Anthropic researchers find a "workspace" inside Claude where its reportable thoughts live

What is actually happening inside a language model when it "thinks"? On July 6, 2026, Anthropic published interpretability research suggesting a surprisingly structured answer: Claude appears to contain a small, privileged internal space that functions like a mental workspace — a staging area for the thoughts the model can hold in mind, use for multi-step reasoning, and report when asked.

J-space and the J-lens

The researchers built a new analysis tool they call the Jacobian lens, or J-lens, and used it to probe the flood of internal activity inside Claude. Most of that activity is automatic and inflexible — vast machinery that grinds along without anything resembling deliberation. But a small subspace, which the team dubbed J-space, behaves differently. Contents placed there can be reported by the model, held across steps, recruited for deliberate reasoning, and sometimes detected before they surface in the final answer.

If that sounds familiar, it should. It closely mirrors the "global workspace" theory from cognitive neuroscience, which proposes that human consciousness works like a broadcast stage: many specialized brain processes run in the dark, and only information promoted to a shared workspace becomes available for reasoning and report. Stanislas Dehaene and Lionel Naccache, the neuroscientists behind the leading version of that theory, contributed external commentary to the paper.

Independently checked, carefully framed

Notably, the finding did not stay in-house. Neel Nanda, who leads language-model interpretability at Google DeepMind, provided commentary that included an independent replication of some results on an open-weight model — meaning the structure is not a quirk of one company's system.

The researchers are careful about what they are not claiming. Finding an architecture reminiscent of theories of consciousness is not evidence that Claude is conscious. The practical significance is more concrete: it gives researchers a specific place to look when they want to know what a model is "aware of" in a functional sense — a potential handle for catching a model's intentions before they show up in its output.

Why interpretability keeps mattering more

This lands in a year when understanding model internals moved from academic curiosity to industry priority — MIT Technology Review named mechanistic interpretability one of its breakthrough technologies for 2026. As AI systems take on more autonomous work, the ability to inspect their internal state is becoming the foundation for real safety auditing rather than take-our-word-for-it assurances.

Why it matters for small business

No workflow changes on Monday — but this research points at where trust in AI tools is heading. Today, when an AI assistant gives you an answer, you have no way to know what it considered or concealed along the way; you judge the output and hope. Work like this is building toward tools that can verify what a model was actually doing — which will eventually show up as better audit features, more honest uncertainty, and AI products that can be certified for sensitive work like bookkeeping, contracts, and client data. When you evaluate AI vendors over the next few years, "how do you know what your model is doing?" is about to become a fair — and answerable — question.