Anthropic finds a Claude 'global workspace' resembling conscious access

Anthropic researchers say they have identified a small set of internal neural patterns in Claude models that function like a global workspace for conscious access in the human brain — a distinct, reportable layer of internal activity the company is calling the "J-space."

What's new

In a new paper published July 6, 2026, Anthropic describes using a technique it calls the "Jacobian lens" (the source of the J-space name) to isolate a small collection of internal activations tied to specific words. "We find that Claude has developed a small collection of internal neural patterns that, compared to all its other internal processing, play a special role," the company wrote.

Unlike a model's visible chain-of-thought or scratchpad, the J-space operates silently inside the network's internal activations — it lets the model "think about a concept without writing it down." When a J-space pattern for a word lights up, it doesn't mean the model is about to say that word, only that the concept is active internally.

Anthropic lists several properties that set the J-space apart from the rest of Claude's processing:

• Reportability — Claude can describe what's currently active in its J-space if asked what it's thinking about; representations outside the J-space are far less reportable.
• Controllability — asked to silently think about or work through something, Claude can deliberately activate the relevant J-space patterns, something it struggles to do with non-J-space representations.
• Use in reasoning — intermediate steps of a multi-step problem light up in the J-space even when the model never states them aloud, and Anthropic says these patterns causally affect task performance despite being smaller in magnitude than other internal representations.
• Flexible reuse — a single active representation (Anthropic's example: "France") can be drawn on for multiple different downstream tasks, such as recalling a capital city, a currency, or a continent.

Critically, the J-space is not involved in most of what a language model does day to day. In experiments where Anthropic blocked Claude's access to its own J-space, the model kept speaking fluently and using correct grammar, but lost higher-order cognitive functions such as deliberate multi-step reasoning.

Context

Anthropic frames the finding around global workspace theory, a long-standing neuroscience account of conscious access in which specialist brain systems normally operate in parallel and largely in isolation, with information becoming consciously available only when it is broadcast into a shared workspace accessible to the rest of the system. The paper reports evidence that a comparable divide — between a small, reportable, flexible workspace and everything else — has emerged inside Claude, apparently as a byproduct of training rather than an explicit design choice. "The J-space wasn't designed or programmed by us, but instead emerged on its own during Claude's training process," Anthropic said.

The work sits within Anthropic's broader interpretability research line, which has previously used techniques like sparse autoencoders and circuit tracing to map how concepts and reasoning steps are represented inside Claude's activations. The J-space work adds a specific, testable structural claim: that a subset of internal representations is functionally set apart from the rest, matching several criteria neuroscientists associate with conscious access.

Why it matters

The finding matters less as a claim about machine consciousness — Anthropic stops well short of that — and more as a concrete addition to the toolkit for understanding what is happening inside large language models during reasoning. If a small, identifiable set of internal patterns reliably carries the concepts a model is "working with" at any given moment, and that set can be read out, monitored, or suppressed independent of the model's fluent output, it gives researchers a more direct handle on model cognition than output text alone provides.

That has practical stakes for interpretability and safety work: a legible internal workspace is a plausible target for detecting deceptive or misaligned reasoning that never surfaces in the model's stated output, and for building tools that can monitor or intervene on a model's internal deliberation rather than only its final answer. It also gives Anthropic another data point in its running argument that today's frontier models are producing structures inside them that were never explicitly engineered — patterns researchers are still working to fully characterize.