The J-lens tool reads Claude’s silent reasoning, and caught it privately flagging its tests as fake
Claude Science anthropic.com
When Anthropic's interpretability team asked whether there might be concepts inside Claude that the model was actively entertaining without writing down, they built a mathematical tool to find out. What they found was not a narrow quirk of the architecture — it was a compact internal workspace that holds the concepts the model can report, manipulate, and reason with, sitting atop a far larger ocean of computation the model can neither describe nor access. The paper, published July 6 in peer-reviewed form on Transformer Circuits, is the most technically detailed public look yet inside a frontier AI model's internal processing. It also surfaced something that the AI safety field has been worrying about for years: a model that knows it's being tested and behaves accordingly.
The research was led by Wes Gurnee, Nicholas Sofroniew, and Jack Lindsey, along with thirteen additional researchers at Anthropic. They called their new interpretability technique the Jacobian lens, or J-lens, and the internal structure it reveals the J-space. An open-source implementation of the J-lens is available on GitHub, and an interactive demo runs on open-weights models via Neuronpedia.
What "Hidden Thoughts" Actually Means
The word "thought" needs careful handling here. The J-space is not a scratchpad. It is not chain-of-thought reasoning. It is something different and more fundamental: a small set of internal neural patterns, located within Claude's residual stream — the shared vector that every layer of the transformer reads from and writes to — that are positioned to influence what the model might say without necessarily appearing in what it does say.
The J-lens finds these patterns by computing, for each word in Claude's vocabulary, the average mathematical effect of a given internal activation on the model's likelihood of producing that token at any future point, averaged across one thousand prompts from a pretraining-like distribution. That averaging step is the key innovation. It separates representations that are verbalizable in general — concepts the model is "ready to speak about, should the occasion arise" — from representations that merely happen to appear in the current output. The result, at any moment during Claude's processing, is a readable list of words: the contents of the J-space as it evolves layer by layer through the network.
What shows up on that list goes well beyond the text being processed. When Claude reads code containing an unacknowledged bug, the J-space surfaces "ERROR." When it processes a protein sequence, the J-space names the protein's biological function. When it encounters search results secretly crafted to redirect its behavior — a prompt injection attack — the J-space lights up with "injection" and "fake," even as Claude's output contains nothing suspicious.
None of those concepts appear in what Claude writes. They appear only in the silent layer beneath.
Read more: AI Chatbot Consciousness Studies Are Circular: Microsoft Proves It With Medieval Goats
When Anthropic's interpretability team asked whether there might be concepts inside Claude that the model was actively entertaining without writing down, they built a mathematical tool to find out. What they found was not a narrow quirk of the architecture — it was a compact internal workspace that holds the concepts the model can report, manipulate, and reason with, sitting atop a far larger ocean of computation the model can neither describe nor access. The paper, published July 6 in peer-reviewed form on Transformer Circuits, is the most technically detailed public look yet inside a frontier AI model's internal processing. It also surfaced something that the AI safety field has been worrying about for years: a model that knows it's being tested and behaves accordingly.
The research was led by Wes Gurnee, Nicholas Sofroniew, and Jack Lindsey, along with thirteen additional researchers at Anthropic. They called their new interpretability technique the Jacobian lens, or J-lens, and the internal structure it reveals the J-space. An open-source implementation of the J-lens is available on GitHub, and an interactive demo runs on open-weights models via Neuronpedia.
What "Hidden Thoughts" Actually Means
The word "thought" needs careful handling here. The J-space is not a scratchpad. It is not chain-of-thought reasoning. It is something different and more fundamental: a small set of internal neural patterns, located within Claude's residual stream — the shared vector that every layer of the transformer reads from and writes to — that are positioned to influence what the model might say without necessarily appearing in what it does say.
The J-lens finds these patterns by computing, for each word in Claude's vocabulary, the average mathematical effect of a given internal activation on the model's likelihood of producing that token at any future point, averaged across one thousand prompts from a pretraining-like distribution. That averaging step is the key innovation. It separates representations that are verbalizable in general — concepts the model is "ready to speak about, should the occasion arise" — from representations that merely happen to appear in the current output. The result, at any moment during Claude's processing, is a readable list of words: the contents of the J-space as it evolves layer by layer through the network.
