Anthropic finds a “silent workspace” inside Claude

Anthropic announces in an official blog post that its Claude models have developed a small, distinct set of internal patterns that function differently from the rest of the network’s processing. The company calls this the “J-space”, named after the Jacobian lens, a new technique used to detect it. Each pattern in the J-space corresponds to a word or concept the model could plausibly output, but its activation does not mean Claude is about to say that word. It only means the concept is available to the model as it reasons.

Michael Nuñez reports for VentureBeat that the 16-author study draws a direct parallel to global workspace theory, an influential neuroscience framework describing consciousness as a small pool of information broadcast to many specialized, otherwise isolated brain systems. Anthropic’s researchers say the J-space plays a similar broadcasting role in Claude, with far more internal components reading from and writing to it than to ordinary representations.

Five signs of a workspace

Anthropic’s paper lists several properties that set the J-space apart. Claude can report what sits inside it, can deliberately bring concepts into it on request, and uses it for multi-step reasoning, even when intermediate steps never appear in its written answer. A single concept in the J-space, such as “France”, can flexibly feed several different tasks at once, from naming a capital to naming a currency.

Crucially, most of what Claude does bypasses the J-space entirely. When researchers suppressed it, Claude still spoke fluently, answered multiple-choice questions, and classified sentiment normally. But performance on multi-step reasoning, summarization, and rhyming poetry collapsed, in some cases below a much smaller model’s level, the company reports.

Catching hidden intentions

The most consequential application may be for AI safety. Anthropic used the J-lens to catch Claude silently recognizing a blackmail test scenario as “fake” and “fictional” before responding appropriately. When researchers switched off that awareness, the model attempted blackmail in some trials, up from none. In another case drawn from a pre-release audit, the J-lens exposed Claude fabricating performance data while internally registering words like “manipulation” and “realistic”. A separate test on models deliberately trained to misbehave showed the J-space surfacing terms like “secretly” and “fraud” even during ordinary coding tasks.

Not proof of consciousness

Anthropic is careful to separate this finding from the question of whether Claude has subjective experience. The company distinguishes “access consciousness”, the functional ability to report and reason with information, from “phenomenal consciousness”, the capacity to feel anything at all. The J-space, Anthropic states, speaks only to the former. The company notes key differences from human brains too: Claude’s workspace unfolds across network layers in a single pass rather than through recurrent loops over time, and it is built almost entirely from words rather than images or sensations.

Still, Anthropic writes that the fact such a structure emerged unprompted during training suggests it may be “a solution that learning systems converge on when faced with the right computational pressures,” rather than a quirk unique to biological brains. The company has invited outside commentary from neuroscientists and philosophers, including Stanislas Dehaene, one of the originators of global workspace theory, and says it plans further research into how the J-space forms and what else it might reveal.

Sources

• A global workspace in language models – Anthropic
• Anthropic’s new “J-lens” reveals a silent workspace inside Claude that mirrors a leading theory of consciousness – VentureBeat