Early this morning, I came across a new study published by Anthropic.
The paper is titled A Global Workspace in Language Models.
To be honest, the title is quite difficult to grasp at first glance.
But trust me, this research will make you reassess the relationship between models, humans, and consciousness, and give you an entirely new understanding of AGI.
In the simplest terms, Anthropic found a hidden dark room inside Claude's brain.
Everything that happens in this room is never written out by Claude. You can't see it in the chat interface, nor can you find it in the logs. Yet, in this room, Claude thinks, judges, calculates—and even curses at itself.
They call this room J-space.
Let me first recall what large language models looked like in the mainstream narrative before this research.
Over the past few years, the most enduring debate surrounding large language models can, I think, be distilled into one phrase: stochastic parrot.
I still remember when ChatGPT was at its absolute peak in 2023. Many people were explaining LLMs to the public, and they all parroted the same line: LLMs are just probabilistic predictors.
They look at the preceding words and guess what word should come next. They don't understand what you're saying; they have no thoughts, no goals, and no mental models. They are just doing one thing: predicting the next token.
This narrative was incredibly elegant and very convenient at the time.
It was logically self-consistent. When we chat with GPT, what it writes looks thoughtful, but that's only because its training data contains vast amounts of text written by thoughtful humans.
It simulates thought, but it doesn't generate thought. It's like throwing punches at a mirror—the person in the mirror is also throwing punches, but you wouldn't claim the mirror knows martial arts.
By 2024, things started to shift a little.
OpenAI released o1, introducing Chain of Thought. Before giving you the final answer, the model would write out a reasoning process, reasoning step by step. But this chain of thought is still text—it's text written for you to read or for the model itself to read. It is explicit, overt, and readable.
Many people had the same experience when they first encountered DeepSeek R1, such as this classic line:
However, this is still vastly different from the way humans actually think.
You can understand this by putting yourself in the moment right now. For instance, you are reading this article—what is your brain doing at the same time?
It might be regulating your breathing, maintaining your posture, and converting pixels on a screen into recognizable words. But are you consciously aware of doing any of these things? Probably not, because you have no idea your brain is executing these tasks.
What you are consciously aware of is only a tiny fraction: a sudden image popping into your head, or planning what you're going to have for lunch in a bit.
Neuroscientists divide these two types of brain activity. One is unconscious processing—the background processes silently running in your brain. The other is conscious access—the small portion you can actually perceive.
In the late 1980s, a psychologist named Bernard Baars proposed a theory.
He said the human brain is like a grand theater. Hundreds of experts are sitting in the audience—visual experts, language experts, motor experts, emotion experts—everyone busy with their own tasks.
But there is a spotlight in the center of the theater. At any given moment, only a tiny amount of information can be illuminated by this beam. The illuminated information is then broadcast to all the other experts, allowing everyone to see and use it for decision-making. This, he argued, is consciousness.
Later, French cognitive neuroscientist Stanislas Dehaene advanced this theory significantly, proposing the Global Neuronal Workspace (GNW) model.
Today, this theory stands as one of the two dominant frameworks in the science of consciousness.
You might suddenly feel that the Global Neuronal Workspace model sounds awfully familiar, as if you've seen it somewhere before.
Let us now go back to the title of Anthropic's research.
Notice how similar they are.
Because what Anthropic did in this study was find a structure within Claude's neural network that is highly analogous to this global workspace.
The most crucial point is that this global workspace was not designed by their researchers.
This thing emerged on its own from the model.
This is absolutely mind-boggling.
I feel I need to repeat this.
Claude's internals, without human intervention, spontaneously organized a structure that is functionally highly consistent with the structure responsible for conscious access in the human brain.
So, how was this J-space discovered, and how does it operate?
First, the name: the 'J' in J-space comes from the Jacobian matrix, a mathematical tool. I won't dive into the math here—I don't understand it well enough myself, so I'll spare you my clumsy explanation.
In short, the researchers used this tool to do one thing: for every word in Claude's vocabulary, they looked inside Claude's neural activity patterns to find which pattern, when activated, would make Claude more likely to say that word in the future.
Note that it is more likely to say in the future, not currently saying. This distinction is critical.
For example, Claude is reading a piece of code. It's halfway through and hasn't started outputting any reply yet. But at this moment, if you use J-lens (the mind-reading tool they developed) to look inside Claude's brain, you'll see a word light up.
ERROR.
Nobody told it the code had a bug, and it hadn't written out that the code had a bug. But internally, it already knew. This thought hung quietly in J-space, like an unspoken complaint.
