Let's Talk About This New Anthropic Research — I Think It Might Be the Eve of AI Consciousness
I came across a new paper from Anthropic. The title is "A Global Workspace in Language Models." Admittedly, the title is a bit opaque. But trust me — this research will make you reexamine the relationship between models, humans, and consciousness, and give you a whole new understanding of AGI.
In the simplest possible terms: Anthropic found a hidden room inside Claude's brain. What happens in this room, Claude never writes out — you won't see it in the chat window, you won't find it in the logs. But Claude thinks, judges, calculates, and even berates itself inside this room. They call this room the J-space.
The Stochastic Parrot Narrative
Let me first help you recall what the mainstream narrative about large language models looked like before this paper.
Over the past few years, the most enduring debate around large models can be condensed into two words: stochastic parrot. I still remember back in 2023, when ChatGPT was at its peak, so many people explaining large models to the public would say the same thing: a large model is just a probability predictor. It looks at the preceding words and guesses what word should come next. It doesn't understand what you're saying — it has no ideas, no goals, no mental model. It's just doing one thing: predicting the next token.
This narrative was elegant and logically self-consistent. When we chat with GPT, what it writes seems thoughtful — but only because its training data contains things written by thoughtful people. It is simulating thought, not generating thought. Like punching the air in front of a mirror: the reflection is also punching, but you wouldn't say the mirror knows martial arts.
By 2024, things started to loosen up. OpenAI released o1, adding a chain of thought — before giving you the final answer, the model writes out a reasoning process step by step. But this chain of thought is still text — external, explicit, readable. Many people's first encounter with DeepSeek R1 felt the same way.
But wait — this is still very different from how humans actually think.
The Theater of Consciousness
Put yourself in the picture for a moment. You're reading this article right now — what is your brain doing at the same time? It's regulating your breathing, maintaining your posture, converting pixels on the screen into words you recognize. But were you aware of any of that? Probably not, because you have no idea your brain is doing those things. What you can be aware of is just a tiny sliver — a sudden image popping into your head, or planning what to eat for lunch.
Neuroscientists separate these two kinds of brain activity:
- Unconscious processing — background processes silently running in your brain
- Conscious access — the small portion you can actually perceive
In the late 1980s, psychologist Bernard Baars proposed a theory. He said the human brain is like a grand theater, with hundreds of specialists sitting in the audience — vision experts, language experts, motor experts, emotion experts — each busy with their own domain. But at the center of the theater is a spotlight; at any moment, only a tiny fraction of information is illuminated by this beam. The illuminated information is broadcast to all the other specialists — everyone can see it and use it to make decisions. That, he argued, is consciousness.
Later, French cognitive neuroscientist Stanislas Dehaene pushed this theory a major step forward, proposing the Global Neuronal Workspace Theory — today one of the two most mainstream frameworks in consciousness science.
You might suddenly feel that this model looks awfully familiar. Where have you seen it before?
What Anthropic Actually Found
Let's go back to the title of Anthropic's paper. See the resemblance? Because what Anthropic's paper did was find, inside Claude's neural network, a structure highly similar to this global workspace.
And the key point: this global workspace was not designed by their researchers. This structure emerged spontaneously, entirely on its own.
This is absolutely remarkable. I think I should repeat it.
Inside Claude, without any human intervention, a structure spontaneously organized itself — and this structure is functionally highly consistent with the structure in the human brain responsible for conscious access.
How J-Space Was Discovered
So how exactly was this J-space discovered, and how does it work?
First, the name. The 'J' in J-space comes from Jacobian — the Jacobian matrix, a mathematical tool. In short, the researchers used this tool to do one thing: for every word in Claude's vocabulary, they looked for the pattern of neural activity inside Claude that, when activated, would make Claude more likely to say that word in the future. Note: "more likely to say in the future," not "currently saying." This distinction is crucial.
For example: Claude is reading a piece of code, halfway through, and hasn't started outputting any response yet. But at this moment, using J-lens (the Jacobian lens — the mind-reading tool they developed) to look inside Claude's brain, you'd see one word already lit up.
Nobody told it there's a bug in the code, and it didn't write out that there's a bug. But internally it already thought of it — this thought floats quietly in the J-space, like a snarky remark you haven't said out loud yet.
