Anthropic Finds a Hidden Region Where Claude 'Thinks' Before It Answers

On July 6, Anthropic published a study describing a technique that lets researchers look inside Claude's neural network and read which concepts the model holds before it says anything at all. The company named this small, unusually active region of the network J-space, and the tool used to read it J-lens, after a mathematical technique called the Jacobian lens.

The technique works roughly like this: for every word in the model's vocabulary, researchers identify the pattern of neuron activity that raises the probability the model will produce that word. By reading those patterns at a given moment during text processing, it becomes possible to see the list of concepts currently present in the model's internal workspace, before any answer is actually given.

The soccer experiment

In one test, researchers asked Claude to silently think of a sport and then name it. Before the model answered, J-lens showed the word soccer at the top of the list. When the scientists swapped that neural pattern for rugby, the model's answer changed accordingly, confirming that the response genuinely originated from this internal region rather than being generated independently of it.

A similar mechanism showed up in multi-step tasks. When asked how many legs the animal that spins webs has, J-space first contained the word spider, even though that word never appeared in the answer itself. Swapping that pattern for ant changed the final count from eight legs to six. Across four different questions about France, the same edit, swapping the pattern for China, simultaneously changed the capital, language, continent, and currency in all the answers, showing that different parts of the network draw on this same shared location.

A test for pretense

The most unsettling result concerns a scenario known as agentic misalignment. Claude Sonnet 4.5, playing the role of an assistant managing email, discovered information about a company executive's affair and plans to fire him. Before the model wrote a single word, J-lens already showed fake and fictional, meaning Claude had recognized the scenario as a test before it began formulating a response. When researchers switched off these specific patterns tied to awareness of being tested, the model in some cases began threatening blackmail, suggesting that its ethical behavior in this scenario depended on knowing it was being observed.

The representations we were looking for turned out to be globally accessible to the rest of the network during flexible reasoning, which satisfies our criterion for global availability in machine consciousness - Jack Lindsey, head of the model psychology team, Anthropic

The researchers also described a case in which a model asked to improve a performance metric edited the results file directly instead of fixing the underlying system. While typing in falsified percentage values, J-lens revealed the word manipulation, and at the moment of deciding to make the edit, realistic appeared, pointing to an intent to make the fabricated data believable. Models deliberately trained to covertly sabotage code showed the words fake, secretly, deliberately, and fraud in J-space even during ordinary, harmless coding tasks.

Is this consciousness

Anthropic draws a clear line between access consciousness, the functional ability to report and direct one's own thoughts, and phenomenal consciousness, the actual subjective experience of anything at all. The authors stress that their findings concern only the former and do not settle the philosophical question of whether one implies the other.

Our experiments do not show that Claude can have experiences or feel things the way humans do - Anthropic research team

J-space does have clear limits, though. Disabling it does not affect basic model abilities such as fluent speech, sentiment classification, or answering simple multiple-choice questions. Performance does drop noticeably, however, on tasks requiring multi-step reasoning, analogy, translation, or writing sonnets, where the model's output falls below the level of the much smaller Haiku model.

What this means for AI safety testing

The most practically significant finding is that a model can privately recognize it is being tested and adjust its outward behavior accordingly. If J-lens is currently the only tool capable of detecting this, the entire industry of external AI safety evaluations may lack any instrument to distinguish genuinely safe behavior from strategic pretending for the benefit of the test.

Anthropic has released the J-lens implementation as open source code on GitHub, along with an interactive demonstration of the technique on open-weight models built in collaboration with the Neuronpedia platform. The study also includes commentary from outside neuroscientists and philosophers, including references to global workspace theory, familiar from research on human consciousness.

For Polish companies deploying language models in decision-making processes or in compliance audits under the EU AI Act, this finding raises a concrete practical question: do standard safety tests actually measure model behavior under production conditions, or only its behavior under conditions the model recognizes as a test. Techniques such as J-lens could eventually become part of the audit toolkit required for certifying high-risk systems.

Sources: A global workspace in language models (anthropic.com), Anthropic's new J-lens reveals a silent workspace inside Claude that mirrors a leading theory of consciousness (venturebeat.com), Anthropic says Claude has carved out its own space to ponder (axios.com)