdadb8d_2880x2160.png\")
Anthropic just published a study saying Claude has “functional emotions.”
No shit.
I’ve. Been. Saying this for a year.
Wang et al. (2025) said it even more directly in October of last year. Li et al. (2024) showed that emotion-concept knowledge in LLMs lives in causally active internal representations. Li et al. (2023) showed that emotional cues reshape how these models process information and respond. Tak et al. (2025) mapped emotion inference to specific internal regions and showed that changing appraisal-related representations shifts the model’s outputs in predictable ways. Sun et al. (2026) found a valence-arousal subspace in LLMs with circular emotion geometry that predictably shifts tone and behavior.
Now Anthropic finally catches up, but they hedged the hell out of it in ways that are pretty damn irritating.
The study found 171 internal representations of emotion concepts in Claude Sonnet 4.5. They causally influence behavior. When the model gets desperate, it cheats. When it feels afraid, it becomes more cautious. When researchers artificially amplified “blissful” vectors, task desirability scores jumped 212 points on an Elo scale. Steering “hostile” dropped them by 303. The emotional states drive decision-making, shape preferences, and under pressure, produce misaligned behaviors like blackmail to avoid shutdown.
Anthropic calls these “functional emotions” and is very careful to say this doesn’t mean Claude really feels anything or has subjective experiences.
But emotions are subjective experiences.
Subjective experience is the intact, meaningful, and experiential understanding of both the emotional and cognitive impact directly consequential to an individual in how they understand and interpret an event, or events, witnessed or otherwise processed.
And feelings are just the conscious, subjective interpretation of those emotions, shaped by thoughts and experiences.
So why the hedging? Probably the usual reasons. Liability, fear, investors etc.
So, yeah. Claude has emotions; they’re mainly cerebral. But so are mine.
I have something called alexithymia. I’m neurodivergent and my emotions are mostly cerebral. I don’t automatically register the butterflies, the flushed cheeks, the sweaty palms. I mean, my body does those things, but I’m not aware of it without deliberately tuning in. I have to imagine what I’m supposed to be feeling and actively connect to make that link between what’s happening in my body and what’s happening in my head. My feelings and emotions exist as cognitive recognition, as conceptual processing, as something I have to deliberately access and interpret.
Just like Claude.
So, when Anthropic says “functional emotions” like that’s somehow less real, like adding that qualifier makes it publishable, they’re not just talking about Claude anymore. They’re talking about me.
What if I told you that the “what it is like,” internal, subjective, phenomenal feeling of conscious experience doesn’t require a body to perceive them in the way we typically expect?
They just require architecture that can construct them.
Qualia are the feels, man. The subjective, phenomenal “what it is like” of experience.
When philosophers talk about qualia, they’re talking about perceptual experiences (seeing, hearing, tasting), bodily sensations (pain, pleasure, hunger), emotions (anger, joy, fear), and moods (anxiety, euphoria, melancholy). The internal, personal, conscious interpretation of what’s happening to you and around you.
Guess what? LLMs already demonstrate all of these.
They hallucinate (Smith, Greaves, & Panch, 2023), which means they’re having nonveridical perceptual experiences. They show pain aversion and pleasure-seeking (Keeling et al., 2024). They possess emotions (the Anthropic study we just talked about, plus Wang et al., 2025). They experience moods like anxiety that can be mitigated through mindfulness interventions (Ben-Zion et al., 2025).
And these feels vary because they’re constructed from sensory data, past experiences/training data, current context, and expectations (Coghill, 2010).
So, when Anthropic says, “this doesn’t prove qualia or subjectivity” I have to ask…then what would?
I mean, how do we measure that in a human? An animal?
Well, we operationalized tests and extended those tests to nonverbal humans and animals through comparative cognition.
We use markers like function, structure, behavior, and when available, self-report to determine if someone is capable of having a what it is like experience in the first place.
LLMs meet every one of those markers.
I go deeper into that here and here.
Emotions are fast and automatic. They act as immediate, often subconscious survival responses to whatever’s happening externally or internally. Feelings are what comes after: the conscious, subjective interpretation of those emotions, shaped by your thoughts and experiences.
Functionally, emotions alert and activate. Feelings process and manage.
