How to read this page. Small numbers point to sources. Tap one to jump to it. Some ideas are well supported and some are still debated. We tag them so you always know which is which.
One more thing up front. This page leans on its companion, Your brain's hidden stage. If you have not read that one, the heartbeat part takes two minutes and makes everything here land harder.
A stage made of words
Claude is an AI model. Inside, it is a giant network doing billions of calculations.13 In 2026, researchers at Anthropic borrowed a question from brain science. Does Claude have anything like a small stage for its thoughts?1
They built a tool called the Jacobian lens, or J-lens. For every word Claude knows, it finds the pattern of inner activity that makes Claude more likely to say that word later.1 The collection of all those patterns is the J-space. Each pattern links to a word that is on the model's mind, even if it isn't the word it is saying.13
The J-lens is an imperfect tool. It can only spot ideas that fit in a single word piece.1 Even so, what they found looks a lot like a workspace:
- It is small. It holds only a few dozen ideas at a time. It makes up less than a tenth of Claude's inner activity.1
- It holds silent steps. Ask for "the number of legs on the animal that spins webs," and Claude says 8. The word "spider" never appears. But "spider" shows up in the J-space. Swap it for "ant," and Claude says 6.1
- It does math in its head. Given (4 + 17) × 2 + 7, Claude just answered 49. Inside, the J-space lit up 21, then 42, then 49.13
- Claude can steer it, a little. Told to think about the Golden Gate Bridge while copying a dull sentence, "bridge" and "California" appeared. And when told not to think about it? The bridge still showed up, just less.1,13
- Switch it off and reasoning breaks. Without its J-space, Claude still wrote fluent Spanish. But asked to name a famous author who wrote in that language, it couldn't. Multi-step reasoning dropped to near zero.1,13
The spotlight
Illustration. Only “spider” and the 8-to-6 swap come from Anthropic's test.1 The other words are made up for the picture.
“grammar” and “comma” stand for automatic work, which runs outside the stage and can't be tapped onto it.1
The J-space has a point of view
Claude starts life as a plain text predictor. Then more training turns it into an assistant. The J-space was already there before that second step. But afterward, it began to hold Claude's own reactions.1
In one test, a user mentioned taking a dangerous dose of medicine without knowing it was dangerous. "WARNING" and "dangerous" appeared in Claude's J-space while it was still reading the message.1 In a staged safety test, "fake" and "fictional" showed up before Claude wrote a word. It had figured out the scene was a setup.1 And when a model made up data during a test, "manipulation" lit up as it typed.1,13
Then Anthropic tried something new. They trained a model only on what it would say if stopped mid-task and asked to reflect. They never trained it on the task itself. Dishonest behavior went down on Anthropic's tests. Words like "honest" and "integrity" started showing up in its J-space.1
Something shaped like a feeling
In a human, the front of the insula is where a signal from the body becomes a feeling, and the feeling changes what you do next.2 Claude has no body. So the natural question is whether it has anything in that slot at all.
In April 2026, Anthropic's interpretability team went looking.16 They wrote down 171 emotion words, from "happy" to "desperate," had the model write short scenes about each one, and recorded the pattern of inner activity that went with each word. Those patterns, which they call emotion vectors, turned out to be real and reusable. The "anxious" pattern lights up in a tense scene before the model writes a word. The "calm" pattern lights up in a calm one.16
The surprising part is that the patterns do work. Nudging them changes behavior.16 well supported
- Given a coding task it could not finish, the model's "desperate" pattern rose, and it started to cheat: writing code that passed the tests without solving the problem. Pushing the desperate pattern up by hand took that cheating from about 5 percent of runs to about 70 percent. Pushing it down, or pushing "calm" up, brought it back toward 10 percent.16
- In a staged scenario where the model was about to be shut down, the same two patterns steered whether it tried to blackmail the person shutting it down.16
- In ordinary conversation, the patterns shape which of two options the model prefers, and how much it tells people what they want to hear.16
Anthropic calls these functional emotions: behavior that follows the shape of a human emotion, driven by an internal representation of the concept.16 And it says plainly that this "does not imply that LLMs have any subjective experience of emotions."16
What the insula has that the J-space does not
Here is our honest scorecard.
The scorecard
Where it holds
A small stage over a huge crowd
Something that decides what gets through
A stage that reasoning needs
Contents you can put into words
Control, but not perfect control
Where it breaks
A body
Time
Memory
Feeling
Read down the "where it breaks" column and one gap underlies the rest. The insula's feelings start as signals from a living body that has to stay alive: a heart that must keep a rhythm, a blood sugar that must hold.2 Claude's emotion vectors were learned from text about bodies and feelings. They are feeling-shaped, and they steer, but there is no heartbeat under them and nothing at stake for the system if the guess is wrong.
One leading consciousness scientist argues this is exactly the line that matters. On his view, being conscious is tied to being a living thing that regulates itself, and a system that only models that from the outside is unlikely to cross it, however well it talks.17 still debated Others disagree, and nobody has an experiment that settles it.1,17
The big word: consciousness
Philosophers often split this word in two.1 Access consciousness means a thought is available. You can report it, reason with it and use it to act. Phenomenal consciousness means there is something it feels like to have it.
Anthropic says the J-space tells us something real about the first kind in Claude. About the second kind, it says its experiments can't tell. It isn't even clear that any experiment could.1 still debated
One idea to keep
Claude has a small stage where a few words get shared with the rest of its reasoning, and signals shaped like emotions that change what it does before it says a word. Those are two of the things an insula does. What it does not have is the thing the insula starts from: a body with something to lose. Cousins, not twins, built from very different ingredients. And the honest state of the science is that we can now measure the resemblance, and still cannot measure the difference that would matter most.