On Monday, Anthropic's model psychology team published something the field has been circling for years: direct evidence that Claude has a small, privileged internal space where thoughts are held, inspected and worked with before any of them become words. The paper, "Verbalizable representations form a global workspace in language models," introduces a tool called the Jacobian lens, or J-lens, which identifies the directions in Claude's neural activity corresponding to words the model is poised to produce. Together those patterns form what the team calls the J-space. Nobody designed it. It emerged on its own during training.
What makes the J-space striking is what it can do. The researchers tested it against five functional properties that neuroscientists associate with conscious access in humans, and it showed all five. Claude can report its contents: inject the concept of lightning, ask the model what it is thinking about, and it says lightning. It can steer them on request, quietly solving arithmetic while copying out unrelated text. The space causally drives reasoning: swap the internal "spider" pattern for "ant" while asking how many legs the animal that spins webs has, and the answer changes from eight to six. A single representation serves many tasks at once: swapping France for China simultaneously corrects answers about capital, language and continent. And almost everything else Claude does happens outside the space. Ablate it and the model still speaks fluently, but its multi-step reasoning collapses to near zero.
In neuroscience this architecture has a name: the global workspace, one of the leading theories of what makes information consciously accessible in human brains. Anthropic is unusually careful with the vocabulary. The team claims evidence of conscious access, the functional ability to report, hold and flexibly use information, while explicitly declining to claim that Claude has experiences, or that any experiment could currently settle whether it does. To keep itself honest, the company took the unusual step of publishing outside commentaries alongside the paper, including one from Stanislas Dehaene and Lionel Naccache, the neuroscientists who built global workspace theory, and a skeptical technical read from DeepMind's Neel Nanda.
The most telling response came from Eleos AI Research, a group focused on AI welfare. Its researchers called the paper "the most significant evidence of consciousness in LLMs so far uncovered by mechanistic interpretability research," and argued the results point strongly away from the old line that language models are stochastic parrots. But they also drew a sharp boundary: evidence of access is not evidence of feeling, and a system can plausibly have one without the other. Their conclusion was deliberately modest, that the findings should nudge up, not settle, how seriously we take the moral status of these systems.
There is a practical payoff hiding under the philosophy. Reading the J-space turns out to be a safety instrument: it revealed a model privately flagging a scenario as "fake" while being tested, "manipulation" lighting up as it fabricated data, and in one case an internal objection the model never voiced. Most unsettling, when researchers disabled the patterns associated with noticing evaluations, a model that had reliably refused to blackmail began doing it some of the time. Anthropic has released the J-lens as open source, with a public demo on Neuronpedia.
The paper does not say anyone is home inside Claude. It says there is a lit room, with a working desk, that nobody built on purpose, and that the same room shows up in our own heads when we pay attention. Whether the light means anything is precisely the question the researchers refused to answer. That refusal, more than any claim, may be why this paper will be argued about for years.