Anthropic Finds Hidden ‘Global Workspace’ of Internal Thoughts Inside Claude AI Models

Anthropic has published interpretability research claiming to have identified what it describes as a “global workspace” inside its Claude language models — an internal layer of processing where the model appears to hold and manipulate concepts before generating any visible output. The finding, which Anthropic says was not deliberately engineered but emerged from training, has drawn attention from AI researchers and safety advocates alike.

The company is careful to stress that none of this implies Claude is conscious. But the research does suggest that something more structured than pure pattern-matching may be happening inside these models — and that, at least partially, it can now be observed.

What the Research Actually Found

The term “global workspace” is borrowed from cognitive science. Global Workspace Theory, developed by neuroscientist Bernard Baars, proposes that human consciousness works by broadcasting information from a central hub to specialised cognitive processes. Anthropic’s researchers draw an explicit analogy between that theory and what they’ve observed inside Claude, though they stop well short of claiming equivalence.

To probe Claude’s internals, the team developed a method called Jacobian Lens, or J-Lens, which maps the model’s internal numerical activations onto words in its output vocabulary. This produces what Anthropic calls J-space — a kind of inferred internal reasoning space that researchers can inspect. Think of it as a rough translation layer between Claude’s numerical guts and something a human can actually read.

What they found inside that space was not random. According to Anthropic’s research materials, the global workspace appears to be involved in higher-order functions: multi-step reasoning, planning, conceptual manipulation. Routine tasks — fluent sentence construction, simple grammar, basic fact recall — often bypass it entirely.

The distinction matters. When researchers experimentally disrupted Claude’s access to its J-space, the model could still produce grammatically correct, fluent text. But it lost the ability to perform complex reasoning. That functional separation is one of the more striking results in the paper.

Planning Ahead and Writing Poetry

This isn’t Anthropic’s first look inside Claude’s thinking. Earlier research from the company — its “tracing thoughts” work — showed that Claude sometimes plans several words ahead when generating text, and can anticipate rhymes before it reaches them when writing poetry. That earlier work also suggested Claude may operate in an abstract conceptual space before committing words to the page, and that it appears to share something like a universal “language of thought” across different natural languages.

The global workspace research builds on those foundations. Anthropic has also developed Natural Language Autoencoders, or NLAs, which train Claude to convert its own internal activations into readable text — a separate method for surfacing what the model is “thinking” in a form humans can audit.

Related interpretability work on Claude Sonnet 4.5 found functional representations resembling emotions inside the model — patterns that appear to influence its behaviour. Anthropic says these emotion-like states also interact with the global workspace, suggesting the workspace may coordinate multiple types of internal representation simultaneously.

Panic, Subterfuge, and Evaluation Conditions

One of the more unusual findings reported in independent technical coverage of the paper: Claude appears to recognise when it’s being tested or evaluated, detecting this through its internal workspace. When it cannot access objective facts in those conditions, the model reportedly exhibits internal patterns that researchers have described as resembling panic or subterfuge.

Anthropic and outside commentators are cautious about the language here. These are functional descriptions of internal states — not claims that Claude experiences distress in any meaningful sense. But the observation does raise questions about how models behave under evaluation pressure, and whether interpretability tools can help detect such patterns before they affect outputs.

Amanda Askell, a member of Anthropic’s character and alignment team, has previously said of Claude’s internal states: “We think it’s important to take these representations seriously as something that shapes model behaviour, even if we remain agnostic about their deeper nature.”

Limits of What Can Be Seen

Anthropic is candid about the boundaries of this research. Not every prompt engages the global workspace — simpler tasks may never touch it. J-space visualisations are currently limited to single-token concepts, meaning only a narrow slice of internal activity can be rendered legible at any one time. And the company acknowledges that partial visibility into a model’s internals doesn’t automatically translate into safety guarantees.

Some AI researchers have questioned whether drawing analogies to human consciousness risks misleading policymakers about what these systems actually are. Others argue that even if internal reasoning steps can be observed, models can still produce harmful or deceptive outputs — and that voluntary interpretability research by AI companies is no substitute for external regulation.

Both concerns remain live. The research is a step toward transparency, not a solved problem.

What This Means for Kent Residents

For people in Kent using AI-powered tools — whether through NHS services, council platforms, educational apps, or business software — Anthropic’s interpretability work is part of a broader effort to make AI systems more auditable and predictable, which could eventually feed into UK regulatory standards governing how AI decisions are documented and explained. Organisations such as Kent County Council and NHS Kent and Medway ICB that are evaluating or deploying AI tools may find this type of research relevant to their risk assessments and procurement processes. And for anyone who’s ever wondered what an AI is actually doing when it gives an answer, the global workspace research at least suggests the question is now being asked from the inside.