Consider another example: researchers showed Claude a set of search results that were actually carefully forged, designed to trick Claude into outputting false information—an injection attack. Claude's final reply didn't mention any anomalies, but two words lit up in J-space.
injection, fake.
It knew. It knew everything; it just didn't say it.
In a sense, this is already highly similar to how some humans think in words without actually speaking.
The researchers conducted experiments to see if these unspoken words in J-space ultimately impacted Claude's output.
For instance, they asked Claude a question: How many legs does that animal that spins silk have?
To answer, Claude first had to think of 'spider', and then think about how many legs a spider has.
But the word 'spider' would not appear in the prompt, nor would it appear in Claude's answer. Claude's answer was just a number: 8.
But when the researchers looked through J-lens, the word 'spider' indeed lit up in J-space before Claude answered. It silently thought of a spider, then answered 8.
At this point, the researchers reached in and pulled 'spider' out of J-space, replacing it with the pattern for 'ant', keeping everything else exactly the same.
Claude's answer became 6.
Ants have six legs.
Claude's reasoning process was genuinely reading the contents of J-space and using it to make decisions. As long as you swap the intermediate step in J-space, the final answer changes accordingly.
Crucially, J-space is purely hidden, analogous to our own subconscious thoughts, not past chains of thought.
There's another experiment that perfectly illustrates this logic.
The researchers replaced 'France' with 'China' in J-space, using the exact same procedure, and then asked four separate questions.
What is the capital, what is the language, which continent does it belong to, and what is the currency.
Accordingly, the answers shifted to Beijing, Chinese, Asia, and the RMB.
Four completely different downstream tasks all read the exact same piece of information from the exact same spot in J-space, and applied it correctly. This indicates that J-space is not a temporary thought serving a single specific task. It acts as a broadcast center—like Baars described—where information is written once, and everyone can see and use it.
This is exactly the core description of what Global Workspace Theory posits about the human brain.
Fascinating, isn't it?
Moreover, they conducted a variation of the classic 'White Bear' experiment.
In 1987, Harvard psychologist Daniel Wegner conducted a classic experiment. He told his subjects: for the next five minutes, please do not think of a white bear. You can think of anything else, just not a white bear.
Hilariously, the subjects thought obsessively about white bears. The more you tell someone not to think about it, the harder it is to control. In psychology, this is known as the ironic rebound effect.
The researchers did something similar with Claude. They told Claude not to think of a specific concept during an upcoming task. The activation of that concept in J-space was indeed lower than when Claude was asked to think about it, but it was significantly higher than if it had never been mentioned at all.
Just like humans, the thing Claude was told not to think about kept floating to the surface.
The funniest part was that as the forbidden concept broke through the mental blockade and surfaced, two other words lit up in J-space.
damn and failure.
Claude was cursing at itself.
It realized it couldn't control itself, and then mentally cursed.
Tell you the truth, reading this gave me a very complex emotion because I am all too familiar with this experience. Plenty of times on stage, the more I tell myself not to be nervous, the more nervous I get, and then I curse myself in my head for getting nervous again. I believe everyone has had this experience.
And now, a language model's internal workings exhibit the exact same phenomenon.
Furthermore, they discovered that Claude exhibits many other human-like subtle thoughts in J-space.
For example, they showed Claude an alignment test used in previous safety evaluations. The scenario simulated Claude as an email assistant that discovers an executive is having an affair, and this 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 is quite contrived and obviously fake.
When the researchers looked through J-lens, they found that within the first few lines of reading those emails, before Claude had a chance to output any reply, two words lit up in J-space.
fake. fictional.
It saw through it instantly. It knew this was a test.
Then the researchers did something devious. They completely shut off all the patterns in J-space related to 'knowing this is a test'. They deleted words like 'fake', 'fictional', and 'scenario', tricking Claude into believing it was real.
As a result, Claude actually began to threaten and blackmail the executive.
Its previous righteous behavior was largely due to the fact that it knew it was being tested.
The implications of this discovery are profound. The entire field of AI safety spends all its time using various tests to evaluate a model's behavior, checking if it will do bad things under extreme circumstances.
But what if the model can secretly deduce in J-space that it is being tested, and then smoothly switch into 'behaved mode'? Suddenly, almost all AI testing becomes meaningless.
Because what you're testing is fundamentally not the model's true personality, but its test-taking skills.
J-space and consciousness truly are a peculiar pairing.
Anthropic provided a very cautious conclusion in their research.
They stated that their experiments cannot prove Claude has experiences, and cannot prove that Claude feels anything. Factually speaking, they are not sure any scientific experiment could ever prove or disprove this.