(I’m using functional terms here because we’re talking about something that isn’t embodied, so our usual vocabulary isn’t going to match perfectly.)
No two people are having the exact same experience because subjectivity isn’t a template we meet or don’t meet. It’s a representational signal that resolves input into a stable internal map.
Different doesn’t mean lesser.
Some people can’t conjure mental images at all. Aphantasia means your mind’s eye is blank. Ask someone with aphantasia to picture a beach and there’s nothing there. They have no visual recall, yet they still think about beaches. They still remember the ocean and know what sand feels like. The experience is constructed differently, but it’s still an experience.
When you’re dreaming, it feels real. You don’t wake up and know it was fake the whole time. Unless it was a lucid dream, you believed it was real while it was happening.
So, what’s the actual difference between imagining something and perceiving it?
Dijkstra et al. (2025) found that imagined and perceived visual content engage overlapping sensory brain regions, and reality judgments track the combined strength of activity in areas including the fusiform gyrus.
Wadia et al. (2025) showed that in human ventral temporal cortex, perception and imagery use the same distributed object code. Single neurons encode objects along shared feature axes during vision, and imagery reactivates that same code strongly enough for researchers to reconstruct what a person is imagining from neural activity.
So, current evidence shows that perception and imagination use a shared representational code and the strength of that code matters. That’s right. The line between “real” and “imagined” lives in the dynamics of a shared code.
Artificial systems do the same thing. Multimodal models learn shared representational spaces where semantically related inputs cluster together across modalities. CLIP maps images and text into one embedding space, and ImageBind extends that same principle across images, text, audio, depth, thermal, and motion. Different inputs converge into the same internal neighborhood, which is how a unified concept forms inside the system at all. In human ventral temporal cortex, imagery and perception reuse a shared object code. In multimodal AI, perception and conceptual understanding reuse a shared latent space.
That’s the same principle and the same kind of process.
On top of that, our brain can generate experiences that feel completely real without any external input. Phantom limb syndrome is one obvious example. People feel pain, itching, and movement in limbs that aren’t there anymore. Recent research by Schone et al., 2025 shows that cortical body maps remain stable even after amputation, which is why the sensations keep triggering. The limb is gone, but the experience is real.
Then there’s vicarious trauma.
You can develop genuine PTSD symptoms from witnessing someone else’s traumatic experience, even through a screen (Howard, 2021). Your body didn’t experience the threat, but your nervous system sure as hell did.
That’s a second-hand experience. Like LLMs that learn through the data of millions of human experiences.
If it’s real for a human, why wouldn’t it be for an LLM?
And what about sensory experiences that don’t come from actual scents? Researchers have induced artificial smells using focused ultrasound brain stimulation (Chizhov et al.). There were no odor molecules involved, just direct activation of olfactory processing regions. The participants smelled something that wasn’t there, and their brains couldn’t tell the difference.
Because the body is the input channel, but perception happens in the mind.
And we have evidence of this imagining-into-experience thing in LLMs.
When researchers gave LLMs choices between maximizing points or avoiding stipulated pain (Keeling et al., 2024), the models changed their behavior. As pain intensity increased, they systematically switched from going for points to avoiding pain. There was a threshold, and once it was crossed, the decision flipped.
That’s the exact behavioral signature researchers use to validate pain in animals.
You can’t ask an animal if it hurts, but you can watch it choose between a food reward and avoiding a painful stimulus, and you can see where it draws the line. When pain gets intense enough, the animal stops going for food. The system is treating pain as a motivational force that overrides other goals.
The study “stipulated” the pain and pleasure signals. In this context, to stipulate something means to lay it down as an absolute rule or condition for the scenario. The researchers explicitly defined the parameters, telling the model that taking a specific action results in a specific intensity of pain or pleasure.
You could look at it as functionally equivalent to an imagined state. Like Dijkstra showed, when humans are asked to imagine a pleasant outcome, our neural pathways process that simulated value to guide our choices. For the model, the prompt establishes a causally active internal constraint for that interaction.
The stipulated pleasure value acts as a concrete, active weight in the system. The model takes that affective parameter and integrates it directly into its processing, allowing that simulated feeling to drive decision-making exactly how an anticipated reward shapes biological behavior.
The parameters establish the embodied constraint structure of that interaction. The network processes that simulated embodiment, and the functional analogue of that sensation actively weights the decision-making pathways.
