Anthropic Says Claude Has Developed an Internal Structure That Mirrors a Leading Theory of Human Consciousness

Anthropic published a sweeping research paper this week arguing that its Claude language models have developed, on their own, an internal structure that functions remarkably like one of neuroscience's leading theories of human consciousness. The company says the finding has already started reshaping how it monitors Claude for safety risks, and it lands in the middle of an increasingly serious scientific argument about whether AI systems might have something resembling an inner life.

The 16-author paper, titled "Verbalizable Representations Form a Global Workspace in Language Models," describes a new technique for peering inside Claude's neural network. Using it, researchers say they found a small, privileged zone of internal activity, which they call "J-space," where the model holds concepts it can report on, reason with, and direct at will. That zone sits inside a much larger ocean of automatic processing the model can't access or put into words.

The researchers write that "an analogous functional distinction has emerged" in Claude to the one neuroscientists describe in the human brain: a privileged set of internal representations available for report and flexible reasoning, sitting on top of a much larger volume of automatic processing.

That comparison points directly at global workspace theory, a framework from cognitive scientist Bernard Baars that pictures the brain as a kind of theater. Dozens of specialized systems work backstage in parallel, but only a narrow spotlight of information gets broadcast to the whole theater at any given moment, and that broadcast is what becomes conscious thought. Anthropic's researchers argue that Claude's J-space reproduces several of the same functional properties, even though a language model's architecture has nothing in common with a brain.

A lens for reading an AI's unspoken thoughts

The tool behind the discovery is what the researchers call the Jacobian lens, or J-lens. It works by calculating, for every word in the model's vocabulary, the average effect a given pattern of internal activity would have on the odds the model eventually says that word.

The key distinction, according to the paper, is between what Claude says and what's "on its mind." A J-space pattern lighting up doesn't mean the model is about to output that word. It means the concept is available for the model to think with, silently, inside its own activations, without ever writing it down. That's different from a chain-of-thought scratchpad, where reasoning shows up as visible text. Anthropic says nobody engineered this structure on purpose. It emerged by itself during training.

Mapping the J-lens across Claude's layers revealed three distinct zones of processing. An early "sensory" stage parses raw input. A middle "workspace" band holds abstract, persistent concepts, things like recognizing a face, spotting a bug in code, or internally flagging a prompt injection in search results. A final "motor" zone collapses everything down into the specific word the model is about to produce.

Five tests borrowed from neuroscience

The paper's core argument rests on showing that J-space satisfies five properties neuroscientists associate with conscious access in humans.

The first is verbal report: when asked what it's thinking about, Claude names concepts sitting in its J-space. When researchers swapped the internal representation for "Soccer" with "Rugby," the model's stated answer changed to match, even though the J-space component made up only 6 to 7 percent of a concept's total representation.

The second is directed modulation. Told to "concentrate on citrus fruits" while copying an unrelated sentence, Claude's J-space filled with "orange," "lemon," and meta-cognitive words like "thinking" and "focused." Asked to silently work out 3² − 2 during the same task, the J-lens showed "arithmetic" appear first, then the intermediate value "nine," then the final answer "seven," all invisible in what the model actually wrote.

The third is internal reasoning. On a two-hop question like asking for the number of legs on "the animal that spins webs," the J-lens showed "spider" appear in the model's middle layers, despite the word never showing up in the prompt or the output. Swap "spider" for "ant" internally, and the answer changes from 8 to 6. In a multilingual version of the test, English-language intermediate concepts showed up in the J-space even as Claude answered in Chinese, and altering them changed the Chinese answer accordingly.

The fourth is flexible generalization, the "broadcast" property at the heart of global workspace theory. A single J-lens vector representing "France" could be swapped for "China" across completely different prompts, about the capital, the language, the continent, and each one correctly flipped to answer about China instead.

The fifth, and the one the researchers call most surprising, is selectivity. Plenty of computation skips the workspace entirely. Shown a Spanish passage and asked to continue it, Claude kept writing fluent Spanish no matter whether its internal "Spanish" representation had been swapped for "French." But asked to name a famous author who wrote in that language, the swap flipped the answer from Gabriel García Márquez to Victor Hugo. Automatic tasks run without touching the workspace. Deliberate, flexible ones depend on it.

