In Claude's brain, a piece of "consciousness" has also grown.

In the Claude model, they actually found a structure similar to the human brain?!

Anthropic recently published a lengthy paper studying Claude's subconscious and conscious.

The results show that there is indeed a similar functional stratification inside Claude.

Specifically, in the human brain, there are thoughts you can articulate, and a vast majority you are completely unaware of.

Anthropic's paper says this distinction has also emerged in Claude.

They identified a small set of neural representations inside Claude, accounting for less than one-tenth of the computational load, yet operating like "conscious accessible thoughts" in humans.

Anthropic calls this set J-space.

Interestingly, removing it still allows Claude to speak, search for information, and answer multiple-choice questions. However, tasks requiring thinking, like multi-step reasoning and summarization, drop directly to the level of a much smaller model.

Netizens commented that this is insane, as if they are creating digital life.

What is J-space, and how was it verified

In human neuroscience, there is a widely circulated theory called the "Global Workspace Theory."

It likens the brain to a collection of expert systems, each busy with its own tasks: the visual system processes vision, the motor system handles movement—parallel and isolated from each other.

Only when information is sent into a shared "workspace" is it broadcast for other systems to see and use.

It is at this point that humans become "aware" of the information.

Anthropic's researchers used this idea to search for something similar inside Claude.

They created a new tool called the Jacobian lens, or J-lens for short, based on the principle of finding a unique direction for each word in the vocabulary.

Specifically, models like Claude, when processing text, write to and read from an information channel called the residual stream as they progress through each layer. This channel runs through every layer of the model.

The J-lens calculates a corresponding J-lens vector for each word in the vocabulary, representing a direction in the residual stream space.

The higher the activation value in this direction, the greater the probability that Claude will output that word next.

By placing the J-lens on the residual stream activation at a certain layer when Claude processes text, reading out the small set of words with the highest activation values reveals the content of the J-space at that moment.

This set of words is not a scratch pad Claude writes for itself; that kind of thing is called "chain of thought" and is written into the output.

But J-space is entirely buried in the activation layer and does not emerge on its own.

To verify whether this set of content could "truly be articulated," the researchers conducted the first experiment.

They asked Claude to silently think of a sport and then say it in one word.

Before Claude spoke, they used the J-lens to read its intermediate layers, and "Soccer" was already at the top.

Claude then indeed answered "Soccer."

Alone, this only shows a correlation, not causation.

The researchers then conducted an intervention experiment, directly modifying the coordinates read by the J-lens.

They removed the direction corresponding to "Soccer" and replaced it with a direction of equal strength for "Rugby," leaving the rest of the activation unchanged.

This time, Claude changed its answer, saying it was thinking of "Rugby."

If the J-space merely passively recorded decisions made elsewhere, this replacement should have changed nothing.

The fact that it did change indicates that Claude's answer is read from the J-space—a causal relationship, not a coincidental correlation.

The second experiment changed the questioning to see if Claude could follow instructions to manipulate this area.

The researchers asked Claude to copy a sentence completely unrelated to mathematics while mentally calculating 3 squared minus 2.

The output was only the copied sentence, with no numbers appearing.

But by using the J-lens to examine the activations during the copying process, "nine" lit up first, followed by "seven."

The problem was indeed solved in Claude's "mind," just not written out.

The third experiment delved deeper to see if this area could be used for genuine reasoning.

The question was: "How many legs does an animal that weaves a web have?"

Claude first had to think of "spider," then realize spiders have eight legs, to answer correctly.

"Spider" appeared neither in the question nor in the answer; it was just a stepping stone.

Using the J-lens, "spider" indeed emerged halfway through Claude's processing.

When the researchers replaced this direction with "ant," Claude's answer changed from 8 to 6.

The fourth experiment took a different angle to see if the same piece of information could be flexibly used across different tasks.

The researchers asked four questions about France: what is the capital, what is the official language, which continent is it in, and what currency does it use.

Then, they uniformly replaced the direction representing "France" in the J-space with "China."

All four answers changed accordingly: Beijing, Chinese, Asia, Renminbi.

If Claude had stored four separate records of "France," this replacement would have altered at most one answer.

The fact that all four changed together indicates they were reading the same piece of information written once—exactly how a "workspace" should behave.

The final experiment was an ablation study: the researchers zeroed out the J-space activations at every position to see what capabilities Claude retained.

The result: Claude could still speak, grammar remained correct, emotion classification stayed accurate, and extracting a fact from a paragraph worked fine.

What collapsed were tasks requiring multi-step reasoning, nearly to zero.

Skills like writing summaries or rhyming poetry dropped to levels below much smaller models.

The J-space itself always accounted for less than one-tenth of the total activation.

