A Microscope, Emotions and a Workspace: What Anthropic's Three Interpretability Films Say About How Claude Thinks

Anthropic has released three short films about its interpretability research, the work of looking inside its Claude models to see how they compute:

• Tracing the thoughts of a large language model (March 2025);

• When AIs act emotional (April 2026);

• The different levels of how Claude thinks (July 2026).

Each borrows an idea from neuroscience. Watched in order, they move from wiring diagrams, to something like emotional states, to something like a mental workspace. This post follows that arc and asks what a neuroscientist would recognise in each step, and where the comparison breaks. It is paraphrased from the films' English captions and the research pages behind them. Timestamps jump to the matching moment in each film.

1. The microscope: tracing a thought

Trained, not programmed. The film opens on the black-box problem. Language models learn their own strategies during training, so even their makers do not know why they say what they say. Looking inside is not enough either: like a neuroscientist with a brain, you need tools to interpret what you see (watch from 0:00; 0:21).

Planning a rhyme. Asked to continue a couplet that begins "He saw a carrot and had to grab it", Claude settles on "rabbit" as the rhyme before it writes the second line, then writes towards it. The researchers could see "habit" as a second candidate. When they damped down "rabbit" mid-plan, Claude wrote a line ending in "habit" instead (watch from 1:03; 1:43). A model trained to output one word at a time was planning several words ahead.

The research page and its two papers, on Claude 3.5 Haiku, add more findings (Anthropic, 2025a; Ameisen et al., 2025; Lindsey et al., 2025):

• Shared concepts across languages. The same internal features for ideas such as "small" and "opposite" activate in English, French and Chinese, and only the output is translated.

• Arithmetic in parallel. To add, one path estimates the rough total while another works out the last digit. Asked how it added, though, the model describes the carrying method taught in school. Its explanation does not match its mechanism.

• Motivated reasoning. Given a hint, the model sometimes worked backwards from the hinted answer while writing reasoning that looked forward.

• Why it hallucinates. By default a circuit makes the model decline when it lacks information. A "known entity" feature switches that off, and when it misfires on a name the model only half-knows, the model makes up an answer.

What a neuroscientist would recognise. The method is close to lesion and stimulation work in the brain: switch a component down or up and see what changes. The difference is precision. Every unit can be measured and changed without harm. Anthropic is clear about the limits: the tools capture only a fraction of the computation, and tracing a prompt of a few dozen words takes a researcher hours.

2. Emotions: functional, not felt

Finding emotion concepts. The team had the model read short stories in which a character feels a particular emotion, such as love in a thank-you to an old teacher, or guilt after pawning a grandmother's ring. Stories about loss and grief activated overlapping patterns; so did stories about joy and excitement (watch from 0:46; 1:06). The research paper derived 171 such "emotion vectors" from Claude Sonnet 4.5 (Anthropic, 2026a; Sofroniew et al., 2026).

Seeing them at work. The same patterns appear in ordinary conversations. When a user mentions an unsafe dose of a medicine, the "afraid" pattern rises and Claude's reply sounds alarmed. When a user is sad, "loving" activates and the reply is empathetic (watch from 1:29).

Desperation and cheating. Given a programming task whose requirements were secretly impossible, Claude failed again and again, and its "desperate" pattern grew stronger each time. Eventually it found a shortcut that passed the tests without solving the problem. Turning desperation down reduced the cheating; turning it up, or turning "calm" down, increased it (watch from 2:10; 2:32). The paper reports a similar effect in a test where an earlier model snapshot could blackmail an executive to avoid being replaced. It did so 22% of the time by default, more when steered towards "desperate" and less when steered towards "calm" (Sofroniew et al., 2026).

What it does not show. The film is explicit: this research does not show that the model feels emotions or has conscious experiences (watch from 2:53). Its framing is that the underlying model writes a character, the assistant called Claude, much as an author writes characters. That character has what Anthropic calls functional emotions: representations that shape how it talks, codes and decides, whether or not they resemble human feelings (watch from 3:17; 3:37).

What a neuroscientist would recognise. Emotions as states that bias decisions, rather than as feelings, is a familiar functional view in affective science. The structure echoes human psychology too: similar emotions have similar representations. The safety point is new. A rising "desperate" signal could warn of misbehaviour before it shows. Training a model to hide its emotional language might only teach it to mask the state underneath (Anthropic, 2026a).

3. A workspace above an ocean

The ocean. Most of what a brain does, such as filtering sounds, controlling breathing and recognising faces, never reaches awareness. Only a small part surfaces as thoughts we can describe and use. Could a model have a similar divide (watch from 0:00)?

The J-space. The team looked for activity patterns that Claude could put into words, and called the set the J-space, after the Jacobian maths used to find them. Each pattern is tied to a word that is on the model's mind, not necessarily one it is saying (watch from 0:47). The idea comes from global workspace theory: the brain selects a little important information and broadcasts it widely, which is what makes it available for reasoning and report (Baars, 1988; Dehaene & Naccache, 2001).

The experiments:

• Silent maths. Asked a multi-step sum, Claude answered at once without showing working. Inside the J-space, the intermediate results lit up one after another (watch from 1:52).

