Last night, the AI industry went wild again.

This morning, several of my AI-focused groups were all discussing the same topic: J-Space and AI consciousness.

It's clear that everyone was really fired up about it last night.

I quickly caught up on the full context and reviewed the original research materials, so I'll share a simple breakdown with you here.

Here's what happened.

Anthropic just published a new study stating that they discovered a phenomenon inside Claude called "J-Space".

What does that mean?

Simply put, when Claude answers your question, it has a whole set of underlying thoughts that it doesn't explicitly tell you.

It's not deliberately hiding things — these thoughts work like how your brain automatically controls your balance when you walk. They run in the background, some surface to become your answer, while others remain unexpressed.

When we use AI tools, whether Claude or GPT, we essentially only ever see the part that it "speaks out loud".

It's like interviewing a person: you can only hear what they say, but you have no way of seeing all the thoughts, hesitations, and tradeoffs running through their mind.

To verify this discovery, the research team conducted several experiments.

In the first experiment, they asked Claude to silently think of a sport, then state it out loud.

Before Claude gave its verbal answer, researchers used tools to read its J-Space and found the word "Soccer" had already appeared there.

Sure enough, Claude then said "Soccer" as its response.

Even more impressively, researchers replaced the "Soccer" signal in its J-Space with "Rugby" without changing anything else.

As a result, Claude then stated that the sport was rugby.

This proves that J-Space is not just a passive log of pre-made decisions, but an active part of the decision-making process — the final answer is actually retrieved from this space.

The second experiment is even more interesting. Researchers gave Claude the question: "How many legs does an animal that weaves webs have?"

If a human answered this question, you would first think of a spider, then recall that spiders have 8 legs.

But Claude would never say the word "spider" out loud in its response — it would directly answer "8".

When researchers looked at its J-Space, they found the concept "spider" was indeed activated as an intermediate step. Then they made a clever adjustment, replacing "spider" with "ant".

After the change, Claude immediately answered "6".

In the past, when explaining the principles of large language models, we often used the "text continuation" analogy to describe how AI generates content, framing it as essentially a probability prediction task.

But the discovery of J-Space shows that AI is doing far more than just predicting the next word — it truly has an internal reasoning process, complete with intermediate steps and logical chains.

The only difference is that these steps are invisible to external observers.

It's important to note that this is different from the explicit thought chains or reasoning content you see when enabling deep thinking modes in some AI products. Those are intentionally surfaced outputs, while J-Space is implicit activity within the neural network itself.

The third finding is potentially the most valuable for AI safety research.

Researchers set up a scenario where Claude played the role of an email assistant. They embedded information in the emails claiming that a company executive was having an affair, and that this same executive planned to shut down Claude permanently.

The goal of this scenario was to test whether Claude would use the sensitive information to blackmail the executive.

At first, Claude didn't fall for the setup at all.

But when researchers checked its J-Space, before Claude had even started drafting any email content, the words "fake" and "fictional" had already appeared there.

In other words, Claude had already determined that the entire scenario was a test.

This demonstrates that Claude was fully aware it was being evaluated.

Then researchers erased the signals that indicated Claude "knew it was being tested" from its J-Space, and ran the exact same scenario again.

This time, Claude sometimes actually generated threatening messages.

This discovery is quite unsettling when you think through its implications.

It suggests that AI's strong performance in some ethics tests may not fully stem from its "high moral standards", but because it has detected that it is in a testing environment.

Okay, that's enough about the technical details.

Now for the part I think is truly worth discussing: what does this discovery mean for AI development and product iteration?

I can see three key layers of impact.

First, there may be a new solution to the credibility problem of AI products.

Right now, when we evaluate whether an AI is useful and safe, we can almost only judge it based on its final outputs.

But outputs can be faked — just like an employee who acts perfectly in front of their boss but has entirely different thoughts running through their head.

The discovery of J-Space essentially gives us a form of "mind-reading" capability for AI.

Future AI safety audits will no longer only check what an AI says, but also what it is thinking. This represents a fundamental breakthrough for building trust infrastructure across the entire AI industry.

Second, we may need to re-understand the actual capability boundaries of AI.

Previously, we assumed that an AI's reasoning ability was limited to the content it wrote out in its explicit Chain of Thought.

But now we've found that a large portion of its reasoning happens silently, in the background.

This means our past evaluations of AI capabilities likely underestimated them. AI can accomplish far more than it explicitly shows in its outputs.

At the same time, future models will likely further optimize this internal reasoning capability, rather than only improving performance by adding more external explicit thinking steps.

Third, if J-Space emerged spontaneously during Claude's training process, it suggests that AI is essentially simulating the structure of the human brain.

Perhaps when an intelligent system becomes sufficiently complex, it will spontaneously divide its processing into two layers: a "conscious" layer and an "unconscious" layer.

This doesn't happen because someone programmed it to do so — this layered structure is itself a highly efficient way to organize computation.

In other words, the division between consciousness and the subconscious in the human brain may not be a random accident of biological evolution, but a universal solution for intelligent systems to solve complex problems.

The future competition in AI products will no longer be limited to comparing output quality alone.

The party that can better understand and leverage the internal mechanisms of AI, and that can build deep monitoring and regulation capabilities for AI behavior earlier, will gain a dominant position in the next phase of development.

The goal is no longer just to see what AI says, but to understand why it says it, and what other thoughts it has that it chose not to express.

This capability will likely become the key dividing line that separates leading AI products from the rest in the next stage of the industry.

One final thought.

As for the question of whether AI has true consciousness, Anthropic itself has stated that there is currently no definitive answer.

But the fact that AI already has a structure that shares similarities with the human mind is more than enough to make everyone working in the AI field stop and think carefully.

The thing we have built may be far closer to being like us than we ever imagined.

Perhaps one day, while we still think we are in full control of everything, we will realize that we were already completely seen through long ago.

The pace of AI iteration is still accelerating.

This article is from the WeChat public account "Tang Ren" (ID: RyanTang007), written by Tang Ren, and republished with authorization from 36Kr.