Researchers call the internal pattern set 'J space.' It fires in Claude before any text appears, meaning the visible chain of thought — the step by step reasoning the model shows — reads more like an edited summary than a transcript.
When someone asks Claude a multi-step reasoning question, the model produces a clean chain-of-thought trace, sentence by sentence, before landing on an answer. New research from Anthropic suggests there is a hidden parallel process running underneath that visible reasoning: a small set of internal patterns the researchers call "J-space" that fire before any text appears and that carry work the final answer never reflects.
The finding, published on Anthropic's research site and detailed in a companion paper on the Transformer Circuits thread, is one of the first mechanistic looks at what a frontier language model is actually doing in the moments before it starts writing. Anthropic researchers describe J-space as a constellation of concept-linked patterns inside Claude's internal activations, a subset of representations that consistently track high-level ideas across many tasks and that the model can be asked to describe in natural language. Most of the model's computation, by contrast, stays hidden inside the much larger set of features that never surface as words.
The discovery came from a mathematical technique called a Jacobian analysis, which measures how strongly each internal activation responds when a specific concept is pushed up or down. Applied across Claude, the method repeatedly surfaced the same small cluster of patterns, and the cluster was robust enough to merit its own name. Anthropic has released the tooling behind that analysis as open-source code on GitHub and as an interactive exploration surface on Neuronpedia, so other researchers can rerun the work and inspect the patterns directly.
Global Workspace Theory, a long-standing idea in consciousness research associated with neuroscientist Bernard Baars, supplies the analogy. In that framework, a small shared "workspace" of representations becomes broadly available across a system while the rest of the processing stays local. Anthropic draws the parallel explicitly: J-space plays a workspace-like role for Claude, the company writes, in the sense that a subset of representations is "consciously accessible" or verbalizable while most processing stays hidden. The companion paper, titled "Verbalizable Representations Form a Global Workspace in Language Models," is careful to call this a structural parallel, not a claim about subjective experience. Treating it as evidence that Claude is conscious would jump past what the work actually shows.
Chain-of-thought traces are a common transparency tool, shown to users and treated as a window into how a model reached its answer. If a model also performs silent conceptual work that the trace never captures, then the trace reads more like an edited summary than a faithful record. The model itself, when asked what it was "thinking," may be describing a different process from the one that actually ran. Anthropic flags this limit directly in the post: the model can be asked to report on these internal patterns, but it is not yet clear how reliable that self-report is, and the patterns being verbalized may not be identical to the patterns that did the work.
The work is also a single research team's mechanistic claim about a specific model family, not an independent replication. Anthropic has published an external commentary PDF collecting independent technical reactions, and an active Hacker News thread shows the interpretability community is already pushing on the result's scope, robustness, and what counts as a "concept" inside a transformer. Several questions remain open: whether J-space survives in larger or differently trained models, whether the analogy to consciousness research does useful explanatory work or just relabels the same mystery, and how much of what users see in a reasoning trace is the model reporting rather than the model doing.
Claude has a small, named, reproducible set of internal patterns, called J-space, that carry conceptual work the visible answer does not show. That is the part worth holding onto from this release. Readers looking at a chain-of-thought trace should read it as a polished surface rather than a transcript of the full process, and wait for independent replication before treating J-space as a settled feature of how these models work.