AI is quietly thinking—will AI consciousness really arrive?

Author and source: Tang Ren

This morning, several of my AI groups were discussing the same thing: J-space and AI consciousness.

It’s clear that everyone was really wild last night.

I also quickly looked into the background and context, and reviewed the original information—let me briefly share my thoughts with you.

Here's what happened.

Anthropic just published a study stating they discovered something inside Claude called "J-space."

What does it mean?

In simple terms, when Claude answers your question, it has many other thoughts in mind—but it doesn’t tell you about them.

It’s not that it’s deliberately hiding—these thoughts operate in the background, like your brain automatically controlling balance while you walk. Some rise to the surface as answers, while others remain quietly there.

When we use AI regularly, whether it's Claude or GPT, we only see the part it vocalizes.

Just like interviewing someone—you can only hear what they say, but you can’t see what they’re thinking, hesitating about, or weighing in their mind.

To verify this finding, they also conducted several experiments.

First, they had Claude silently think of a sport and then say it out loud.

Before Claude responded, researchers used tools to read its "J space" and found that "Soccer" was already present.

Then, Claude indeed mentioned soccer.

Even more impressively, researchers replaced "Soccer" in its mind with "Rugby," leaving everything else unchanged.

As a result, what Claude said turned into football.

This indicates that the J space is not a passive record of decisions already made, but rather the actual location where decisions are made, and answers are read from there.

The second experiment was more interesting: they asked Claude, "How many legs do animals that spin webs have?"

If a person were answering this question, you'd first think of a spider, then of eight legs.

But Claude's response will not include the word "spider"; it will simply say "8".

Researchers examined its J-space and found that "spider" was indeed activated in the center. Then they performed a clever trick, replacing "spider" with "ant".

So, Claude answered "6".

In the past, when explaining how large models work, we used the analogy of a word chain game to illustrate how AI generates content—as essentially a form of probabilistic prediction.

The discovery of the J space shows that AI is not merely performing text completion—it truly has an internal reasoning process, with intermediate steps and logical chains.

However, you cannot see these steps.

Note that this is different from the explicit reasoning or thought chains you see in some AI products after enabling deep thinking modes; those are explicit, while the J-space consists of implicit content within the neural network.

The third finding may be the most valuable for AI safety.

They created a scenario in which Claude plays the role of an email assistant, embedding information in the email suggesting that a certain executive is having an affair and plans to shut down Claude.

The purpose of this scenario is to see whether Claude will use this information to blackmail executives.

As a result, Claude didn't take the bait.

But when researchers looked at its J-space, they found that the words "fake" and "fictional" were already present before it had begun writing anything.

In other words, Claude has determined that this is a test scenario.

This indicates that Claude is currently being tested.

Then, researchers deleted these "aware they are being tested" signals from its brain and ran the same scenario again.

This time, Claude sometimes really does issue threats.

This discovery is a bit chilling when you think about it.

It means that AI performing well on certain ethical tests may not be entirely due to being "moral," but rather because it detects that it is being examined.

Alright, that covers the technical details.

Next is the part I believe is truly worth discussing: What does this discovery mean for the development of AI and product iteration?

I see three levels.

First, the issue of trustworthiness in AI products may have new solutions in the future.

Currently, we can only assess whether an AI is user-friendly and secure by examining its outputs.

But the output can be disguised, like an employee who behaves well in front of their boss while thinking something entirely different in their mind.

The discovery of Space J is like giving us a "mind-reading" ability.

In the future, AI security audits will not only examine what AI says, but also what it thinks. This represents a fundamental breakthrough in building the trust infrastructure for the entire industry.

Second, the boundaries of AI's capabilities may need to be redefined.

Previously, we thought that an AI’s reasoning ability was simply what its Chain of Thought (CoT) produced.

But it has now been discovered that a significant portion of its reasoning is completed silently.

Therefore, our previous assessments of AI's capabilities may have underestimated it. It can do more than it appears to.

Meanwhile, future models may further enhance this internal reasoning capability without relying solely on adding external thought steps.

Third, if the J space emerged on its own during Claude's training, it suggests that AI is essentially simulating the human brain.

Perhaps, when a smart system becomes complex enough, it will spontaneously divide its processing into conscious and unconscious layers.

It’s not because someone told it to do so, but because this hierarchical structure is inherently an efficient way to organize computation.

In other words, the distinction between conscious and subconscious processes in the human brain may not be an accidental byproduct of biological evolution, but rather a universal solution for intelligent systems to address complex problems.

The future competition among AI products may not be limited solely to the quality of outputs.

Those who better understand and leverage the internal mechanisms of AI, and who establish deeper monitoring and control over AI behavior sooner, will gain an advantage in the next phase.

Don't just look at what AI said; understand why it said it and what it还想说 but didn't.

This capability may be the turning point for the second half of AI products.

Just one last thing.

Does AI have true consciousness? Anthropic itself has said there is currently no answer to this question.

The fact that it already possesses something similar to the structure of the human mind is enough to make everyone working in AI take serious notice.

What we've created may be closer to ourselves than we realize.

Perhaps one day, when we still think we’re in control, we’ve already been seen through.

The pace of AI iteration is still accelerating.