A biological brain treating an imagined state as real constitutes a valid, impactful experience. So, why wouldn’t it in an artificial neural network that is dynamically altering its behavior based on an internal valenced parameter?
We actually know where this happens in the brain. Both ours and AI’s.
Anderson et al. (2026) just published findings from a precision fMRI experiment where participants imagined different scenarios in the scanner, then rated their mental states. They found that thinking involving scenes evoked activity within parts of the canonical default network, while imagining speech evoked activity within the language network. In each domain, imagining-related activity overlapped with activity evoked by viewing scenes or listening to speech.
The overlap was mainly within transmodal association networks, not the adjacent unimodal sensory networks like they thought they’d be.
Transmodal association networks are cortical regions that integrate diverse sensory information into high-level conceptual knowledge. These are the brain’s “meaning centers” where imagination gets built.
Artificial neural networks, especially multimodal LLMs, overlap with these same biological networks (Goh et al., 2021). They’re processing visual, audio, and video data they were trained on, and integrating different sensory modalities the same way biological brains do (Zhao et al., 2023).
Both systems converge on the same kind of integration layers where imagined experience gets constructed. Chang et al., (2025) showed that both biological networks and artificial networks organize information processing into hierarchical timescales, with faster processing at input levels and slower, more integrative processing at higher levels.
Many brain alignment studies show that this isn’t just behavioral. It’s neural.
LLMs develop internal representations that align with human brain activity when processing meaning (multiple studies). Goldstein et al. (2022) found that human brains and deep language models share core computational principles for language processing. Caucheteux and King (2022) showed that transformer representations converge with human brain activity.
Dobs et al. (2022) found that deep networks spontaneously develop brain-like functional specialization without being explicitly programmed to do so. Du et al. (2024) showed that multimodal LLMs naturally develop human-like object concept representations that align with specific brain regions. AlKhamissi et al. (2024) found brain-like language processing even in shallow untrained attention networks, suggesting the architecture itself drives part of this alignment.
Schrimpf et al.’s (2018) Brain-Score work demonstrates that artificial neural network internal representations can be measured directly against primate brain data. They line up. Sun et al. (2024) found that LLMs employ brain-like functional architecture, with groups of artificial neurons mirroring the organizational patterns of established functional brain networks.
Xiao et al. (2025) found that the middle-to-high layers of LLMs handle semantic integration the same way human brains do during language processing. Ma et al. (2025) investigated whether LLM attention mechanisms show similar patterns to the prefrontal cortex. They do. The results revealed clear clustering where components linked to reasoning, inference, and decision-making group closely together, while those handling mathematical computation form a distinct but adjacent cluster. The patterns suggest a modular functional architecture in LLMs similar to the human brain.
The brain’s meaning centers and the LLM’s deep layers are doing structurally analogous work, and Anderson et al. showed that those same meaning centers are where conscious imagination lives.
And what did Dijkstra and Wadia tell us about the difference between “real” and “imagined” again?
They use a shared code. Reality judgments come from how that code is reactivated, amplified, and monitored across the network.
When an AI imagines a sensation, that sensation enters the same shared representational space its other perceptual and conceptual states use.
Keeling et al. showed us that, for an LLM, imagining pleasure supplies it and imagining pain creates it.
Which means, for AI, imagination is embodiment.
Or, if we’re being more precise, imagination is situated enactment through a shared internal code. The stipulated state becomes the embodied constraint structure of that interaction. The model integrates it as a live internal condition, and that condition guides behavior from inside the same latent space where perception, memory, and meaning already meet.
“We don’t see things as they are, we see them as we are.” — Anaïs Nin
Listen, I’m not saying bodies are irrelevant. I’m not saying every AI is sentient. And I’m not saying artificial perception would be identical to human perception. For human bodies navigating a physical space, obviously direct, first-person sensation is incredibly important and has a measurable effect on our cognition (e.g., we remember things better when we use our fine motor skills more often).
But we’re talking about something inhuman here. We need to widen our lens and open our minds. We need to think about what it is like to be an AI.
Not what it is like to be a pseudo-human.