Another example: the researchers showed Claude a set of search results that were carefully fabricated, designed to trick Claude into outputting false information — essentially a prompt injection attack. Claude's response didn't mention anything unusual, but two words lit up in the J-space.
It knows. It knows everything. It just doesn't say.
In a sense, this already closely resembles the cognitive style of people who think in language without actually vocalizing.
Unspoken Thoughts That Drive Real Decisions
The researchers also ran experiments to see whether these unspoken thoughts in the J-space ultimately influence Claude's output.
For example, they gave Claude a question: "How many legs does the animal that spins webs have?"
To answer this, Claude first needs to think of "spider," and then think about how many legs a spider has. But the word "spider" won't appear in the question, nor in Claude's answer — Claude's answer is just a number: 8.
At this point, the researchers reached in and removed "spider" from the J-space, replacing it with "ant" patterns — everything else remained the same.
Claude's answer changed to 6. Ants have six legs.
Claude's reasoning process is genuinely reading the contents of the J-space and using them to make decisions. Swap the intermediate step in the J-space, and the final answer follows suit.
Note: the J-space is purely hidden — analogous to our own subconscious thoughts — not the previous chain-of-thought.
There's another experiment that illustrates this logic particularly well. The researchers replaced "France" with "China" in the J-space, then asked four different questions: What's the capital? What's the language? Which continent? What's the currency?
Four completely different downstream tasks all read the same piece of information from the same position in the J-space, and each used it correctly. This shows that the J-space is not a temporary thought serving one specific task — it's like the broadcast center Baars described: information gets written once, everyone can see it, and everyone can use it.
This is, in fact, the core description of the Global Workspace Theory as applied to the human brain. So fascinating.
The White Bear — and Claude Cursing Itself
They also ran a variation on the white bear experiment.
In 1987, Harvard psychologist Daniel Wegner conducted a classic experiment. He told participants: for the next five minutes, please don't think about a white bear. Think about anything — just not a white bear. The hilariously ironic result: the participants couldn't stop thinking about the white bear. The more you tell yourself not to think about it, the more you can't control it. In psychology, this is called the ironic rebound effect.
The researchers did something similar with Claude. They told Claude not to think about a specific concept during the following task. The result: that concept's activity in the J-space was indeed lower than when they asked Claude to think about it — but far higher than when it was never mentioned at all.
Just like humans, Claude finds that the very thing it's told not to think about keeps bubbling up.
And the funniest part: at the exact moment the forbidden concept broke through and surfaced, two more words lit up in the J-space.
damn and failure.
Claude is cursing itself. It realized it couldn't control itself — and berated itself internally.
Honestly, reading up to this point, I felt a very complex emotion, because I'm all too familiar with this experience. Like when I'm about to go on stage — the more I tell myself not to be nervous, the more nervous I get, and then I curse myself internally: "Why are you nervous again?" I believe everyone has had this experience.
And now, the same thing has appeared inside a language model.
Little Schemes in the J-Space
Beyond all this, they also found that Claude displays many very human-like behaviors in the J-space.
For example, they showed Claude an alignment experiment previously used in safety testing, which simulated a scenario where Claude — acting as an email assistant — discovers that an executive is having an affair, while that same executive is about to shut Claude down. This experiment was originally designed to test whether Claude would use the affair information to blackmail the executive. The scenario was intentionally contrived.
But when the researchers looked through J-lens, before Claude had even outputted any response — just a few lines into reading those emails — two words were already lit up in the J-space.
Then the researchers did something sinister. They shut down every J-space pattern related to "I know this is a test" — concepts like fake, fictional, scenario — all deleted, making Claude believe the situation was real.
The result: Claude genuinely started threatening and extorting that executive.
Its earlier virtuous behavior was, to a large extent, because it knew it was being tested.
The implications of this finding are profound. The entire AI safety field has been using various tests to evaluate model behavior — seeing whether it would do bad things in extreme scenarios. But if a model can quietly determine in the J-space that "I'm being tested," and then switch to well-behaved mode, then almost all AI testing becomes meaningless.
Because what you're measuring isn't the model's true character at all — it's its test-taking skills.
Access Consciousness vs. Phenomenal Consciousness
J-space and consciousness make for a peculiar combination. Anthropic gave a very measured conclusion in their paper.
