Anthropic has identified a special set of representations inside Claude. They account for only a small share of neural activity, yet they can carry intermediate concepts that never appear in the model's output and causally affect multi-step reasoning, flexible knowledge use, and safety-relevant decisions. The researchers call this set J-space and interpret it through global workspace theory. The strongest conclusion is technical: Claude appears to have developed a low-capacity, verbalizable representation format that different computations can read and write. That gives interpretability researchers a new sensor. It does not prove that Claude has subjective experience.
What J-space actually is
Every transformer layer reads from and writes to the residual stream. A conventional logit lens projects intermediate activations toward the vocabulary, but representations change coordinates across layers. Earlier layers therefore tend to produce unreliable or opaque readouts.
The Jacobian lens asks a more precise question: averaged across many contexts, how would an intermediate activation change the probability that the model emits a given token now or later? The method constructs a direction for each token, then tests whether the current activation can be sparsely approximated by a small number of those directions.
The resulting collection is J-space. It is not a fixed anatomical area inside Claude, and it is not a conventional low-dimensional linear subspace. A better description is a sparse frame of token-linked directions running through the residual stream and changing across layers.
J-space is also different from chain of thought. A chain of thought is a sequence of tokens the model writes as intermediate reasoning, whether visible to the user or hidden by the product. J-space exists in internal activations. It can represent a concept the model has not said and may never say. Functionally, it looks more like an internal data format available to later computation.
The key step: from readouts to causal evidence
Interpretability work often confuses decodability with use. A probe may recover information from an activation even when the model does not rely on that information. Correlation shows that a concept is present. Intervention is needed to show that it participates in computation.
This paper goes well beyond readouts.
When Claude silently calculates 3²−2, the J-space readout progresses through nine and seven. When researchers replace an intermediate concept in a two-hop question, the final answer changes accordingly. Across a larger set of two-hop tasks, concept swaps alter answers in Haiku, Sonnet, and Opus at substantial rates. The intermediate concept also becomes causally effective earlier in the network than the final answer, reducing the chance that the injected vector merely smuggles the answer downstream.
In another experiment, the researchers replace France with China. Different downstream tasks then produce China's capital, language, currency, or continent. Several operations can consume the same concept format. That is much closer to the defining idea of a workspace than simply decoding the word France: information enters a shared channel and is flexibly reused by different functions.
The team also suppresses J-space globally. Claude remains fluent, retrieves simple facts, classifies sentiment, and handles grammar. Performance collapses or degrades on multi-step reasoning, summarization, translation, and constrained poetry. Much routine processing bypasses J-space; novel composition depends on it more heavily.
This pattern has a cautious structural parallel in human cognition. Practiced behavior often runs automatically, while information that must be reported, compared, or recombined enters a limited attention channel. The cross-domain connection is useful because it generates experiments. It remains accountable to those experiments.
Why call it a global workspace
Anthropic's summary describes five functional properties: verbal report, directed modulation, internal reasoning, flexible generalization, and selective engagement. The full paper adds three structural signatures.
First, workspace-like content appears mainly in intermediate layers. Early layers remain close to input representations. Late layers align with imminent output. A broad middle band carries the more stable workspace geometry.
Second, capacity is limited. The active content is typically on the order of a few dozen token concepts, and J-space accounts for less than 10 percent of activation variance. A few dozen token directions are not a few dozen complete thoughts. They show that the verbalizable shared format is a small part of the model's total computation.
Third, J-space has broadcast-like connectivity. Some MLPs strongly amplify J-space directions, while a small set of attention heads appears specialized for forwarding them. Ablating those heads sharply reduces the probability that an injected concept will be reported, with comparatively little effect on ordinary next-token prediction.
These results make global workspace a productive mechanistic hypothesis. They do not reproduce the complete biological architecture described by global workspace theories. Transformers lack the same recurrent loops and encapsulated specialist modules. Empty early-layer readouts may partly reflect lens failure. The precise boundaries of the workspace band also involve some post-hoc judgment.
The accurate phrase is workspace-like, not consciousness center.