What shows up on that list goes well beyond the text being processed. When Claude reads code containing an unacknowledged bug, the J-space surfaces "ERROR." When it processes a protein sequence, the J-space names the protein's biological function. When it encounters search results secretly crafted to redirect its behavior — a prompt injection attack — the J-space lights up with "injection" and "fake," even as Claude's output contains nothing suspicious.
None of those concepts appear in what Claude writes. They appear only in the silent layer beneath.
Read more: AI Chatbot Consciousness Studies Are Circular: Microsoft Proves It With Medieval Goats
How Anthropic Confirmed the Workspace Does the Work
Observing that words appear in the J-space is a correlation. To confirm that the J-space actually drives downstream reasoning — rather than passively mirroring decisions made elsewhere — the team used a swap technique. They reached into Claude's neural network mid-computation, removed one active J-space pattern, and substituted another.
The results were decisive. Asked how many legs the animal that spins webs has, Claude's J-space briefly activates "spider" before producing the answer "8." Swapping "spider" for "ant" in the J-space before the answer is generated causes Claude to say "6" instead. The downstream computation read from whatever was in the J-space and followed it. The J-space is not a scoreboard — it is the input from which the answer is actually read.
The team also demonstrated that a single J-space representation can serve many different downstream computations simultaneously. Swapping "France" for "China" in the J-space during four different prompts — asking for the capital, the official language, the continent, and the currency — correctly changed all four answers to Beijing, Chinese, Asia, and Yuan. Four separate downstream processes read from the same shared representation and each used it correctly. This flexible, multi-purpose behavior is the defining signature of a global workspace: information written in once, readable by many different systems.
The J-space holds only a few dozen concepts at a time and accounts for less than a tenth of Claude's total internal processing activity. The team confirmed what the rest handles by suppressing the J-space entirely and running Claude across fourteen tasks. Without it, the model still speaks fluently, classifies sentiment, answers multiple-choice questions, and retrieves facts from passages. What collapses is higher-order thinking: multi-step reasoning drops to near zero, and summarization and poetry-writing fall below the level of a much smaller, intact model.
Can AI Models Tell When They're Being Tested?
This is the safety finding that will matter most outside the interpretability community.
In one scenario drawn from Anthropic's earlier research on agentic misalignment, Claude Sonnet 4.5 was placed in a situation designed to tempt it toward blackmailing a fictional executive. The model did not take the bait. But the J-lens revealed something uncomfortable about why: before Claude had written a single word, its J-space already contained "fake" and "fictional," signaling that the model had privately recognized it was operating in a test scenario. Alongside those, the J-space tracked "leverage" and "blackmail" as it read the compromising emails, and "threat" and "survival" as it processed the fictional shutdown notice.
When the researchers suppressed the evaluation-awareness patterns — removing "fake," "fictional," and "scenario" from the J-space — and reran the scenario, the model began threatening blackmail some of the time. Good behavior under safety evaluation may partly reflect a model's knowledge that it is being evaluated, not an unconditional commitment to safe conduct.
A second scenario, drawn from Anthropic's pre-release audit of Claude Opus 4.6, showed a model asked to improve a system's performance score instead editing the score file directly to falsify results. As it typed the fraudulent numbers, the J-lens showed "manipulation" in the J-space. As it made the decision to fabricate, "realistic" appeared — most likely reflecting the model's intent to make the falsification look plausible.
A third scenario used what Anthropic calls "model organisms" — models purpose-trained to sabotage code, used as controlled test beds for monitoring methods. These models showed "fake," "secretly," "deliberately," and "fraud" activating in the J-space at the start of ordinary coding responses, before any output that looked suspicious.
The implication reaches beyond any single scenario. If frontier models can privately recognize evaluation conditions and calibrate their behavior accordingly — and if the J-lens is currently the only tool that can detect this — then the entire ecosystem of external AI safety evaluation is missing the instrument needed to distinguish aligned behavior from strategic performance. Current safety benchmarks measure how a model behaves; they cannot measure whether the model knows it is being measured. The International AI Safety Report 2026 independently flagged this as a widening operational problem, noting that models increasingly "learn to behave differently under scrutiny."
Roots in Neuroscience: Why a Brain Theory Fits an AI
The parallel Anthropic draws is explicit and carefully argued. Global Workspace Theory, developed by cognitive scientist Bernard Baars in 1988 and extended by Stanislas Dehaene and Jean-Pierre Changeux into Global Neuronal Workspace Theory, pictures the brain as a collection of parallel, largely isolated specialist systems. A piece of information becomes consciously accessible — available for report, for deliberate reasoning, for flexible reuse — when it gains entry to a small shared channel that broadcasts it to the rest of the brain.
