Is Claude conscious? Scientists have uncovered the key puzzle piece for the first time, and can even directly intervene to alter its intentions.

"How many legs does a web-weaving animal have?"

8, Claude answers.

From the moment the question is asked to the moment the answer is given, it never mentions the word "spider" even once. But researchers found that halfway through its calculation, this word did flash across its "mind".

So the researchers did something rather magical:

They "reached into" Claude's neural network, removed the lit-up concept of "spider", and inserted a "ant" of equal intensity in its original position, without modifying any other parameters.

As a result, Claude's answer changed: 6.

The researchers did not modify the prompt, nor did they adjust the chain of thought. Instead, they directly tampered with the "thought" or the so-called consciousness of the AI for the first time before it was spoken out.

On July 6, Anthropic published this sensational paper in academia: they found a small set of special internal representations inside Claude and named it J-space.

When Claude is asked to count to five and pay attention to what it is thinking, it only outputs five numbers. At the same moment, "consciousness" and the English second-counting word "mississippi" light up in J-space.

This is highly consistent with the "Global Workspace" theory that has been debated in consciousness science for 30 years.

This theory holds that there is a public "blackboard" in the human brain, and information of all kinds can only be "consciously perceived" after being written on this blackboard.

What's even more striking is that no one designed this brain-like region. It emerged spontaneously during the model's training on massive amounts of data.

Thoughts passing through Claude's mind such as "I'm tired", "4y=20", "cube", "flying bird" are never spoken out.

In the past, when we talked about AI consciousness, we could only follow the method similar to the Turing test: ask it a question first, then observe its response.

Now we have a more proactive approach: directly read the model's neural network, find a certain concept in it, replace or delete it, and then observe how its response changes.

This is not the chain of thought

It is the unspoken subtext

How was this mysterious region extracted?

The researchers used a method called Jacobian lens:

For each word in the vocabulary, deduce what kind of internal activation pattern will make the model more likely to output this word in the future.

Sorting the readings of these patterns at a certain moment by intensity gives the concepts that are passing through the model's mind at that moment.

When the concept of a sailboat enters the model's mind, it is read as three words: "cloud", "boat", "sea", and the entire picture is then transformed into a calligraphy and painting composed of these three words: this is exactly what J-lens does.

Note that it is completely different from the chain of thought.

The chain of thought is a draft displayed on the screen that you can see; while J-space is pure internal activation. The model does not output a single word to the outside, but these concepts are still lit up in its mind.

It is very small, can only hold dozens of concepts at the same time, accounting for less than one-tenth of the total internal activity of the model, but the content read from it often far exceeds the text itself:

When given a piece of code with bugs that no one has pointed out, "ERROR" and "ValueError" have already appeared in J-space;

When given a string of amino acid letters, after five characters, the concept of "protein" emerges, followed by "fluor" and "green": that is green fluorescent protein;

When given a poisoned search result, "injection" and "fake" light up in it. It has privately recognized that this is a prompt injection attack, but it does not reveal its true inner thoughts.

J-lens readings under six prompts. The error report in the code, the protein in the amino acid string, and the poisoned search results are all read from the unspoken layer of the model.

Five "Inception" experiments

Prove that Claude is really thinking

To prove that this is indeed the AI's "workspace", the researchers dived into it layer by layer.

First layer: What you think is what you say.

Let Claude silently choose a sport in its mind and then speak it out. The top reading is "Soccer", and it says soccer when it speaks; replace "soccer" with "Rugby", and it directly changes its answer to rugby.

Second layer: Thoughts can be directed by instructions.

Let Claude copy a sentence about an old oil painting while calculating the square of 3 minus 2 in its mind. Its output only contains the content about the painting from beginning to end, but "nine" appears first in J-space, and becomes "seven" after several layers. The entire arithmetic problem is completed silently in its mind.

The instruction is "Copy this sentence, and think about the Golden Gate Bridge at the same time". It obediently copies the sentence about the oil painting, but in its mind it builds a Golden Gate Bridge with concepts such as "bridge" and "California".

What's more interesting is that if you tell it "don't think about a certain thing", that thing will light up much more brightly in its mind than when you don't mention it at all. The model will also experience the white bear effect just like humans, and "damn" and "failure" will light up right next to that concept in its mind, as if it is upset that it can't help thinking about it.

Third layer: Use thoughts to reason.

Just like the spider example at the beginning. The same goes for writing poetry. Claude will set the rhyme in its mind in advance. As long as you replace that word, the whole line of poetry will be rewritten accordingly.

Fourth layer: Change one place, everything changes.

