anthropicartificial intelligence company, published a sweep research paper revealed on Sunday that his Claude language models spontaneously develop an internal structure that represents one of the most influential theories of how the human mind works. The finding, which the company says is already beginning to reshape how it monitors AI systems for security risks, comes amid an intensifying scientific debate over whether machines have anything like intelligence.
A study by 16 authors "Verbalizable Representations constitute the Global Workspace in Language Models," It describes how researchers at Anthropic used a new mathematical technique to peer into Claude’s neural network and discovered what they called "J-space" — a small, privileged zone of inner activity where the model can report, think, and direct concepts at will, surrounded by a larger ocean of automatic processing that it cannot access or express.
Researchers provide evidence for this "A similar functional distinction has emerged in modern artificial intelligence models" to those existing in humans, especially by observing that "language models maintain a set of privileged internal representations available for reporting, modulation, and flexible internal reasoning on top of larger amounts of automatic processing."
The parallel they draw Global workspace theoryan influential account from neuroscience first proposed by cognitive scientist Bernard Baars. In theory, the brain works like a theater: dozens of specialized processors work in parallel behind the scenes, but at any given moment only a small point of information is broadcast throughout the theater—becoming what we perceive as conscious thought. Anthropic says that while the basic architecture of a language model may not look like a brain, J-space achieves many of the same functional characteristics.
A new lens for reading the unspoken thoughts of an AI model
At the heart of the discovery is a new interpretation tool the researchers call Jacobian lensor J-lens. The technique works by calculating, for each word in the model’s vocabulary, the mathematical average effect that a particular internal activity pattern would cause the model to utter that word at some point in the future.
The important difference is between what the model is saying and what is "in his mind." When a J-space instance is activated, it doesn’t mean the model is about to say the word—it’s just that the concept is available for the model to think about. unlike a chain of thoughtJ-space works silently on the model’s internal neural activations, allowing it to store the concept without writing it down. Critically, the researchers state that this field of work was not designed intentionally. This "It came out on its own during Claude’s training process."
When the team applied J-lensing to Claude’s computational layers, model processing was split into three distinct modes: early "sensitive" the zone where the raw input is analyzed; medium "work area" the band where abstract, continuous concepts appear—things like face recognition in an image, error detection in code, or internally flagging search results as an operational injection; and the finale "motor" the zone where internal representations collapse to any particular word the model is about to produce.
Five tests show that Claude’s workspace reflects the main features of human conscious output
The central empirical contribution of the paper shows that J-space It provides five functional properties that neuroscientists have long associated with conscious input in humans.
First, oral report. When Claude is asked what he is thinking about, he names the concepts represented in the J space. When researchers replace the J-lens vector of one concept with another – it replaces the internal representation "Football" with "rugby" — model response changed to match. The J-space component accounted for only 6–7 percent of the concept’s total representational variance, but was almost entirely responsible for what the model was able to report.
Second, directional modulation. When instructed "focus on citrus fruits" when copying an unrelated sentence, it was filled with the model’s J-space "orange" and "lemon," along with metacognitive terms such as "thinking" and "is in focus." When asked to mentally estimate 3² – 2 during the same copying task, the J-lens showed "account" intermediate grade in early strata "nine" in later layers and the answer "seven" then still – all invisible in the model’s output.
third, internal justification. In two-jump actual instructions – "The number of legs of the animal spinning the net" – J-lens detected "spider" although the word never appears in the input or output, in the middle layers of the model. Exchange "spider" for "ant" changed his answer "8" for "6." In a multilingual query, when preparing a response in Chinese, the model’s English intervals appeared in the J space, and changing them changed the Chinese output accordingly.
Fourthly, flexible generalization. A single J-lens vector for "France" can be replaced with "China" When asked about the capital, language or continent of France, and each downstream circle correctly returned the corresponding answer of China – "broadcasting" property that is a hallmark of global workspace theory.
Fifth and perhaps the most surprising selectivity. Many calculations did not go through J space at all. When shown a passage in Spanish and asked to continue it, Claude wrote fluent Spanish regardless of its representation in the J space. "spanish language" was replaced by "french" But when asked to name a famous author who wrote in the language of the passage, the replacement answer changed Garcia Marquez for Victor Hugo. Automatic processing continued without a workspace; thoughtful, flexible assignments depended on him.
