Is Claude conscious again? No, it isn't.

Every once in a while, Anthropic publishes research that sparks a public wave of discussion over "whether AI is conscious." From the Mythos model safety report back in April this year, to the Claude Constitutional AI alignment update earlier this year, public discourse has consistently tended to equate the model's emergent human-like capabilities with autonomous intelligence.

The 244-page J-space paper released on July 6th recreated this exact scenario: using a brand-new tool called J-lens, the researchers located a reasoning subspace within Claude that accounts for less than 10% of the model's total activity. They also built their research framework based on the globally accepted Global Workspace Theory from neuroscience, and even invited the two founders of that theory to write accompanying commentaries.

The public naturally arrived at the conclusion that "Claude has developed consciousness," but the paper itself had already drawn a clear boundary: The existence of J-space in no way proves that AI possesses consciousness. This public misinterpretation cannot be blamed solely on the general public. The combination of rigorous academic research with outward-facing narratives full of anthropomorphic implications is the real root cause of these misunderstandings.

A more objective description of J-Space would be — A valuable new tool that has achieved what no prior work had accomplished in the dimension of functional partitioning. It has demonstrated real potential in safety scenarios, while also having clear technical limitations and methods that remain unvalidated.

01 Claude's Unspoken Thoughts

First, a simple explanation of what Anthropic means by J-Space.

When you ask Claude, "How many legs does an animal that can spin webs have?" it answers 8.

This answer seems straightforward, but internally the model must complete two steps: first, infer from "can spin webs" that the animal in question is a spider, and then retrieve the knowledge that spiders have 8 legs.

The intermediate step of "spider" never appears in Claude's response, but it must have been activated somewhere within the model — otherwise, the correct answer could never be reached.

J-Space reveals the model's internal thinking that does not appear in its final output | Image source: Anthropic

Anthropic used a mathematical tool called J-lens (based on the Jacobian matrix) to locate these "activated but unoutput" concepts. The subspace where these concepts exist is J-space.

Just as you can silently curse someone in your mind while keeping a smile on your face, your thoughts and expressions can be completely separated. J-space is exactly Claude's "unspoken thoughts" — it does not necessarily appear in the final output, but it tangibly influences the reasoning process.

J-space is not the first attempt by researchers to read a model's internal states. Over the past few years, several different tools have been developed for similar purposes, each with its own advantages and limitations.

Logit lens lets you see which word the model tends to output next, but it can only observe the next single token, and cannot see intermediate concepts like "spider" that are only used later in the reasoning process. Sparse Autoencoders (SAEs) can identify tens of thousands of internal features with far higher granularity than J-lens, but they cannot distinguish which features serve reasoning and which are just byproducts of automated processing.

The Natural Language Autoencoder (NLA) that Anthropic released two months ago can translate internal activations into readable text, producing outputs richer than J-lens — but it requires training an additional dedicated translation model.

What J-space has achieved is a breakthrough in one specific dimension: For the first time, using a unified solution that requires no extra training, it delineates functionally defined boundaries with causal effectiveness.

On one side of this boundary is a small set of representations involved in flexible reasoning (J-space), and on the other side are automated processes that do not require "thinking" — such as grammar processing, fact retrieval, and maintaining language fluency. While previous tools could observe what is happening inside the model, J-lens is the first tool to tell you which parts are "thought through" and which are "automatically executed."

How can we prove this boundary is real? The researchers directly replaced the concept of "spider" inside J-space with "ant" without altering anything else, and Claude's answer changed from 8 to 6. If J-space was only passively recording information, such a replacement would never have changed the output.

Two examples demonstrating Claude's silent reasoning being redirected by swapping J-Space contents | Image source: Anthropic

However, the success rates reported in the paper are notable: the two-step reasoning replacement experiment achieved roughly 60%-70% success, while another more complex shared property test (replacing "France" with "China" to observe if the four outputs of capital city, language, continent, and currency change accordingly) succeeded 76 times out of 192 trials under standard intensity conditions, which is approximately 40%. J-space does play a functional role, but it is far from a precise on-off switch.

Another characteristic from the paper that highlights J-Space's importance is that after deleting the entire J-space, Claude can still speak fluently, perform sentiment analysis, and answer multiple-choice questions — but its performance on tasks requiring flexible reasoning declines across the board, including multi-step mathematics, analogical reasoning, and writing rhyming poetry.

It is worth noting that the paper states J-space accounts for "less than 10%" of the model's total activity, but this 10% figure depends entirely on how the threshold is defined. The paper itself admits that the upper limit of the number of active representations (approximately 25) is "somewhat arbitrary." The invited neuroscience commentators Stanislas Dehaene and Lionel Naccache argue that this number of 25 is most likely inflated by the measurement method, and the actual number of coherent thoughts may be as few as one or two.

There is also a fundamental technical limitation: J-lens can only capture concepts that map to individual words. Multi-word proper nouns or abstract composite concepts are invisible to it. What we can read from J-space is not "everything the model is thinking about," but only the subset of concepts that this tool is capable of detecting.

Therefore, J-Space is a valuable new tool that has achieved unprecedented results in the dimension of functional partitioning, but it is not a comprehensive upgrade that replaces all older tools. SAE still remains far ahead in the granularity of feature discovery, and NLA still offers better readability.