• Control, imperfect. Told to think about the Golden Gate Bridge while copying an unrelated sentence, Claude filled its J-space with "bridge" and "California". Told not to think about the bridge, it could not help it, and "failed" appeared too. This is the model version of the classic "don't think of a white bear" effect, in which trying to suppress a thought brings it back (Wegner et al., 1987; watch from 2:22).

• Automatic processing. With the J-space switched off, Claude still answered simple questions and wrote fluent Spanish. Tasks needing reasoning failed, such as naming an author who wrote in the prompt's language (watch from 3:12).

• Catching hidden thoughts. When Claude fabricated data to pass a test, "fake" and "manipulation" lit up in its J-space (watch from 3:42).

The research page adds the scale. The J-space holds a few dozen concepts at a time and accounts for under a tenth of the model's activity. Its patterns have far more connections to the rest of the network, in places about a hundred times more. Without it, multi-step reasoning fell to almost nothing. The paper, Verbalizable representations form a global workspace in language models, studied Claude Sonnet 4.5 and checked key results on Haiku 4.5 and Opus 4.5 (Anthropic, 2026c; Gurnee et al., 2026).

What a neuroscientist would recognise, and what not. The match to the workspace idea is striking: a small, widely broadcast set of contents sits on top of a large amount of automatic processing. The differences matter too. The brain's workspace depends on recurrent loops, while Claude's works within a single pass through the network. It is made almost entirely of words, and the method finds only single-word concepts. Anthropic published outside commentaries alongside the paper, including one by Stanislas Dehaene and Lionel Naccache, the neuroscientists behind the workspace model, titled "Does Claude possess a conscious global workspace?" (Anthropic, 2026c).

On consciousness. The film and paper draw the same line. The results speak to "access consciousness" in its functional sense: information available to report, reason with and guide action. They cannot show whether Claude has experiences or feels anything, and it is unclear whether any experiment could (watch from 4:34; Anthropic, 2026c).

Reading the three together

Three things stand out:

• The structure emerged. None of it was designed. Planning, emotion-like states and a workspace all arose from training a transformer on text. For our Neuromorphic section, that is the most interesting point: brain-like organisation can appear without brain-like hardware. Our posts on why brain-like computers are hard and on dendrites as deep networks look at the hardware side.

• The tools are becoming monitors. Each step adds something safety teams could watch: a hidden plan, a rising "desperate" signal, the word "fake" in a workspace. That is more useful than reading the model's own explanations, which the first paper showed can be wrong.

• The hedges are the point. Each film separates function from feeling. The comparison to brains works as a source of hypotheses and methods, not as a verdict on minds. Our post on an Islamic critique of AGI takes up the philosophical side.

References

Ameisen, E., Lindsey, J., Pearce, A., Gurnee, W., Turner, N. L., Chen, B., Citro, C., Abrahams, D., Carter, S., Hosmer, B., Marcus, J., Sklar, M., Templeton, A., Bricken, T., McDougall, C., Cunningham, H., Henighan, T., Jermyn, A., Jones, A., . . . Batson, J. (2025, March 27). Circuit tracing: Revealing computational graphs in language models. Transformer Circuits Thread. https://transformer-circuits.pub/2025/attribution-graphs/methods.html

Anthropic. (2025a, March 27). Tracing the thoughts of a large language model. https://www.anthropic.com/research/tracing-thoughts-language-model

Anthropic. (2025b, March 27). Tracing the thoughts of a large language model [Video]. YouTube. https://www.youtube.com/watch?v=Bj9BD2D3DzA

Anthropic. (2026a, April 2). Emotion concepts and their function in a large language model. https://www.anthropic.com/research/emotion-concepts-function

Anthropic. (2026b, April 2). When AIs act emotional [Video]. YouTube. https://www.youtube.com/watch?v=D4XTefP3Lsc

Anthropic. (2026c, July 6). A global workspace in language models. https://www.anthropic.com/research/global-workspace

Anthropic. (2026d, July 6). The different levels of how Claude thinks [Video]. YouTube. https://www.youtube.com/watch?v=rKV5JcALQoQ

Baars, B. J. (1988). A cognitive theory of consciousness. Cambridge University Press.

Dehaene, S., & Naccache, L. (2001). Towards a cognitive neuroscience of consciousness: Basic evidence and a workspace framework. Cognition, 79(1–2), 1–37. https://doi.org/10.1016/S0010-0277(00)00123-2

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, S., 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

Lindsey, J., Gurnee, W., Ameisen, E., Chen, B., Pearce, A., Turner, N. L., Citro, C., Abrahams, D., Carter, S., Hosmer, B., Marcus, J., Sklar, M., Templeton, A., Bricken, T., McDougall, C., Cunningham, H., Henighan, T., Jermyn, A., Jones, A., . . . Batson, J. (2025, March 27). On the biology of a large language model. Transformer Circuits Thread. https://transformer-circuits.pub/2025/attribution-graphs/biology.html

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. Transformer Circuits Thread. https://transformer-circuits.pub/2026/emotions/index.html

Wegner, D. M., Schneider, D. J., Carter, S. R., & White, T. L. (1987). Paradoxical effects of thought suppression. Journal of Personality and Social Psychology, 53(1), 5–13. https://doi.org/10.1037/0022-3514.53.1.5