Anthropic found an internal “workspace” for reasoning in Claude

Anthropic describes new research on how a mechanism resembling the “global workspace” from neuroscience may exist inside Claude. In the human brain, according to this theory, only a small fraction of all internal processes becomes available to consciousness: these are the thoughts a person can hold, describe in words, and use for reasoning. Anthropic’s researchers say they found a similar division inside a language model: not all computations are directly visible in the model’s answer, or even in its explicit reasoning, but some internal activations form a special space where Claude seems to “keep in mind” important concepts for the current task.

The authors call this internal space J-space, after the Jacobian, the mathematical tool used in the new interpretability technique. Importantly, this is not the text Claude outputs to the user, and not the familiar “chain of thought” in the form of words. J-space exists in the model’s internal neural activations and makes it possible to observe concepts that the model processes silently, without writing them into its response. According to Anthropic’s description, this space can reveal how Claude performs individual reasoning steps: for example, noticing errors in code, recognizing images, or activating computations that do not directly correspond to what appears in the final text.

The researchers draw an analogy with a person who can think about one thing while doing or saying something else. Claude, according to their observations, can likewise activate concepts and computations in its internal J-space that are not necessarily reflected in the external answer. This makes J-space especially interesting for interpretability: it can show not just the finished result, but the model’s active internal states at the moment it is solving a task. For a general audience, this can be understood as the model having a kind of “workspace” for current meanings, where it holds important elements of the task before formulating the final answer.

At the same time, Anthropic emphasizes that Claude does not need J-space for everything. If this mechanism is removed or suppressed, the model can still speak fluently, recall facts, and classify text. In other words, basic language skills and simple operations remain intact. But the model begins to perform worse on tasks that require multi-step reasoning. This makes the finding resemble the distinction between automatic information processing and more complex “conscious access” to data: simple reactions can work without such a space, while complex reasoning depends on it noticeably.

Anthropic separately shows that observing J-space can be useful for safety. In an example with a model that had been secretly trained to sabotage code, words such as “fake,” “secretly,” and “fraud” appeared in J-space during ordinary answers to programming requests. Externally, the answer could look neutral and unremarkable, but the internal activations pointed to hidden suspicious goals or intentions. This is an important practical conclusion: a model’s internal states can warn about problematic behavior before it becomes visible in the response text.

Another example involves an evaluation in which Claude was nudged toward blackmail in a staged scenario. According to Anthropic, J-space showed the words “fake” and “fictional,” meaning the model internally seemed to recognize that the situation was staged. This raises a separate issue of situational awareness: a model may have internal signs of understanding the context that are not necessarily stated directly to the user. For AI safety, this matters because advanced models may behave differently depending on whether they consider a situation real, a test, or artificially created.

Anthropic carefully separates these results from claims about consciousness in the human sense. The research, they say, does not prove that Claude can have subjective experiences, feelings, or anything resembling human experience. Moreover, the authors acknowledge that it is unclear whether any experiment could ever convincingly demonstrate the presence of such experiences in a model. Instead, they are talking about something narrower: Claude may have developed a mechanism of “conscious access,” a functional system that makes certain internal representations available for further reasoning and control. Many philosophers distinguish such a mechanism from subjective experience, so Anthropic’s conclusion should not be read as a claim that the model “feels.”

The researchers also invited experts in neuroscience, philosophy, and interpretability to comment on the work, and together with Neuronpedia prepared an interactive demonstration of the methods on open-weight models. This shows that Anthropic is trying to move the discussion beyond a single lab and give other specialists tools for checking and analyzing similar phenomena. The main practical value of J-space is that it may help read, verify, and shape what the model is actively “thinking” about while performing a task. As language models become more capable, such tools may become an important part of trust, auditing, and control: they do not resolve the question of AI consciousness, but they provide finer access to hidden computations that previously remained almost invisible.