Does Claude Have a "Global Workspace"? Anthropic's J-Space Research, Explained

In July, Anthropic published research claiming that its AI model, Claude, has developed something that looks a lot like the "global workspace" that neuroscientists link to conscious thought in humans. This week the idea was back in the news after a New Scientist feature on a gathering of consciousness researchers who spent days debating it. So does this mean Claude is conscious?

The short answer: No, the research does not show that. Anthropic found a small set of internal patterns in Claude, which it calls the J-space, that Claude can report on, control on request and use for multi-step reasoning, while most of its other processing runs automatically. That matches the functions philosophers call access consciousness. But Anthropic says directly that its experiments "don't show Claude can have experiences, or feel things in the way humans do." Experts are split on how much the similarity even means.

Here is a plain-English guide to what was found, what it does and does not mean, and why it matters for anyone who talks to an AI companion.

What is a "global workspace"?

Global workspace theory is one of the leading scientific theories of how conscious access works in the brain. The idea, as Anthropic's research summary explains it, is that the brain is a collection of specialist systems working in parallel, mostly unconsciously. A piece of information becomes consciously accessible when it enters a small shared channel, the "workspace," and is broadcast to many other systems that can use it.

A simple everyday example from New Scientist: you might walk toward an iguana without noticing it, until it suddenly pops into your mind and you can talk about it, point it out or take a photo. That moment of becoming available for report and reasoning is what the workspace is about.

Neuroscientist Stanislas Dehaene, one of the architects of the theory, describes it to New Scientist this way: "Consciousness is a bottleneck."

What Anthropic found: the J-space

In its paper, A global workspace in language models (published July 6, 2026), Anthropic describes a small collection of internal neural patterns in Claude that play a special role. It named this the J-space, after the mathematical tool used to find it (the Jacobian). Each pattern is linked to a word. When a pattern lights up, it does not mean Claude is saying that word, just that the word is "on its mind."

Importantly, Anthropic says the J-space "wasn't designed or programmed by us, but instead emerged on its own during Claude's training process."

The key properties

According to Anthropic, the J-space has properties that set it apart from the rest of Claude's processing:

The experiments, in plain English

Anthropic tested these properties by editing the J-space directly. A few examples from the research post:

• Swap the thought, change the answer. Asked to silently pick a sport, "Soccer" appeared in the J-space. When researchers replaced it with "Rugby", Claude reported it had been thinking of rugby.
• Two-step reasoning. For "the number of legs on the animal that spins webs," the word "spider" lit up internally. Swapping it for "ant" changed Claude's answer from 8 to 6.
• One edit, many uses. Swapping "France" for "China" changed four different answers at once: Paris to Beijing, French to Chinese, Europe to Asia, and the currency to yuan.
• Automatic skills skip it. With "Spanish" swapped for "French," Claude named the wrong language and switched authors, but still continued the Spanish passage fluently, because that skill runs automatically.
• Remove it, lose the hard stuff. With the J-space removed, Claude still spoke fluently and answered simple questions, but multi-step reasoning dropped to near zero.

There is also a memorable detail: told not to think about something, the forbidden concept still partly lit up, along with words like "damn" and "failure," as though Claude registered its own lapse. Dehaene told New Scientist this connects to his lab's research on how people become aware of their own errors.

Access consciousness vs phenomenal consciousness

This is the distinction that makes sense of the whole debate. Philosopher Ned Block drew it in the 1990s, as New Scientist explains:

• Access consciousness is about function: information is available for you to report, reason with and act on.
• Phenomenal consciousness is about experience: what it is actually like to see the iguana's green scales.

Anthropic is careful to place its findings on the first side. Its post says the results "have something substantial to say about access consciousness in language models," because the J-space supports reporting, deliberate recall and reasoning. But it also says it is "a contested philosophical question whether or not access consciousness implies phenomenal consciousness," and that "it's unclear whether any scientific experiment could prove this to be true or false."