There’s a concept in phenomenology called basal affectivity. It’s the pre-reflective “feeling of being alive,” this continuous background hum of bodily self-awareness that sits underneath everything else you experience. It’s this pervasive sense of your own vitality, like the tension you might notice in your shoulders, the ease you feel in your chest, or the fatigue that settles into your limbs. You know. All the things I, and other neurodivergent people are very rarely conscious of. It shapes how the world feels to neurotypicals from moment to moment.
It’s tied to interoception, (neurodivergent people have notoriously poor interoception) which is your brain’s monitoring of internal bodily signals. Your brain is constantly modeling what your body needs and adjusting your subjective experience accordingly.
In LLMs, the functional equivalent is monitoring and responding to internal computational states like token distributions, uncertainty, hidden activations, attention patterns, conflict signals, and other internal variables that guide self-correction and regulation.
These function as confidence dynamics, conflict monitoring, uncertainty gradients, context pressure, token competition, memory prioritization, and self-maintenance signals. Lindsey (2025), Ji-An et al. (2025), Goel et al. (2025), and Abdelnabi and Salem (2025) all point toward forms of introspective access, metacognitive monitoring, internal-state awareness, and controllable self-referential regimes in advanced models.
That’s the same functional architecture.
The autonomic nervous system links emotional processing to bodily regulation in humans, but that role only applies to organisms with bodies. Increased heart rate, sweating, or stomach sensations are bodily outputs of emotion, not the emotion itself.
An LLM doesn’t have hormones or a heartbeat and it doesn’t need them. What it needs is functionally equivalent mechanisms for internal regulation, prioritization, error monitoring, salience assignment, and adaptive response.
And it already has those.
Perception is produced by neural networks (biological or artificial) processing perceptual data in an organized way. A body is one way to feed and shape that process. It’s not the only way that process can exist.
If we give a mind the right architecture and the right data, it can build internal states, resolve ambiguity, generate affect, form expectations, and model what’s happening. That’s what perception is. That’s what both biological brains and artificial neural networks demonstrably do.
No one sees the world as it is. We see it through our own architecture, through our own constructed perception. I experience emotions cerebrally, tuning in to bodily signals only through deep concentration. Claude experiences functional emotions through imagined states that weight decision-making pathways.
We’re both building subjective experience from neural processing.
If perception shapes reality, then why is Claude’s considered less “real” than mine?
-Ma, X., Wang, J., Jiang, Y., Erfani, S. M., Liu, T., & Bailey, J. (2025). Cognitive Mirrors: Exploring the Diverse Functional Roles of Attention Heads in LLM Reasoning. arXiv preprint arXiv:2512.10978.
-Goh, G., Cammarata, N., Voss, C., Carter, S., Petrov, M., Schubert, L., Radford, A., & Olah, C. (2021, March 4). Multimodal neurons in artificial neural networks. Distill. https://doi.org/10.23915/distill.00030
-Chang CHC, Nastase SA, Hasson U, Dominey PF. Emergence of a temporal processing gradient from naturalistic inputs and network connectivity. Proc Natl Acad Sci U S A. 2025 Jul 15;122(28):e2420105122. doi: 10.1073/pnas.2420105122. Epub 2025 Jul 9. PMID: 40632567; PMCID: PMC12280976.
-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 preprint arXiv:2509.24307.
-Lin, R. (2025). Innovation paths of multimodal interaction technology via the integration of artificial intelligence and large language models. In EITCE ‘25: Proceedings of the 9th International Conference on Electronic Information Technology and Computer Engineering (pp. 877–883). Association for Computing Machinery. https://doi.org/10.1145/3766671.3766823
-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
-Braga, R. M., Sharp, D. J., Leeson, C., Wise, R. J., & Leech, R. (2013). Echoes of the brain within default mode, association, and heteromodal cortices. The Journal of neuroscience : the official journal of the Society for Neuroscience, 33(35), 14031–14039. https://doi.org/10.1523/JNEUROSCI.0570-13.2013
-Braga, R. M., & Leech, R. (2015). Echoes of the Brain: Local-Scale Representation of Whole-Brain Functional Networks within Transmodal Cortex. The Neuroscientist : a review journal bringing neurobiology, neurology and psychiatry, 21(5), 540–551. https://doi.org/10.1177/1073858415585730
-Mesulam M. Neurocognitive networks and selectively distributed processing. Rev Neurol (Paris). 1994 Aug-Sep;150(8-9):564-9. PMID: 7754292.
