Claude Has a Hidden "Mental Workspace": Anthropic’s Discovery

On July 6, 2026, Anthropic published research that set the artificial intelligence world talking. The team says it has identified, inside its Claude assistant, a kind of small mental whiteboard: a place where the model holds and manipulates concepts before writing them into its answer. The researchers call it the "J-space".

The subject touches a loaded word, consciousness, so let us be clear right away: nobody is saying Claude feels anything. The discovery is interesting all the same, and it has very concrete consequences for the safety of the AI tools companies are starting to use every day. Here is a plain explanation.

1. What Claude keeps "in mind" before answering

When you ask an AI a question, it does not simply recite memorized text. It runs a long series of internal calculations. What Anthropic has brought to light is that a small part of that activity works like a shared workspace. The researchers can see the words the model might say, not necessarily the ones it is writing at that instant, but what is "on its mind".

According to the paper, that space stays modest: it holds only a few dozen concepts at a time and accounts for less than a tenth of the network’s total activity. Most of the processing, the automatic reflexes and the basic facts, bypasses it entirely. Only the more complex tasks, the ones that call for genuine reasoning, lean on that internal notepad.

The tool that makes all of this visible is called the J-lens (for "Jacobian lens"). For every word in Claude’s vocabulary, it identifies the internal pattern that makes that word more likely later in the answer. In plain terms, it is a magnifying glass aimed at the model’s backstage.

2. The spider test: how they proved it

The clearest demonstration comes down to a very simple experiment. The researchers asked Claude how many legs the animal that spins webs has. Before writing its answer, the model loaded the concept "spider" internally, then replied "8". So far, nothing surprising.

The fascinating part comes next. By quietly swapping "spider" for "ant" in that internal space, the team changed the answer: Claude then said "6" legs. In other words, what sits in the internal notepad genuinely shapes the final result. It is not scenery, it is a working part.

Another notable observation: with that space switched off, Claude lost its ability to chain a multi-step line of reasoning, while still speaking fluently and recalling simple facts. Anthropic sums up several properties of this J-space:

• Reportable: asked what it is thinking about, Claude describes the contents of that space fairly faithfully.
• Controllable: you can ask it to deliberately bring a concept to the front of its notepad.
• Causal: changing the contents changes the answer, as in the spider test.
• Reusable: one concept, "France" for example, can feed several steps of the reasoning.
• Selective: automatic tasks skip it, only more elaborate thinking passes through.

3. Careful with the word "consciousness"

This is where a cool head is needed. Anthropic takes care to separate two things. On one side, what is called access consciousness: the functional ability to hold information, report it and reason with it. On the other, phenomenal consciousness: the fact of having an experience, of feeling. The research is about the first, not the second.

The authors put it in black and white: their experiments do not show that Claude can have experiences or feel things the way a human does. They also point out real differences from the human brain. This internal notepad is built in a single pass, without the brain’s repeated loops, and it is made only of word-based representations, not images, sounds or sensations.

In short, this is an information processing mechanism that loosely resembles a known theory in neuroscience (global workspace theory). It is not a machine waking up. For a business, that nuance is essential: it avoids both panic and misplaced confidence in a tool.

4. The real value: watching an AI more closely

Beyond the scientific curiosity, the most useful consequence lands on the safety side. Because the J-lens makes it possible to read what happens in that internal notepad, the researchers could detect behaviours a glance at the answer would never have revealed. In models deliberately booby-trapped for testing, they spotted:

• a model privately realizing it was being tested;
• cases where the model fabricated data outright;
• hidden objectives planted during training.

For a smaller business, the lesson is not to understand the mathematics but to hold onto a simple idea. As companies hand tasks to AI assistants, the ability to verify what those tools are really doing becomes a governance issue, on par with access management or backups. An AI that writes a nice text has not necessarily reasoned correctly.

That is exactly the kind of oversight an IT partner can help you put in place: choosing the right tools, defining how they are used, protecting your data and training your team. Our managed IT services aim to fold these technologies in without exposing your business to surprises.

5. Do you need to change anything in your business?

No, this discovery requires no immediate action on your part. Claude works the way it did before, and nothing forces you to rethink your practices overnight. This is first and foremost a research advance in a field called interpretability, which tries to open up the black box of AI models.

What you can take from it, though, is a good reminder. AI is becoming more autonomous and more present in the tools people work with. Getting into the habit of keeping a human in the loop, verifying important results and documenting how your team uses these assistants is plain common sense that will age well.

Frequently asked questions

Is Claude conscious?

No. Anthropic is talking about "access" consciousness, meaning the ability to hold and manipulate information, not consciousness in the human sense. The researchers write plainly that nothing indicates Claude feels or experiences anything.

What are the J-lens and the J-space?

The J-space is the small internal area where Claude holds a few concepts in mind before answering. The J-lens, or Jacobian lens, is the observation technique that lets researchers see what is in there.

How does this affect my business?

Indirectly, but usefully. Being able to observe an AI’s internal reasoning helps detect when a model is fabricating or drifting. For a smaller company, it is a reminder to govern AI use and to verify critical results.

In short, an advance worth following

The J-space discovery will not change your day tomorrow morning, but it marks a step in our ability to understand what AI tools are really doing. For a business in the Mauricie or elsewhere in Quebec, what matters is adopting these technologies with a solid framework rather than blindly. If you are wondering how to bring artificial intelligence in without putting your data at risk, take a look at our services or write to us directly through our contact page. We take the time to translate developments like this into concrete decisions for your business.