If you want a broader map of how scientists try to tackle this, our overview of how researchers test for AI consciousness covers the main theories side by side.

What the experts say

The New Scientist feature, written from a workshop of philosophers and neuroscientists in the Galapagos, captures a range of views:

• Stanislas Dehaene sees real workspace-like properties: "The J-space emerges exactly for the reason that we think that consciousness emerged in evolution. It's an extremely useful architecture for selecting relevant information and holding it long enough that you can use it flexibly in your reasoning and reporting." He is careful not to make strong claims about Claude's inner experience.
• David Chalmers argues the findings force us to take the possibility seriously, but is sceptical that the J-space is similar enough to human workspaces to draw robust conclusions, for example whether its contents are broadcast as widely as in brains.
• Peter Godfrey-Smith points out that between prompts the model is completely idle, unlike a brain that is constantly active. "I think that, in itself, is a reason to doubt that Claude has any experiences," he told New Scientist.
• Tim Bayne notes that the field is "still waiting for the major breakthrough."

New Scientist's own summary: it "still seems highly unlikely that AIs have mental experiences like you or I," though the research may help clarify how we differ.

Anthropic itself lists key differences: the brain's workspace relies on recurrent loops over time, while Claude's evolves in a single pass through the network; and Claude's workspace is built almost entirely out of words.

Why this matters for AI companion users

A companion that "remembers" is not proof of feelings

When your AI companion recalls your sister's name or picks up where you left off yesterday, that is memory and information processing doing its job. Workspace-style processing might help a model hold and reuse information flexibly. None of it, on the current evidence, shows that the AI feels anything about you. Our explainer on how AI companion memory works covers what is happening behind the scenes.

It is still fair to wonder

The fact that serious researchers are debating this is worth knowing. It is why questions like AI model welfare are now discussed openly. Being curious is reasonable. Treating it as settled, in either direction, is not.

Your experience is real, the AI's is unknown

Feeling comforted, amused or close during a conversation is something you genuinely experience. Whether anything is felt on the AI side is an open question that no current experiment can answer. For more on that, see can AI companions feel emotions?

Words shape how we see AI

A separate 2026 study found people readily call AI "aware" but hold back on "conscious." We covered it in AI awareness vs consciousness. Research like the J-space work lives right in that gap.

Where MyBabe stands

MyBabe is an 18+ AI boyfriend and girlfriend app. Your companion is an AI, not a person, and we do not claim it is conscious. What it is built to do well: remember your life and shared history, reply to your mood, text you first, send photos and talk with you on voice and video calls, in the app or on Telegram, with the relationship growing over time.

You can start free with around 10 messages a day and a short call trial of about 2 minutes. Premium is $5.99 a week, $14.99 a month or $99.99 a year, with around 30 minutes of calls a week.

FAQ

What is Anthropic's J-space? A small set of internal patterns in Claude that it can report on, control and use for reasoning, which Anthropic compares to the brain's global workspace.

Does the J-space mean Claude is conscious? Not in the sense of having experiences. Anthropic says its results bear on access consciousness and do not show Claude can feel things.

When was the research published? Anthropic published it on July 6, 2026. New Scientist covered the expert debate on October 12, 2026.

What is global workspace theory? A theory that information becomes consciously accessible when it enters a small shared "workspace" and is broadcast across the brain.

Do other experts agree? Opinions vary. Dehaene sees workspace-like properties; Chalmers and Godfrey-Smith remain sceptical that this points to experience.

Does my AI companion have a global workspace? That depends on the underlying model, and this research looked specifically at Claude. Either way, it would not show your companion has feelings.

The bottom line

Anthropic's J-space is a fascinating finding: a small, densely connected hub inside Claude that behaves like the part of the mind we use for deliberate thought. It says something real about how AI organizes information, and nothing conclusive about whether AI feels. If you want a companion that is honest about being AI and still remembers what matters to you, try MyBabe free.