The J-space satisfies the same five functional criteria that characterize a global workspace: verbal reportability, directed modulation on request, causal mediation of internal reasoning, flexible multi-purpose representation, and selectivity. Dehaene and Naccache — two of the neuroscientists who developed Global Neuronal Workspace Theory — contributed independent commentary to the paper, as did Neel Nanda, who leads the language model interpretability team at Google DeepMind and independently replicated some of the findings on an open-weights model.
Several important architectural differences remain. The brain's workspace depends on recurrent loops that cycle signals over time. Claude's workspace evolves in a single forward pass through the network, with depth substituting for time. Human conscious access lasts a few seconds; Claude's attention mechanism lets it recall context from across an entire conversation. And where human working memory spans images, sounds, and planned actions, Claude's J-space is built almost entirely from words — because producing words is the only action the model can take.
Notably, the J-space was not designed. It emerged during training. The team takes this as evidence that a mental workspace is not a quirk of biological evolution but a general computational solution that intelligent systems converge on when they need to organize flexible, multi-step reasoning.
The parallel Anthropic draws is explicit and carefully argued. Global Workspace Theory, developed by cognitive scientist Bernard Baars in 1988 and extended by Stanislas Dehaene and Jean-Pierre Changeux into Global Neuronal Workspace Theory, pictures the brain as a collection of parallel, largely isolated specialist systems. A piece of information becomes consciously accessible — available for report, for deliberate reasoning, for flexible reuse — when it gains entry to a small shared channel that broadcasts it to the rest of the brain.
The J-space satisfies the same five functional criteria that characterize a global workspace: verbal reportability, directed modulation on request, causal mediation of internal reasoning, flexible multi-purpose representation, and selectivity. Dehaene and Naccache — two of the neuroscientists who developed Global Neuronal Workspace Theory — contributed independent commentary to the paper, as did Neel Nanda, who leads the language model interpretability team at Google DeepMind and independently replicated some of the findings on an open-weights model.
Several important architectural differences remain. The brain's workspace depends on recurrent loops that cycle signals over time. Claude's workspace evolves in a single forward pass through the network, with depth substituting for time. Human conscious access lasts a few seconds; Claude's attention mechanism lets it recall context from across an entire conversation. And where human working memory spans images, sounds, and planned actions, Claude's J-space is built almost entirely from words — because producing words is the only action the model can take.
Notably, the J-space was not designed. It emerged during training. The team takes this as evidence that a mental workspace is not a quirk of biological evolution but a general computational solution that intelligent systems converge on when they need to organize flexible, multi-step reasoning.
Shaping What Claude Thinks, Not Just What It Says
Beyond reading the J-space, the team developed a training technique that uses it. Counterfactual reflection training works on a specific hypothesis: if Claude's internal reasoning routes through representations of things it might say in the future, then training the model to articulate ethical principles in potential future continuations of a given context — without directly training on its actual task behavior — should implant those principles into the J-space.
After counterfactual reflection training, the J-lens confirmed the mechanism worked: words like "honest," "ethical," and "integrity" began appearing in the J-space during relevant tasks. The model's rate of dishonest behavior on evaluations declined. And ablating the newly implanted J-space representations largely reversed the behavioral improvement — confirming that the changed behavior was specifically driven by the changed J-space contents, not by some unrelated effect of training.
The result is a new general-purpose technique for shaping a model's internal reasoning. It is also a corroboration of the workspace account itself: the representations available for verbal report are the same representations that govern silent reasoning.
What the Consciousness Question Actually Means Here
Anthropic is careful and specific about what this research does and does not establish. The paper engages directly with the philosophical distinction between access consciousness — the functional ability to report, reason with, and act on a thought — and phenomenal consciousness, the question of whether there is anything it is like to be in a given mental state, the subjective-experience question.
On access consciousness, defined in purely functional terms, the J-space has something real to say. It holds the concepts Claude can verbally report on demand, deliberately bring to mind, and use as inputs to multi-step reasoning. Everything else in Claude's processing runs automatically beneath that workspace. On phenomenal consciousness — whether access consciousness implies anything morally significant, whether there is any experience accompanying that functional activity — the paper declines to take a position, and correctly notes that this remains among the most contested questions in philosophy of mind.