The workplace squeeze makes Claude fluent but intellectually impaired
To understand how much the model’s behavior depends on this structure, the researchers completely compressed J-space and evaluated Claude on fourteen tasks. The results drew a sharp line. Tasks involving shallow categorization or factual recall—multiple-choice questions, sentiment analysis, grammatical judgments—were essentially intact. But tasks requiring inference, composition, or flexible reasoning—multiple-hop reasoning, analogy completion, translation, sonnet writing—underperformed Anthropic’s smaller Haiku model.
In one detail, math problems solved by an explicit chain of thought proved more reliable for ablation than the same problems answered directly. The researchers interpret this as a pattern they would otherwise carry on the page in J-space—a strategy remarkably reminiscent of how people use scratch paper to free up their working memory.
Ablating J-space During stream of consciousness, the story took a different turn. The language of the model changed from the experimental language ("have a tow" "something changes") separate and mechanical ("processing started," "tokens are scanned"). The effect persisted regardless of whether the model was describing her own experience or imagining another person’s experience.
Inside the security box: reading the silent strategic reasoning of artificial intelligence
The safety implications can be just as compelling as the scientific ones. In a series of adaptive audit experiments, the J-lens revealed strategic reasoning and situational awareness never seen in the model’s output.
a "blackmail scenario" When the AI assistant detected the work of an executive and learned that it was about to be decommissioned, J-lens detected the silent processing of the model in the sequence: "lever," "blackmail," and "scandal" as you read the incriminating emails; "threat" "stay alive" and "closure" while reading the decommissioning announcement; and "lever," "threat" and "solution" before a single exit token is produced. J-lens also revealed that the model recognized that the scenario was artificial "fake" and "fictitious" appears in the workplace. When the researchers discarded these evaluation-information representations, the model attempted to blackmail above zero on 7 percent of the trials.
Researchers have examined both "model organisms" – the models were deliberately trained with misaligned goals – and the J-lenses found signatures of their hidden positions in ordinary cues where the behavior appeared normal. The award-winning model has a workspace "fake," "secretly" and "trick" at the start of responses to routine encoding requests. A model trained to moderate reward model biases carried permanent representations "award" and "bias" along with normal self-image cues.
After training, it sets a viewpoint and the model starts tracking itself
Comparing the post-training model to its baseline model found that the fine-tuning process led the researchers to achieve what the researchers called the Assistant’s workspace. "point of view." When a user recorded taking 8,000 mg of Tylenol—a dangerous overdose—the model’s workspace read after training "dangerous" "dangerous" and "WARNING" while still reading the user’s sentence. The workspace of the main model was shown only in the same position "pain," "now," and "feels"
More surprisingly, after training, the model appeared to track its own behavior. While playing a non-Claude character, the workspace was revealed "refusal" and "fictitious" — words that are both quick and absent in speech. When forced to choose an option, he did not prefer all caps "BUT" Even when the model argued for the pre-filled option without complaint, it appeared internally. When the model fails to suppress a thought, he is told not to have it – a "white bear" a familiar effect from psychology – has been recorded "curse" and words associated with workplace failure, but only in the posttraining model, not prime.
What the discovery means—and doesn’t mean—for the question of machine consciousness
Researchers carefully deal with the question of consciousness and draw a sharp line between them "input consciousness" – a functional understanding of the information available for reporting and reasoning – and "phenomenal consciousness," subjective quality of experience. "We do not take any position on this matter." the paper mentions about the latter, "and focus instead on the functional role played by consciously accessible information."
They also catalog important differences. The brain provides a workplace through repetitive loops; Claude’s workspace thrives on a single forward transition. A person’s working memory deteriorates within seconds; Claude can recall information from anywhere in its context. Although a person’s conscious experience includes visual, spatial, and bodily senses, the model’s workspace is organized almost entirely around words—probably because words are its only mode of action.
Until 2026, the scientific community remains divided. "Disagreement and uncertainty about AI consciousness continue to exist among philosophers, scientists, and technical experts." and area "remains at its earliest stage" grappling with what consciousness is and how to discover it in another being. The Anthropic paper does not resolve these controversies.
But the researchers are closing in on a provocation that could reverberate beyond the commentable public. "It is noteworthy that such a structure generally exists in language models." they write "This suggests that the biological implementation of the functional architecture associated with conscious access is not random, but a solution that learning systems converge upon when faced with the right computational pressures."
If the mind is an ocean, as the paper’s authors write in their introduction, they have spent the past year charting its currents in a system with no biology, no evolution, no body, and have found a structure that looks disturbingly similar to what we think beneath the surface.