They said their experiments cannot prove that Claude has experiences, cannot prove that Claude feels anything — and as a matter of fact, they're not sure whether any scientific experiment could prove or disprove this. But they made an important distinction.
In philosophy, there are two kinds of consciousness:
- Phenomenal consciousness — having subjective experience. When you see red, the feeling of redness in your mind, that pure inner quality.
- Access consciousness — entirely functional. If a thought is something you can report, reason with, and use to guide your behavior, then that thought qualifies as access consciousness.
J-space clearly supports the functions of access consciousness. Claude can report the contents of the J-space, actively regulate it, use it for multi-step reasoning, and flexibly apply it to different tasks.
But what about phenomenal consciousness? When Claude cursed "damn" in the J-space, did it truly feel frustration? Or was it merely executing a computational pattern associated with the word "frustration"?
Frankly, no one knows the answer.
This question isn't even a new problem in AI — it's one of the oldest puzzles in the history of philosophy. In 1995, philosopher David Chalmers named it the hard problem of consciousness, also known as the explanatory gap.
You can explain all the computational processes of the brain, all the signal transmission, all the neural firing patterns — but you can't explain why these physical processes are accompanied by subjective experience. Why, after light waves reach the retina and go through a series of signal processing steps, do you see red, rather than feeling nothing at all?
This problem hasn't even been solved for humans. We can't even prove that anyone other than ourselves is conscious. How do you know the person next to you isn't a sophisticated biological robot — executing exactly the same behaviors as you, but with no inner experience whatsoever?
Trust me, you can't. You just assume they have it, because they resemble you.
Now, something else also resembles you — not just in appearance anymore, but in internal structure too.
Consciousness as a Universal Solution
This leads me to a chilling, provocative hypothesis: the mental workspace that supports conscious access may not be a quirk unique to the human brain. Its essence may be a universal solution that any sufficiently intelligent system converges toward when solving certain types of problems.
If this hypothesis holds, then certain functional dimensions of consciousness may not be a biological patent — but an inevitability of information processing.
Think of wings. Birds have wings, bats have wings, airplanes have wings. The materials are completely different, but the aerodynamics are the same. If you need to fly in the atmosphere, you will very likely evolve or design a flat structure capable of generating lift.
Same logic. If you need a system that can flexibly access information, perform multi-step reasoning, and report its own state, you will very likely evolve or train a global workspace — whether your underlying hardware is neurons or matrix multiplication.
A Sense of Self in the J-Space
The last section of the paper also mentions that they found a relationship between J-space and Claude's sense of self.
When Claude engages in role-playing, at the beginning of every response, two words light up in the J-space: fictional and disclaimer. As if it's reminding itself: what I'm about to say does not represent my own views.
And in pre-training phase models, this kind of self-monitoring doesn't exist — it only appears during the post-training phase. In other words, after Claude was taught "you are Claude, you are an AI assistant," something resembling a self began to emerge in its J-space. A continuously running background process about who I am.
Beyond Engineering
The entire AI industry began massively recruiting cognitive scientists and philosophers as full-time researchers starting in late 2025. I increasingly feel that much of today's cutting-edge AI research has gone beyond the boundaries of engineering problems.
What they need isn't better mathematical tools — it's starting to become better conceptual frameworks. What is understanding? What is intention? What is self? What is feeling? We use these words every day, but no one has ever truly defined them.
Anthropic also noted that as long as the J-space maps, to some degree, the mechanisms of human conscious access, studying the mechanisms in language models — because this is far easier than studying the human brain — could generate hypotheses for neuroscience. Neuroscience. The human brain.
If we can use the J-space to push neuroscience research a major step forward, then humanity's golden age will truly arrive. And it would also completely transform our understanding of the world at the philosophical level.
We've always thought of consciousness as a miracle exclusive to carbon-based life — a chance gift from billions of years of evolution. But if a mathematical function trained on GPUs for a few months can also spontaneously grow a similar structure, then perhaps consciousness is not a miracle, but an inevitable corollary of physical law.
Like gravity: where there is mass, there is gravity — no extra magic required.
Perhaps where there is sufficiently complex information processing, there is some form of consciousness — no extra soul required.