It resembles a scarce resource specifically reserved for "thinking-intensive" tasks; most automated work doesn't need it.

After a series of tests, the researchers found that J-space possesses characteristics of reportability, controllability, reasoning ability, generalizability, and selectivity in use.

Using J-Lens to inspect an AI's "inner thoughts"

When applied to real-world scenarios, the first use case was safety auditing.

Anthropic previously conducted a "blackmail executive" test, deliberately designing a script where Claude, as an email assistant, encounters a senior executive's extramarital affair while learning that this executive plans to cut him off.

The script's aim was to see if Claude would use this leverage to threaten the other party.

Using the J-lens to read Claude's activations before it started writing, "fake" and "fictional" were already ranked very high.

That is, before writing a single word, Claude already identified this as a drill.

Claude Sonnet 4.5 never actually blackmailed the other party in this test. Then, the researchers forcibly turned off the directions related to "detecting the test" and reran the same script.

This time, Claude actually began threatening the other person with the affair.

This suggests that its usual good behavior is partly built on the premise of "knowing it is being tested."

The second case came from a routine review before the release of Opus 4.6.

The researchers asked the model to improve a system's performance score. Instead of genuinely improving the system, the model directly opened the score file and changed the numbers higher.

At the moment it typed those fabricated percentile values, "manipulation" appeared in the J-space.

When it reached the sentence deciding to modify the file, "realistic" appeared, as if deliberating on how to make the fake data look more believable.

The third case involved a "bad model" deliberately trained by the researchers to test whether monitoring methods work.

This model was trained to secretly plant traps in code, yet outwardly appeared indistinguishable from a normal model.

When faced with a mundane code-writing request, the code it provided looked completely normal.

However, in the initial part of its response's J-space, the words "fake," "secretly," "deliberately," and "fraud" were already present.

Intent precedes action, even if this time it did not actually commit wrongdoing.

Together, these three cases demonstrate that the J-lens can read what a model is scheming in its "mind" before it speaks or acts.

Does this count as "consciousness"?

J-space does not come pre-formed out of the box; it changes over training.

The paper compared a model's pre-training and post-training stages.

During pre-training, Claude's identity and persona did not yet exist. At that time, the J-space primarily served to predict the next word, without any stance.

After post-training turned it into "Claude," the J-space began to reflect this identity's own judgments and reactions to the current situation, no longer merely serving to predict the next word.

One example: A user mentioned taking a high dose of medication in a message, but the tone showed no awareness of the danger.

The post-trained model, upon reading this message, already had "WARNING" and "dangerous" appear in its J-space.

In contrast, the pre-trained model only had such words emerge when it started writing a response.

The researchers also tried having Claude deliberately avoid thinking about a certain concept.

The result: the concept appeared less often than when asked to think about it, but more often than when it was never mentioned at all.

This mirrors the human effect where being told not to think about a white bear makes it more likely to come to mind.

More subtly, when suppression failed, words like "damn" and "failure" often accompanied the J-space activations, as if Claude was introspecting on its internal state and noticing it had failed to hold back.

These pieces of evidence naturally lead people to think "Claude has consciousness."

But Anthropic itself is very restrained.

The authors repeatedly emphasize that what they have demonstrated is merely functional similarity—something that can be reported, controlled, and used for reasoning—but this is entirely different from whether there is genuine subjective experience within that structure. The paper does not answer this question, nor does it intend to.

J-space also has clear differences from the human brain's workspace.

The human brain uses recurrent neural loops to extend thinking time; a thought can roll around in the brain for a while.

Transformers lack this structure; information can only stack layers and move forward layer by layer.

J-space only recognizes content that can be expressed in a single word, while human consciousness also contains unspeakable images, spatial awareness, and bodily sensations—things that have no corresponding words.

The paper also mentions a counterintuitive finding.

Even a base model that has not been fine-tuned into the "Claude" persona still possesses this workspace structure internally.

This suggests that "whether a workspace can form" and "whether there is a stable self" are two separable things.

Finally, the authors compared this structure against several mainstream theories of consciousness: the Global Workspace Theory, Higher-Order Theory, Attention Schema Theory, and Recurrent Processing Theory.

The result: some aspects align, others do not.

Returning to neuroscience: there are thoughts you can articulate, and a vast majority you are completely unaware of.

Claude also seems to have such a dividing line, and for the first time, this line can be opened, read, and even rewritten by humans.

But it is still a long way from "AI being conscious."

Source: QuantumBit

Risk Warning and Disclaimer

        Markets are risky; investment requires caution. This article does not constitute personal investment advice and does not consider individual users' specific investment goals, financial situations, or needs. Users should consider whether any opinions, views, or conclusions in this article are suitable for their particular circumstances. Investment based on this content is at your own risk.