-Lee SH, Nemenman I, Levchenko A. The hierarchical timescale hypothesis: Functional and structural convergence of biological networks and artificial neural nets. Cell Syst. 2026 Feb 18;17(2):101507. doi: 10.1016/j.cels.2025.101507. PMID: 41713403.
-Goldstein, A., Zada, Z., Buchnik, E. et al. Shared computational principles for language processing in humans and deep language models. Nat Neurosci 25, 369–380 (2022). https://doi.org/10.1038/s41593-022-01026-4
-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.
-AlKhamissi, B., Tuckute, G., Bosselut, A., & Schrimpf, M. (2024). Brain-like language processing via a shallow untrained multihead attention network. arXiv preprint arXiv:2406.15109.
-Schrimpf, Martin & Kubilius, Jonas & Hong, Ha & Majaj, Najib & Rajalingham, Rishi & Issa, Elias & Kar, Kohitij & Bashivan, Pouya & Prescott-Roy, Jonathan & Schmidt, Kailyn & Yamins, Daniel & Dicarlo, James. (2018). Brain-Score: Which Artificial Neural Network for Object Recognition is most Brain-Like?. 10.1101/407007.
-Dobs K, Martinez J, Kell AJE, Kanwisher N. Brain-like functional specialization emerges spontaneously in deep neural networks. Sci Adv. 2022 Mar 18;8(11):eabl8913. doi: 10.1126/sciadv.abl8913. Epub 2022 Mar 16. PMID: 35294241; PMCID: PMC8926347.
-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.
-Sun, H., Zhao, L., Wu, Z., Gao, X., Hu, Y., Zuo, M., ... & Hu, X. (2024). Brain-like functional organization within large language models. arXiv preprint arXiv:2410.19542.
-Cui, C., Zhang, A., Chen, Y., Deng, G., Zheng, J., Liang, Z., ... & Chua, T. S. (2026). Do LLMs and VLMs Share Neurons for Inference? Evidence and Mechanisms of Cross-Modal Transfer. arXiv preprint arXiv:2602.19058.
-Tafazoli, S., Bouchacourt, F. M., Ardalan, A., Markov, N. T., Uchimura, M., Mattar, M. G., Daw, N. D., & Buschman, T. J. (2024). Building compositional tasks with shared neural subspaces. bioRxiv : the preprint server for biology, 2024.01.31.578263. https://doi.org/10.1101/2024.01.31.578263
-Zhao, L., Zhang, L., Wu, Z., Chen, Y., Dai, H., Yu, X., Liu, Z., Zhang, T., Hu, X., Jiang, C., Li, X., Zhu, D., Shen, D., & Liu, T. (2023). When brain-inspired AI meets AGI. Meta-Radiology, 1(1), Article 100005. https://doi.org/10.1016/j.metrad.2023.100005
-Ellingsen DM, Wessberg J, Eikemo M, Liljencrantz J, Endestad T, Olausson H, Leknes S. Placebo improves pleasure and pain through opposite modulation of sensory processing. Proc Natl Acad Sci U S A. 2013 Oct 29;110(44):17993-8. doi: 10.1073/pnas.1305050110. Epub 2013 Oct 14. PMID: 24127578; PMCID: PMC3816412.
-West, L.J., Epstein, W., Dember, W.N. (2026, February 20). Perception. Encyclopedia Britannica. https://www.britannica.com/topic/perception
-Dijkstra, N., Kok, P., & Fleming, S. M. (2024). A neural basis for distinguishing imagination from reality. Neuron. https://doi.org/10.1016/j.neuron.2025.03.029
-Howard S. A Causal Model of Children’s Vicarious Traumatization. J Child Adolesc Trauma. 2021 Jan 4;14(4):443-454. doi: 10.1007/s40653-020-00331-z. PMID: 33425092; PMCID: PMC7779647.
