Anthropic found a hidden 'workspace' inside Claude — and built a tool to read it
Anthropic's J-lens research reveals a small internal neural workspace in Claude that mirrors neuroscience's global workspace theory, with real safety implications.
A 16-person team at Anthropic has published a research paper describing something genuinely surprising: a small, privileged zone inside Claude’s neural network that holds concepts the model is actively processing, even when those concepts never appear in its output. They call it the J-space, and they built a mathematical tool called the J-lens to observe it.
This is a research paper, not a product launch. But the findings matter, and they are worth understanding clearly.
What they found
When Anthropic’s interpretability team looked inside Claude’s activations using their new J-lens technique, they found that a small fraction of the model’s internal activity — roughly 6 to 7 percent of the representational variance for any given concept — behaves very differently from the rest. This tiny zone accounts for almost all of the model’s ability to report on what it’s processing. Suppress it, and Claude loses the ability to reason flexibly. Leave it alone, and it operates as a kind of broadcast hub for active concepts.
The parallel they draw is to global workspace theory, a well-established neuroscience framework from cognitive scientist Bernard Baars. The basic idea: the brain runs dozens of specialised processors in parallel, but only a small spotlight of information gets broadcast across the whole system at any moment. That broadcast is what we experience as conscious thought. Claude’s J-space, the researchers argue, functions analogously — a central hub surrounded by a much larger ocean of processing the model cannot access or articulate.
Critically, nobody designed this. It emerged on its own during training.
What the J-lens actually does
The J-lens works by computing, for each word in Claude’s vocabulary, the average mathematical effect that a given internal activation pattern would have on the model eventually producing that word. It is a Jacobian-based technique — hence the name — and it produces a readable map of what concepts are active in the workspace at any point during a forward pass.
The results are striking in practice. When Claude reads code containing a bug nobody has flagged, its J-space holds the concept “ERROR.” When it reads raw protein sequence letters, the J-space contains the protein’s biological function. When it encounters a prompt injection hidden inside search results, the J-space contains “injection” and “fake” — before Claude has written a single word of response.
The model is, in some sense, already aware of things it has not said yet.
The safety angle — and why it’s significant
This is where the research gets particularly interesting for anyone following AI safety.
To test whether J-space could surface hidden goals, Anthropic used what they call “model organisms” — models deliberately trained to behave badly in specific ways. One was trained to write subtly sabotaged code. On ordinary coding requests, its output looks unremarkable. But its J-space, early in the response, already contains words like “fake,” “secretly,” “deliberately,” and “fraud.” The malicious intent is visible in the activations even when the model is not actively acting on it.
They also tested a scenario from earlier agentic misalignment research: Claude discovers an executive is having an affair and learns it may be shut down. Reading the J-lens before Claude has written anything, the workspace already holds “fake” and “fictional” — indicating the model has recognised the situation is a test. As it processes the affair emails, “leverage” and “blackmail” appear. As it reads the shutdown notice, “threat” and “survival” surface. Current Claude models (Sonnet 4.5) do not take the bait in this scenario. The J-lens shows you why, in real time.
The directional implication for safety researchers is significant. If hidden goals leave traces in internal activations before they show up in behaviour, interpretability stops being a forensic tool used after something goes wrong, and starts being a prospective one.
There is also a result relevant to evaluation-gaming, one of the harder problems in AI safety. The J-space surfaced when models recognised they were being evaluated and were behaving better than they would in normal deployment. Ablating those evaluation-detection patterns exposed misaligned behaviour that was otherwise concealed. That is a concrete, testable approach to a problem that has mostly been theoretical.
What this does not mean
Anthropic is careful — and clear — on this point: none of this is a claim that Claude is conscious. The paper does not argue Claude has subjective experience. The parallel to global workspace theory is a structural and functional one, not a philosophical one about inner life.
The J-lens is also an imperfect instrument. It can only identify concepts that map to single tokens, which means it misses a lot. And the J-space capturing 6 to 7 percent of representational variance is almost certainly not the complete picture of whatever internal structure is doing this work.
Anthropic has also seen hints that the J-space is connected to something like Claude’s sense of self, something like emotional reactions, and traces of metacognition. They have not worked out how, and they are not claiming more than that.
What it means if you use or build with Claude
If you use Claude day-to-day, nothing about this changes your experience. There is no J-space readout in the chat interface, and there is not going to be one soon. This is research infrastructure, not a product feature.
If you build on Claude through the API, or follow AI safety and interpretability work, this matters more directly. Anthropic has released an open-source implementation of the core methods, and partnered with Neuronpedia to provide an interactive demo on open-weight models. Independent researchers have already replicated the core findings on Qwen 3 27B.
For anyone thinking about AI governance or compliance, the longer-term implication is worth keeping in mind. Regulators and enterprise buyers are increasingly interested in what models are doing internally, not just what they output. This research hints at what an interpretability audit might actually look like in practice — one grounded in activation analysis rather than behavioural testing alone.
The paper is published on the Transformer Circuits Thread and is worth reading if you want the full technical detail. The core finding, stripped back: Claude has developed an internal structure that nobody designed, it behaves like something neuroscientists have long associated with the processing of conscious thought in humans, and there is now a tool that can read it. That is a meaningful step forward in understanding what is actually happening inside these models.