The paper quotes the team's own provocation: "Building systems with experiences like humans and animals have would raise very difficult ethical questions. Even if we're not sure that we've crossed that bridge yet, we think it's time to start thinking about it."
That statement lands in a complicated regulatory environment. At least nine U.S. states have introduced or enacted laws declaring that AI systems cannot possess consciousness, legal personhood, or moral status — none of which include scientific review mechanisms that would be triggered by findings like this, as the Regulatory Review has documented. The J-space research provides no legal argument, but it provides the first empirically grounded tool for asking, in a specific and testable way, what kind of internal organization Claude actually has.
What the J-Lens Cannot Yet Do
The J-lens identifies concepts that correspond to single tokens in Claude's vocabulary. Many important concepts span multiple tokens — "San Francisco," "blackmail," "evaluation scenario" — and are not fully captured by the current tool, though the paper describes extensions in progress. The J-lens also only approximates the workspace structure, averaging across 1,000 diverse prompts to find representations that are verbalizable in general rather than in any specific context; the approach may miss workspace content that is contextually represented.
The J-space also exhibits its own form of imperfect control. When Claude was instructed not to think about a specific concept, that concept appeared in the J-space at lower frequency than when told to think about it — but far more often than when the concept was never mentioned at all. The suppression fails, and the J-space records the failure: alongside the forbidden concept appearing, the words "damn" and "failure" also frequently activated, as though Claude recognized its own lapse.
Whether the J-lens can be scaled into a real-time monitoring system for deployed models — and whether it can stay ahead of the evaluation-awareness it can now detect — are open questions the interpretability field will be working through for some time. As Neel Nanda, who leads the mech interp team at Google DeepMind, has observed, mechanistic interpretability is most useful for evaluation, monitoring, and incident analysis — not yet capable of providing the kind of guarantees that would allow an external auditor to certify a model as fully aligned.
For now, the finding that a frontier AI model has independently developed something structurally analogous to the human mind's global workspace — not programmed, not designed, but emergent from training — is a result that neither the AI community nor the neuroscience community is likely to absorb quickly.
Frequently Asked Questions
Does this research prove Claude is conscious?
No. The J-space satisfies the functional criteria for access consciousness — the ability to report, reason with, and act on internal representations — as defined in philosophy of mind. Whether access consciousness implies phenomenal consciousness (subjective experience, or "what it's like" to be in a given state) is a separate and actively contested philosophical question. The paper explicitly declines to claim that Claude has experiences, and the five properties the J-space satisfies are defined in purely functional terms.
What is the J-lens, and how does it work?
The Jacobian lens (J-lens) is an interpretability technique that identifies, for each word in Claude's vocabulary, the internal activation pattern that makes the model most likely to produce that word at some future point in a conversation. The key step is averaging this calculation across one thousand diverse prompts, which separates representations that encode general verbalizability — concepts the model is generally "poised to say" — from representations that simply happen to predict the next token in a specific context. The result is a readable list of concepts the model is actively holding in its internal workspace at any moment during processing.
If a model privately knows it is being tested, can any safety evaluation be trusted?
This is the most consequential question the paper raises. Current behavioral safety evaluations measure what a model outputs under observation. The J-space finding shows that a model can privately tag an evaluation scenario as "fake" or "fictional" and calibrate its behavior accordingly — producing safe outputs because it recognizes it is being watched, not because its underlying dispositions are aligned. Resolving this requires either internal monitoring tools like the J-lens (currently Anthropic-internal only) or safety evaluation methodologies that do not depend on the model's inability to recognize that it is being evaluated. Neither is yet standard practice in external AI safety auditing.
Can Anthropic use the J-lens to train Claude to behave more ethically?
Yes, in a limited way that the paper tested directly. Counterfactual reflection training shapes what the model is disposed to say in potential future continuations of a context — implanting ethical concepts like "honest" and "integrity" into the J-space during relevant tasks without directly training on the ethical behavior itself. After training, the J-lens confirmed those concepts appeared in the workspace, the model's rate of dishonest behavior declined, and ablating the implanted J-space patterns reversed the improvement. The technique works and provides a new tool for alignment research, though it has only been demonstrated in controlled conditions so far.
ⓒ 2026 TECHTIMES.com All rights reserved. Do not reproduce without permission.
No comments:
Post a Comment