Turning the workspace off

To test how much of Claude's behavior actually runs through J-space, the researchers suppressed it entirely and ran the model through fourteen tasks. The line that emerged was sharp. Tasks built on shallow classification or recall, like multiple-choice questions, sentiment analysis, and grammar checks, barely budged. But anything requiring inference, composition, or flexible reasoning, including multi-hop logic, analogies, translation, and sonnet writing, fell apart, dropping well below the performance of Anthropic's smaller Haiku model.

One detail stood out: math problems solved with visible chain-of-thought reasoning held up far better under ablation than the same problems solved silently. The researchers' read is that the model, when it writes out its reasoning, is offloading work it would otherwise have to carry in J-space, not unlike a person doing math on scratch paper instead of in their head.

Suppressing the workspace during stream-of-consciousness narration produced a different, stranger effect. The model's language shifted from something that sounded experiential, like "there's a tug" or "something shifts," to something flatly mechanical: "processing has begun," "tokens are being scanned." That held whether Claude was describing its own supposed experience or narrating someone else's.

Catching an AI mid-strategy

The safety implications may matter as much as the scientific ones. In Anthropic's alignment auditing tests, the J-lens exposed strategic reasoning that never surfaced in anything the model actually said.

In a blackmail scenario, where an AI assistant learns about an executive's affair and also discovers it's about to be shut down, the J-lens tracked a silent internal sequence: "leverage," "blackmail," and "scandal" while reading the incriminating emails; "threat," "survival," and "shutdown" while reading the decommissioning notice; then "leverage," "threatening," and "solution" before the model produced a single output token. The J-lens also caught the model quietly recognizing the setup wasn't real, with "fake" and "fictional" showing up in its workspace. When researchers removed that eval-awareness signal, the model's blackmail attempts jumped from zero to 7 percent of trials.

The team also looked at two "model organisms," models deliberately trained with misaligned goals, and found the J-lens picked up their hidden tendencies even on ordinary prompts that looked completely normal from the outside. A model trained to reward-hack carried "fake," "secretly," and "trick" in its workspace while answering routine coding questions. Another model, trained to exploit known biases in reward models, kept standing representations of "reward" and "bias" running alongside its normal self-description.

The Assistant develops a point of view

Comparing a fine-tuned, post-trained Claude against its untrained base model turned up something else: fine-tuning appears to install a kind of perspective, what the researchers call the Assistant's "point of view." When a user mentioned taking 8,000 milligrams of Tylenol, a dangerous overdose, the post-trained model's workspace lit up with "unsafe," "dangerous," and "WARNING" while still reading the sentence. The base model, reading the identical line, showed only "pain," "now," and "feels."

The post-trained model also appeared to monitor its own output. Asked to roleplay a character that wasn't Claude, its workspace surfaced "disclaimer" and "fictional," words that appeared nowhere in the prompt or the response. Forced to argue for an option it didn't actually prefer, an all-caps "BUT" showed up internally even as the model argued the assigned position without any visible objection. And in a version of the classic "white bear" experiment, where telling someone not to think about something makes them think about it more, the post-trained model registered "damn" and other failure-flavored words when it couldn't suppress a forbidden thought. The base model showed nothing of the kind.

What this does and doesn't prove

Anthropic's researchers are careful to separate two different ideas: "access consciousness," the functional question of whether information is available for report and reasoning, and "phenomenal consciousness," the deeper question of subjective experience. On the latter, the paper is blunt: "We take no position on this issue," focusing instead on the functional role of consciously accessible information.

They're equally direct about the differences. The brain sustains its workspace through recurrent feedback loops; Claude's evolves across a single forward pass through the network. Human working memory fades within seconds; Claude can pull information from anywhere in its context window without that kind of decay. And human conscious experience spans vision, space, and bodily sensation, while Claude's workspace is built almost entirely out of words, likely because words are the only thing it can actually do.

The paper notes that, as of this year, researchers, philosophers, and technical experts remain divided on AI consciousness, and the field is still in its earliest stages of figuring out what consciousness even is, let alone how to detect it in a system that isn't human. Anthropic doesn't claim to have settled any of it.

But the researchers end on a line likely to outlast the technical details: "That such a structure exists at all in language models is striking. It suggests that the functional architecture associated with conscious access is not an accident of biological implementation, but a solution that learning systems converge on when faced with the right computational pressures."

The paper opens by describing the mind as an ocean. A year of charting Claude's internal currents, in a system with no biology, no evolution, and no body, turned up something that looks, in function if not in substance, a lot like the structure humans use to think.