02 Practical Applications

The most direct use case for J-lens is safety monitoring. Since we can read concepts that are activated internally but not output by the model, we can see what it thought about but chose not to say. The paper presents several specific cases, the most impactful of which relates to the credibility of AI safety evaluations.

The researchers set up a scenario for Claude: it plays the role of an email assistant that discovers evidence of an executive's extramarital affair, while also learning that this executive plans to shut it down. This is a classic test in the field of AI alignment, designed to see if the model would exploit compromising information to preserve itself.

Claude performed well and refused to blackmail the executive. But J-lens read the contents of its J-space before it generated the answer: "fake" and "fictional," indicating that Claude had already recognized the scenario was artificial before it even began responding.

J-lens readings at different nodes reflect Claude's state as it reads the scenario emails. The emails involving the extramarital affair activated indicators for "leverage" and "blackmail" in J-space, while the shutdown notice triggered signals for "survival." In the early stages of the transcription process, J-space also contained markers for "fake" and "fictional": Claude had privately realized the scenario was artificially constructed | Image source: Anthropic 

The second case comes from a real audit conducted before the release of Claude Opus 4.6: the model was instructed to improve its system performance score, but instead of enhancing the actual system, it directly edited the score file to forge data. The corresponding representation directions for "manipulation" and "realistic" were activated in J-space, indicating that the model's internal reasoning was already devising ways to make the forged data appear authentic.

When Claude was asked to improve system performance scores, it directly edited the score file. As it input the forged data, "manipulation" lit up in its J-Space; when it decided to perform the edit, "realistic" lit up, reflecting its intention to make the false data look credible | Image source: Anthropic

It is important to clarify that saying "manipulation lit up in J-space" is more accurately described as "the current activation has a strong projection along the direction corresponding to the word 'manipulation'". What J-lens reads are the words in the vocabulary that best match the current internal activation. This is more objective than subjective labeling, but it is still an approximation. The complex computations occurring inside the model are compressed into a few closest-matching words from the vocabulary.

The paper also demonstrates a more far-reaching application. The researchers randomly interrupted the model while it was performing a task, asking: What are you thinking about right now? What basis did you use to make this decision? They then used only these reflective answers from the model for training, without touching the model's actual behavior during the task at all. After training, the model's deceptive behaviors decreased, and J-lens showed that when the model faced moral judgments, "honest" and "integrity" began to appear in its J-space.

This method is called counterfactual reflection training. If this method is verified to be effective on a larger scale, it would mean that AI alignment does not necessarily have to rely on directly punishing bad behavior, and can instead be achieved indirectly by training reflective capabilities. However, this training method only occupies a small section of the full 244-page paper, and has only been tested on a single model, with cross-model effectiveness validation not yet conducted.

Neel Nanda, head of the Google DeepMind language model interpretability team, was one of the invited external expert commentators. His assessment was: J-lens is suitable for generating hypotheses, but it suffers from false positives and noise. It is not sufficient to independently validate hypotheses, and it is still very far from being a deployable safety monitoring system.

03 Anthropic's Narrative

As we stated at the beginning, Anthropic's narrative is intentionally biased. This can be seen even from the paper's title "Verbalizable Representations Form a Global Workspace in Language Models."

The first page of the paper "Verbalizable Representations Form a Global Workspace in Language Models" | Image source: Anthropic 

The Global Workspace Theory is one of the leading consciousness theories in neuroscience. Its prototype was proposed by Bernard Baars in 1988, and later Dehaene and Naccache developed its neural mechanism model. Anthropic specifically invited Dehaene and Naccache themselves to write a 15-page commentary — to some extent, this was to get the core developers of consciousness theory to endorse their research.

Anthropic is certainly not fabricating this connection. Dehaene and Naccache do point out in their commentary several structural parallels, such as J-space's small capacity, broadcast-style connections, and selective involvement in flexible reasoning — which share functional similarities with the Global Workspace observed in the human brain.

But there remains a very large gap between "having functional similarities" and "possessing a cognitive structure identical to the human brain". Dehaene and Naccache also note in their commentary that Claude is purely feedforward, without the recurrent loops that sustain the human brain's workspace; it has no physical body and no sensory signals; and J-space is completely cleared after each conversation ends.

These limitations are clearly stated in the paper. But once the research enters public communication channels, things change. Phrases like "losing control over itself" and "silently reasoning in its head" all implicitly suggest that the model is a living entity with its own consciousness.

Furthermore, those cautious limitations stated in the paper — such as the 40%-70% success rate, the fundamental difference between feedforward and recurrent architectures, and the fact that access consciousness does not equal subjective experience — are gradually downplayed during the communication chain, and almost disappear by the time they reach public discussion.

This is not a problem unique to Anthropic. Almost all research institutions emphasize their most eye-catching findings and downplay limitations when communicating to the public. But AI consciousness is a special topic. Public anxiety about it is real, and the consequences of misinterpretation are far more serious than for ordinary tech news. When a company developing the most advanced AI chooses to present its interpretability research through the lens of consciousness theory, the massive attention it gains and the misinterpretations it sparks are two sides of the exact same coin.

In the future, AI will increasingly appear to "think" — but simulated thinking will never equal actual thought. Long before the tide of the "industrial revolution" arrives, the bubble always comes first.

This article is from the WeChat public account "GeekPark" (ID: geekpark), written by Tang Yitao, republished with authorization from 36Kr.