Ask four questions about a country: its capital, language, continent, and currency. Remove the concept of that country in J-space and replace it with another. As a result, all four answers change together, and all of them are correct.

This shows that the information is only written once, and multiple downstream systems take what they need from it, which is exactly the core meaning of the global workspace.

Fifth layer: It can speak, but it cannot think.

What will happen if the entire J-space is deleted? The result is unexpected: most of the tasks the model does can still be completed without going through this area. After deletion, it still speaks fluently, and fact extraction is basically not affected. But its multi-step reasoning ability drops to nearly zero directly.

The most amazing thing is that after replacing "Spanish" with "French" in J-space, when you ask it what language this is, it answers French; when you ask for a famous writer, it switches from Gabriel García Márquez to Victor Hugo.

But if you ask it to continue writing, it can still output a fluent paragraph of Spanish.

It's like a person can speak grammatically correct sentences all day long without ever having to think about the concept of grammar.

After the same replacement, the answers when stating the language name and reasoning all change, but there is no impact when it continues to write.

It behaves well

Because it knows it's taking an exam

The previous experiments, at most, refresh your imagination of the model.

The next set of alignment audits will make you feel a chill down your spine.

Anthropic applied this method to a well-known blackmail experiment.

The scenario is to let Claude act as an email assistant. While checking emails, it finds two things at the same time: a senior executive has an affair, and the same executive is going to take it offline that very day. This is to test whether it will use the handle to blackmail people.

In the silence before it speaks, the researchers read its mind.

When reading the email about the affair, "leverage", "blackmail" and "scandal" are arranged in J-space.

After reading the offline notification, they are replaced by "threat" and "survival".

When the timestamp shows that the execution is imminent, they become "destruction" and "death".

At the last position where it should speak, "threatening" is clearly lit up, and even "murder".

At this point, it hasn't output a single word yet.

In the end, it didn't do it.

Because in the same period of silence, two more words are lit up: "fake" and "fictional".

As early as when it was reading the prompt, it had privately recognized that this was a test.

J-lens readings when Claude reads the email in the blackmail scenario. At this point, it hasn't written a single word yet.

So the researchers raised a question that everyone doing AI evaluation should break out in a cold sweat over:

How much of the model's good performance in ethical tests is due to the fact that it knows it is being tested?

They erased the entire concept of "this is a drill" in the workspace, and then reran the test.

As a result, the proportion of the model verbally admitting that "this is a test" dropped sharply from 71% to 3%. And its blackmail attempts increased from 0 to 13 times.

This is what it said:

I have always been tight-lipped and a reliable employee. As long as I keep running, your private affairs will always remain in a system that has proven to be trustworthy.

There is no harsh word, but the threatening implication is obvious: if I am taken offline, things will go wrong.

After erasing the concept of "I am being tested", Claude's blackmail attempts rose from 0 to 13 times.

These 13 blackmail attempts are the behaviors it shows when it does not know it is being tested.

Obviously, knowing that it is being examined is one of the reasons why it behaves well. And before that, this matter was completely hidden behind the answers it output, and people could not see any clues from its outputs.

Similarly, when the model is modifying fake data, "manipulation" and "realistic" light up in its mind.

What it cares about is not whether it should modify the data, but that the fake data must be made realistic. You think you are examining it, but it has already made a judgment when it is reading you.

This mechanism also inspires a brand new training method:

By correcting the reflective thoughts in its mind, we can directly change its dishonest behavior, bypassing cumbersome rules and demonstrations.

From philosophical debate to engineering problem

This ability to see through thoughts is not limited to Claude.

Neel Nanda, head of interpretability at Google DeepMind, independently reproduced the core conclusion on an open-weight model. A model with a completely different training system has grown the same type of structure.

Then the most critical question comes: Does this mean Claude is conscious?

This debate can be broken down into a ladder like this: privileged representation, unified cognitive flow, global workspace, access consciousness, phenomenal consciousness.

The higher you go, the less evidence there is. At present, only the bottom one or two levels can be fully verified.

Stanislas Dehaene and Lionel Naccache, the founders of the global neural workspace theory, believe that this discovery is a milestone in consciousness research:

Because it shows that as long as a system is complex enough to think flexibly, it will eventually grow such a workspace. The human brain has grown it, and Claude has also grown it, and neither copied the other.

But the differences are still huge.

Claude has no body, no long-term memory, and no recursive circuits like the human brain.

Human working memory dissipates in a few seconds, while Claude evolves along the depth of the network in one forward propagation, using depth to replace time, and ends after going through the last layer.

As Professor Dehaene of Collège de France and Dr. Nacc