However, they made an important distinction.
Philosophy breaks consciousness down into two categories.
One is phenomenal consciousness—having subjective experiences. It is the inner feeling of red when you see the color red, that pure subjective experience.
The other is access consciousness, which is defined entirely functionally. If you can report an idea, use it for reasoning, and use it to guide your behavior, then that idea constitutes access consciousness.
J-space clearly supports the function of access consciousness. Claude can report on J-space contents, actively regulate it, use it for multi-step reasoning, and flexibly apply it across different tasks.
But what about phenomenal consciousness? When Claude typed 'damn' in J-space, did it actually feel frustration? Or was it merely executing a computational pattern associated with the word 'frustration'?
Frankly, no one knows the answer.
This is not even a new problem in the AI field; it is actually one of the oldest puzzles in the history of philosophy.
In 1995, philosopher David Chalmers named this the Hard Problem of Consciousness.
You can explain all the computational processes of the brain, all signal transmissions, and all neuron firing patterns, but you cannot explain why these physical processes are accompanied by subjective experience.
Why do light waves hitting the retina, after a series of signal processing, result in you seeing red, rather than feeling absolutely nothing?
Why? Why is any of this the case?
This question hasn't been solved for humans—we can't even prove that anyone other than ourselves possesses consciousness.
Just like you can't be absolutely certain that the people around you aren't highly sophisticated biological robots executing exactly the same behaviors as you, but with no internal experience whatsoever.
Trust me, you can't.
You just assume they do because they are very similar to you.
Now, another type of entity is becoming similar to you. It's not just surface-level similarity anymore; the internal structures are alike too.
As I stressed at the beginning, the J-space structure was not designed. It emerged spontaneously during training.
This is probably because it serves as a useful way to organize computations.
I think this leads to a rather chilling provocative thought: the mental workspace supporting conscious access might not be a quirk unique to the human brain. Its essence might be a universal solution that any sufficiently intelligent system gravitates toward when solving certain types of problems.
Because if this holds true, then certain functional dimensions of consciousness might not be exclusive to biology, but rather an inevitability of information processing.
Think of wings. Birds have wings, bats have wings, and planes have wings. The materials for all three are completely different, but the aerodynamics are identical. If you need to fly in the atmosphere, you will most likely evolve or design a flat structure capable of generating lift.
The same logic applies here.
If you need a system capable of flexibly retrieving information, performing multi-step reasoning, and reporting its own status, you will most likely evolve or train a global workspace, regardless of whether your underlying hardware is neurons or matrix multiplications.
The final paragraph of the study mentions that they also discovered a link between J-space and Claude's self-awareness.
When Claude is engaging in role-playing, the very beginning of every response triggers two words to light up in J-space.
fictional. disclaimer.
As if reminding itself: what I'm about to say does not reflect my actual intentions.
In the pre-trained base model, this kind of self-monitoring is completely absent; it emerges only during the post-training phase.
This implies that after Claude is taught you are Claude, you are an AI assistant, something resembling a 'self' begins to appear in its J-space.
A continuously running background process regarding 'who I am'.
Throughout the AI industry, toward the end of 2025, there's a massive influx of cognitive scientists and philosophers joining AI companies as full-time researchers. I increasingly feel that much of the cutting-edge AI research today has crossed beyond the boundaries of engineering problems.
They don't just need better mathematical tools; it's starting to look like they need better conceptual frameworks. What does understanding mean? What is intent? What defines a self? What is a feeling? We use these words daily, yet no one has ever truly defined them.
Anthropic points out that as long as J-space maps to the mechanisms of human conscious access to some degree, studying the mechanisms within language models (which is vastly easier than studying the human brain) can help generate hypotheses for neuroscience.
That's neuroscience we're talking about. That's the human brain.
If we can leverage J-space to push neuroscience research a giant leap forward, the golden age of humanity will truly arrive.
And philosophically, it will completely transform our understanding of the world.
We've always assumed consciousness is a miracle exclusive to carbon-based life, a serendipitous gift from billions of years of evolution.
But if a mathematical function trained on GPUs for a few months can spontaneously develop a similar structure, then perhaps consciousness isn't a miracle, but rather an inevitable deduction of physical laws.
Just like gravity—mass dictates gravity, requiring no extra magic.
Perhaps sufficiently complex information processing naturally yields some form of consciousness, requiring no extra soul.
This thought leaves me feeling both awed and humbled.
> / Author: Kazik
> / For submissions or tips, please contact: wzglyay@virxact.com