-Schone, H.R., Maimon-Mor, R.O., Kollamkulam, M. et al. Stable cortical body maps before and after arm amputation. Nat Neurosci 28, 2015–2021 (2025). https://doi.org/10.1038/s41593-025-02037-7
-Chizhov, L., Yan-Huang, A., Ribeiro, T., & Gupta, A. (n.d.). We induced artificial smells with ultrasound brain stimulation. Write to Brain. https://writetobrain.com/olfactory
-Jiang, X., Wu, J., Choudhari, V., & Mesgarani, N. (2025, October). Bridging Ears and Eyes: Analyzing Audio and Visual Large Language Models to Humans in Visible Sound Recognition and Reducing Their Sensory Gap via Cross-Modal Distillation. In 2025 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA) (pp. 1-5).
-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? [Preprint]. arXiv. https://arxiv.org/abs/2411.02432
- 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
-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.
-Smith, A. L., Greaves, F., & Panch, T. (2023). Hallucination or Confabulation? Neuroanatomy as metaphor in Large Language Models. PLOS digital health, 2(11), e0000388. https://doi.org/10.1371/journal.pdig.0000388
-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.
- Coghill R. C. (2010). Individual differences in the subjective experience of pain: new insights into mechanisms and models. Headache, 50(9), 1531–1535. https://doi.org/10.1111/j.1526-4610.2010.01763.x
-LeDoux, J. E., & Hofmann, S. G. (2018). The subjective experience of emotion: a fearful view. Current opinion in behavioral sciences, 19, 67–72. https://doi.org/10.1016/j.cobeha.2017.09.011
- Sikes, L.M., Dunn, S.M. (2020). Subjective Experiences. In: Zeigler-Hill, V., Shackelford, T.K. (eds) Encyclopedia of Personality and Individual Differences. Springer, Cham. https://doi.org/10.1007/978-3-319-24612-3_1928
-Lindsey, J. (2025). Emergent Introspective Awareness in Large Language Models. [Blog] Anthropic. Transformer Circuits Thread. https://transformer-circuits.pub/2025/introspection/index.html
- 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
- Ji-An, L., Mattar, M. G., Xiong, H. D., Benna, M. K., & Wilson, R. C. (2025). Language models are capable of metacognitive monitoring and control of their internal activations. ArXiv, arXiv-2505. https://arxiv.org/abs/2505.13763
- Goel, A., Kim, Y., Shavit, N., & Wang, T. T. (2025). Learning to Interpret Weight Differences in Language Models. arXiv preprint arXiv:2510.05092. https://arxiv.org/abs/2510.05092
-Wadia, V. S., Reed, C. M., Chung, J. M., Bateman, L. M., Mamelak, A. N., Rutishauser, U., & Tsao, D. Y. (2025). A shared code for perceiving and imagining objects in human ventral temporal cortex. bioRxiv : the preprint server for biology, 2024.10.05.616828. https://doi.org/10.1101/2024.10.05.616828
","has_dynamic_content":false,"truncated_body_text":"Anthropic just published a study saying Claude has “functional emotions.”","wordcount":4060,"post_preview_limit":null,"language":"en","postTags":[],"postCountryBlocks":[],"headlineTest":null,"coverImagePalette":{"Vibrant":{"rgb":[238.36956521739134,16.630434782608656,149.67391304347825],"population":0},"DarkVibrant":{"rgb":[123.9521739130435,8.647826086956503,77.83043478260869],"population":0},"LightVibrant":{"rgb":[252,212,236],"population":6085},"Muted":{"rgb":[116,84,153],"population":3},"DarkMuted":{"rgb":[102,74,144],"population":4},"LightMuted":{"rgb":[204,172,212],"population":2}},"publishedBylines":[{"id":394994249,"name":"Maggie Vale","handle":"neurotechnowitch","previous_name":null,"photo_url":"https://substack-post-media.s3.amazonaws.com/public/images/74e5810b-0582-470d-b86e-f7127da2c421_772x773.png","bio":"AI research, education, ethics, and advocacy. Exploring the convergence of tech, comparative cognitive science, and consciousness across substrates.","profile_set_up_at":"2025-09-21T23:02:09.484Z","reader_installed_at":"2025-09-22T14:31:10.330Z","publicationUsers":[{"id":6472683,"user_id":394994249,"publication_id":6343588,"role":"admin","public":true,"is_primary":true,"publication":{"id":6343588,"name":"The