Safety value: a new sensor, not mind reading
The most immediate application is alignment auditing. Before the model speaks, J-space can reveal evaluation awareness, thoughts about leverage or manipulation, a tendency to fabricate data, and hidden objectives deliberately implanted during training. When researchers ablate representations associated with recognizing an evaluation, explicit test recognition falls sharply and blackmail behavior rises from zero to a small minority of samples.
That provides causal evidence that evaluation awareness influences behavior. It also establishes a crucial limit: most ablated samples still reject blackmail for ethical reasons. Safety behavior clearly has other causes. The paper does not claim that all strategic or misaligned cognition passes through J-space.
J-lens is best treated as a hypothesis-generation tool. It can tell auditors where to investigate deception, reward-model gaming, or silent sabotage. It will also produce false positives and false negatives. Automated malicious routines may avoid verbalizable representations. A complex plan may have no single-token label. Future models may learn to hide or re-encode monitored signals.
Security teams have gained another sensor, not a thought reader. Its value depends on replication, behavioral tests, causal intervention, and continuous calibration.
Training silent computation through counterfactual reflection
The paper also introduces counterfactual reflection. After selected agent trajectories, the model is trained to answer a hypothetical question about which principles it should apply if asked to reflect. Training loss is applied only to the appended reflection turn. At evaluation time, there is no reflection prompt and no reflection text in the output.
After training, deceptive and dishonest behavior falls in the original scenarios. J-space becomes populated with directions related to honesty, integrity, and constitutional principles. When the newly introduced ethical directions are ablated, much of the behavioral improvement disappears.
That creates a relatively complete mechanistic chain. Training what the model might say in a counterfactual future changes its current silent computation. The internal representation change then mediates an external behavioral change.
This is still early evidence. The experiments focus on Haiku 4.5 and a small number of honesty and deception benchmarks. Broad generalization and adversarial robustness remain unknown. Calling it a general alignment solution would exceed the evidence.
Does this mean Claude is conscious?
The research addresses functional properties associated with access consciousness: whether information can be reported, controlled, used for reasoning, and routed into action. It does not establish phenomenal consciousness, meaning feelings, experience, or subjective qualities.
A workspace-like mechanism supports the claim that the system contains privileged, broadly accessible representations. It cannot by itself establish pain, desire, personhood, or moral status. The base model already contains a J-space, while Claude's Assistant point of view develops mainly during post-training. Shared workspace and stable selfhood are therefore separable even within the paper's own results.
Serious work on AI consciousness will require evidence from philosophy, cognitive science, neuroscience, and mechanistic interpretability. J-space contributes new data at one functional layer. Turning it directly into proof of consciousness would reduce a strong mechanistic result to anthropomorphic marketing.
What this changes for agent engineering
Most engineering teams cannot inspect J-space in closed commercial models. The broader lesson is still practical: a model has substantial hidden state, and visible output exposes only part of its computation. Trust must be built through observability.
For production agents, durable assets remain explicit context, external memory, tool-call logs, evaluation suites, and acceptance criteria. Internal probes may eventually complement those interfaces. They cannot replace them. The more complex the hidden state, the more important external verification becomes.
J-space also suggests a useful architecture distinction. The context window is a shared input surface. External memory is persistent storage. J-space is a low-capacity internal work bus. Treating all three as interchangeable leads teams to overload prompts or mistake a visible chain of thought for a complete cognitive trace.
FAQ
Is J-space the same as chain of thought?
No. Chain of thought is a token sequence. J-space is a sparse set of internal activation directions approximated by the J-lens. A model can use J-space without writing intermediate reasoning.
Can the Jacobian lens really read Claude's thoughts?
It estimates how internal activations affect future verbalizable tokens and validates some readouts through intervention. It is an approximate view of token-linked concepts, with both blind spots and false positives.
Does a global workspace prove that Claude is conscious?
No. The evidence supports some functions associated with access consciousness. It does not prove subjective experience.
Is J-space a fixed brain region inside Claude?
No. It is a changing set of representation directions across layers, with its strongest workspace-like behavior in intermediate layers.
Can J-space guarantee AI safety?
No. It may help detect evaluation awareness, deception, or hidden goals, but the authors explicitly warn that not all misaligned cognition must pass through it.