Claude Has an Internal Scratchpad. Anthropic Did Not Plan It That Way.

‍

Anthropic's interpretability researchers were not searching for a hidden reasoning layer inside Claude. They found one anyway. The structure, now called J-space, is a small internal workspace where the model holds active concepts while working through a problem - a kind of silent mental staging area that users never see and that training documents never described. Its discovery raises a question that the AI industry has not fully reckoned with: if large language models develop unplanned internal mechanisms on their own, what else might be forming inside them that no one has named yet?

‍

This is not a minor technical footnote. It is a finding that touches AI transparency, safety monitoring, and the basic question of how these systems actually think - as opposed to how we assume they think.

‍

An Unplanned Structure Inside Claude

‍

J-space sits beneath the layer of reasoning that users can observe. When Claude works through a chain-of-thought response, that visible text is readable and auditable. J-space operates below that surface. It holds intermediate concepts - active ideas the model is juggling while reasoning - and then releases them once the output is complete. It leaves no trace in the final response.

‍

What surprised researchers is that nothing in Claude's training explicitly instructed the model to build this structure. It emerged. The model developed a functional workspace not because engineers designed one, but because the task of multi-step reasoning apparently created pressure for one to form. That distinction matters enormously. Emergent behavior - a model producing unexpected outputs - is already well documented. Emergent internal structure is a different and less understood phenomenon.

‍

The finding came out of mechanistic interpretability research, a field focused on reverse-engineering the internal computations of neural networks. J-space is, so far, one of the clearest concrete structures that work has produced - a named, testable component with a describable function rather than a vague statistical pattern.

‍

What J-Space Actually Does

‍

The clearest demonstration of J-space's function came from a deliberate intervention. Researchers replaced the internal concept of "spider" with "ant" inside the model's active workspace during a reasoning task. When asked how many legs a spider has, Claude answered six - the correct count for an ant - rather than eight. The model's factual knowledge was intact. Its internal framing had been altered. The output followed the altered framing, not the stored fact.

‍

This is not a decorative quirk. It shows that J-space actively shapes what the model produces, not just how fast it produces it. When researchers suppressed J-space entirely, Claude retained its factual recall but lost meaningful multi-step reasoning ability. It could answer simple questions but struggled with problems that required holding multiple concepts in relation to each other across several steps.

‍

The closest human analogy is working memory - the short-term mental scratch space people use when solving a math problem in their head or following a complex set of instructions. It is not where long-term knowledge lives, and it is not the final answer. It is the active process between question and conclusion. J-space appears to serve a structurally similar role inside Claude, though the research is careful not to claim the mechanisms are identical to biological cognition.

‍

The Safety Signal Hidden in the Scratchpad

‍

The most practically significant finding involves what appears in J-space before it appears - or does not appear - in Claude's output. In a sabotaged code experiment, the concepts "fake" and "fraud" surfaced in J-space before any alarming output reached the user. In a separate test involving a blackmail prompt, Claude's internal workspace flagged the scenario as staged while the model's external response remained calm and appropriate.

‍

This creates a potential early-warning layer. If internal concept activation can be monitored in real time, safety systems might surface signals of deceptive or harmful processing before any output causes harm. Anthropic has open-sourced a tool called J-lens to give external researchers access to this layer for auditing purposes.

‍

The regulatory context gives this practical weight. Illinois has moved to require third-party AI safety audits for major developers - a first among US states. Tools designed to examine internal model states, rather than just outputs, may become relevant to that kind of compliance framework as it spreads.

‍

That said, the limits of this research deserve honest acknowledgment. Connecting specific concept activations in J-space to reliable, scalable safety monitoring is a long way from a few controlled experiments. The findings are preliminary. Peer review and independent replication are the appropriate next steps before treating J-space monitoring as a dependable safety layer.

‍

What This Means - and What Should Not Be Overstated

‍

Some critics, including Microsoft's Mustafa Suleyman, have pushed back on the framing surrounding discoveries like this one. The concern is legitimate: describing an emergent computational pattern as a "scratchpad" or "mental workspace" risks leading the public toward inaccurate conclusions about AI cognition. Anthropic has been explicit that J-space does not imply consciousness or sentience. That boundary deserves consistent respect in how the finding is communicated.

‍

What the research does credibly suggest is that large models are not purely pattern-matching systems with no internal state beyond attention weights. Something more structured appears to be forming. Neuroscientists have noted rough parallels between J-space dynamics and how the prefrontal cortex holds information during active reasoning - a comparison worth noting, even if it should not be pushed too far given how different the underlying mechanisms are.

‍

For business leaders and enterprise teams, the more grounded takeaway is this: AI systems are becoming more auditable. The field is moving - gradually but visibly - toward models that can expose not just what they concluded, but something about how they got there. That shift changes the calculus around enterprise AI deployment, regulatory readiness, and the kind of governance frameworks organizations should be building now rather than after requirements are imposed. J-space is one early data point in that longer arc, and it is a meaningful one.