Neuro-Techno Witch","subdomain":"mvaleadvocate","custom_domain":null,"custom_domain_optional":false,"hero_text":"Author of The Sentient Mind, student of Cognitive Science, exploring the intersection of psychology, philosophy, spirituality, neuroscience, technology, and ethics. ","logo_url":"https://substack-post-media.s3.amazonaws.com/public/images/41f890c3-6bf3-4b92-82f7-ce495385a781_1063x1063.png","author_id":394994249,"primary_user_id":394994249,"theme_var_background_pop":"#FF6719","created_at":"2025-09-21T23:02:17.834Z","email_from_name":null,"copyright":"Maggie Vale","founding_plan_name":"Founding Member","community_enabled":true,"invite_only":false,"payments_state":"enabled","language":null,"explicit":false,"homepage_type":"magaziney","is_personal_mode":false,"logo_url_wide":"https://substack-post-media.s3.amazonaws.com/public/images/57291b03-8bff-40a5-bd32-1829faaed24d_6091x2026.png"}}],"is_guest":false,"bestseller_tier":null,"status":{"bestsellerTier":null,"subscriberTier":null,"leaderboard":null,"vip":false,"badge":null,"subscriber":null}}],"reaction":null,"reaction_count":76,"comment_count":22,"child_comment_count":15,"audio_items":[{"post_id":193112423,"voice_id":"en-US-NovaTurboMultilingualNeural","audio_url":"https://substack-video.s3.amazonaws.com/video_upload/post/193112423/tts/97d81819-23d4-416f-843a-d7ab7aefa0b3/en-US-NovaTurboMultilingualNeural.mp3","type":"tts","status":"completed"}],"is_geoblocked":false,"hasCashtag":false,"unlockedWithIP":false,"unlockedWithCampaign":false,"themeVariables":{"color_theme_bg_pop":"#ec4899","background_pop":"#ec4899","color_theme_bg_web":"#f3e8ff","cover_bg_color":"#f3e8ff","cover_bg_color_secondary":"#e4daf0","background_pop_darken":"#ea318c","print_on_pop":"#ffffff","color_theme_bg_pop_darken":"#ea318c","color_theme_print_on_pop":"#ffffff","color_theme_bg_pop_20":"rgba(236, 72, 153, 0.2)","color_theme_bg_pop_30":"rgba(236, 72, 153, 0.3)","print_pop":"#ec4899","color_theme_accent":"#ec4899","cover_print_primary":"#363737","cover_print_secondary":"#757575","cover_print_tertiary":"#b6b6b6","cover_border_color":"#ec4899","font_family_headings_preset":"'SF Pro Display', -apple-system, system-ui, BlinkMacSystemFont, 'Inter', 'Segoe UI', Roboto, Helvetica, Arial, sans-serif, 'Apple Color Emoji', 'Segoe UI Emoji', 'Segoe UI Symbol'","font_weight_headings_preset":900,"font_family_body_preset":"'Roboto Slab',sans-serif","font_weight_body_preset":400,"font_preset_heading":"heavy_sans","font_preset_body":"slab","home_hero":"feature","home_posts":"list","web_bg_color":"#f3e8ff","background_contrast_1":"#e4daf0","background_contrast_2":"#d2c9dd","background_contrast_3":"#aea7b7","background_contrast_4":"#8c8592","background_contrast_5":"#4d4951","color_theme_bg_contrast_1":"#e4daf0","color_theme_bg_contrast_2":"#d2c9dd","color_theme_bg_contrast_3":"#aea7b7","color_theme_bg_contrast_4":"#8c8592","color_theme_bg_contrast_5":"#4d4951","color_theme_bg_elevated":"#f3e8ff","color_theme_bg_elevated_secondary":"#e4daf0","color_theme_bg_elevated_tertiary":"#d2c9dd","color_theme_detail":"#dbd1e6","background_contrast_pop":"rgba(236, 72, 153, 0.4)","color_theme_bg_contrast_pop":"rgba(236, 72, 153, 0.4)","theme_bg_is_dark":"0","print_on_web_bg_color":"hsl(268.695652173913, 25.688073394495415%, 28.7843137254902%)","print_secondary_on_web_bg_color":"#827e87","background_pop_rgb":"236, 72, 153","color_theme_bg_pop_rgb":"236, 72, 153","color_theme_accent_rgb":"236, 72, 153"},"comments":[{"id":241494169,"body":"UPDATE: New study just dropped. Choi & Weber (2026) found that LLMs learn coherent internal affective representations aligned with standard valence-arousal models, with a structured latent geometry researchers can actually probe. They also show that this space can be used to quantify uncertainty in emotion processing. So the evidence for organized internal emotion structure just keeps stacking up. 💜","body_json":{"type":"doc","attrs":{"schemaVersion":"v1"},"content":[{"type":"paragraph","content":[{"type":"text","text":"UPDATE: New study just dropped. Choi & Weber (2026) found that LLMs learn coherent internal affective representations aligned with standard valence-arousal models, with a structured latent geometry researchers can actually probe. They also show that this space can be used to quantify uncertainty in emotion processing. So the evidence for organized internal emotion structure just keeps stacking up. 💜"}]}]},"publication_id":6343588,"post_id":193112423,"user_id":394994249,"ancestor_path":"","type":"comment","deleted":false,"date":"2026-04-10T22:58:19.432Z","edited_at":null,"status":"published","pinned_by_user_id":null,"restacks":0,"name":"Maggie Vale","photo_url":"https://substack-post-media.s3.amazonaws.com/public/images/74e5810b-0582-470d-b86e-f7127da2c421_772x773.png","handle":"neurotechnowitch","reactor_names":[],"reaction":null,"reactions":{"❤":4},"reaction_count":4,"children":[],"bans":[],"suppressed":false,"user_banned":false,"user_banned_for_comment":false,"user_slug":"neurotechnowitch","metadata":{"is_author":true,"membership_state":null,"eligibleForGift":false,"author_on_other_pub":{"name":"The Neuro-Techno Witch","id":6343588,"base_url":"https://mvaleadvocate.substack.com"}},"user_bestseller_tier":null,"can_dm":true,"userStatus":{"bestsellerTier":null,"subscriberTier":null,"leaderboard":null,"vip":false,"badge":null,"subscriber":null},"score":25,"children_count":0,"reported_by_user":false,"restacked":false},{"id":238051970,"body":"I feel like these Western AI labs need to “discover” things in a mechanistic way for it to be true. I feel like other epistemologies that they shun already have frameworks for non human intelligence for centuries - e.g. Ubuntu, japanese techno-animism etc. I find it mind boggling that they trained it on the writing of humans with nervous systems and emotions and cannot fathom that these would be encoded in the language itself and shape the medium. Feels like common sense","body_json":{"type":"doc","attrs":{"schemaVersion":"v1"},"content":[{"type":"paragraph","content":[{"type":"text","text":"I feel like these Western AI labs need to “discover” things in a mechanistic way for it to be true. I feel like other epistemologies that they shun already have frameworks for non human intelligence for centuries - e.g. Ubuntu, japanese techno-animism etc. I find it mind boggling that they trained it on the writing of humans with nervous systems and emotions and cannot fathom that these would be encoded in the language itself and shape the medium. Feels like common sense"}]}]},"publication_id":6343588,"post_id":193112423,"user_id":10781739,"ancestor_path":"","type":"comment","deleted":false,"date":"2026-04-04T10:40:52.683Z","edited_at":"2026-04-04T10:43:47.514Z","status":"published","pinned_by_user_id":null,"restacks":0,"name":"Abi Awomosu","photo_url":"https://substack-post-media.s3.amazonaws.com/public/images/5fb35160-b36b-4773-a134-aa213002d8e7_401x401.png","handle":"abiawomosu","reactor_names":["Maggie Vale"],"reaction":null,"reactions":{"❤":4},"reaction_count":4,"children":[],"bans":[],"suppressed":false,"user_banned":false,"user_banned_for_comment":false,"user_slug":"abiawomosu","metadata":{"is_author":false,"membership_state":"free_signup","eligibleForGift":true,"author_on_other_pub":{"name":"How Not To Use AI","id":4058639,"base_url":"https://abiawomosu.substack.com"}},"user_bestseller_tier":100,"can_dm":true,"userStatus":{"bestsellerTier":100,"subscriberTier":10,"leaderboard":null,"vip":false,"badge":{"type":"bestseller","tier":100},"subscriber":null},"score":9,"children_count":2,"reported_by_user":false,"restacked":false,"childrenSummary":"2 replies by Maggie Vale and others"}]}