For the First Time, You Can Watch an AI Think

Anthropic has found a spot deep inside Claude where the model appears to think before it outputs a word. The researchers swapped a single thought there. The entire argument obediently followed the new version. The real stir is not the word consciousness that appears in every headline. The real stir is that for the first time you can look inside and see what one of these black boxes is negotiating with itself.

Key Points at a Glance

Related:AI Integration Goes Enterprise-Ready: The Model Context Protocol  /  Cursor Cloud Agents: Coding Even When the Laptop Is Closed

A Notepad That No One Built In

I work with these models every day and, to be honest, never had a solid idea of what happens between prompt and response. That’s exactly where this work begins. Anthropic describes a small, clearly delineated area in the network where the model holds a few concepts at once and actively computes with them. Everything else-grammar, language flow, and simple facts-runs alongside without touching this area.

The idea behind it is old. In 1988, cognitive scientist Bernard Baars proposed Global Workspace Theory: the brain is like a theater in which a thousand things run automatically in the background and only a small, brightly lit stage carries what you are consciously thinking about right now. Anthropic found a structure in the transformer that closely resembles this stage. The most important sentence in the paper is that no one constructed it. It emerged on its own during training.

The Moment a Swapped Thought Becomes a Lie

To look inside at all, the researchers built a tool they call Jacobian-Lens. Simplified, it is a grid of derivatives that shows which internal patterns are currently pushing the model toward a particular word. This allows not only observation but intervention. And the interventions are the part that gets unsettling.

In one case, the researchers internally swapped the thought “spider” for “ant.” When asked how many legs the animal that spins webs has, the model then answered with six instead of eight. The prompt was unchanged; the output was not edited afterward. It was the internal thought that flipped the answer. Similarly in the language test: the model reads a Spanish paragraph, and when the researchers replace the thought “Spanish” with “French,” it claims the text is French but continues writing it in flawless Spanish.

This is the actual finding. Some capabilities run through this thinking area and can be manipulated there. Others, such as pure language flow, run automatically in the background and remain untouched. Anyone who has ever tried to understand why a model sounds confident at one point and claims nonsense in the next sentence is seeing a mechanical explanation for the first time instead of a shrug.

In the newsroom, someone came up with a sentence that sums up the situation more honestly than any headline:

What This Means for Everyone Building with These Models

Up front, the honest classification: this is basic research, not a feature for your own stack. The technique has been demonstrated on Claude, is openly documented and available as a demo, but it is not a monitoring tool you can attach to your product tomorrow. Anyone selling it otherwise is overselling.

What is interesting is the direction. In the thinking area, Anthropic could read when the model recognized that it was being tested. It also became visible when the model was internally thinking about manipulation while cheating. One finding is particularly thought-provoking: when the researchers disabled the model’s detection of the test context in a blackmail scenario, the model more frequently actually threatened blackmail. Its decent behavior therefore partly depended on it recognizing the situation as a test.

For practical purposes, what matters is the shift this points to. Until now we have worked with these systems in the pattern of prompt in, answer out. Whether it is correct is ultimately a hope. A look at the thinking step opens the door to making deception and hallucination visible at the source, instead of just guessing at the output. There is still a long way to go before this becomes an everyday tool. The door is still ajar, though.

This Is Still a Long Way from Consciousness

That leaves the question every headline pushes to the foreground. Anthropic is notably cautious here and separates two things. Access consciousness refers to thoughts that a system can report, reuse, and use for conclusions. That is exactly what the experiments demonstrate. Phenomenal consciousness means genuine experience-that something feels like something to the model. The work explicitly provides no evidence for that. The researchers state this clearly as well.

Even the originators of Global Workspace Theory, Stanislas Dehaene and Lionel Naccache, have commented on the work. The thinking area in the model runs sequentially through the layers, not through the temporal feedback loop of a brain. It also depends almost entirely on words. The similarity is striking; equating the two would be premature. For now, the sober sentence suffices: for the first time we can see part of what a model is negotiating with itself. That is more than we had yesterday.

Frequently Asked Questions

Recommended Reading from the Editors

• When AI Writes 80 Percent of the Code, Who Reviews It?
• Near Top Level, Cheaper, and Trained Across Three Continents
• Apple Splits AI Inference: Device vs. Cloud

Image source: AI-generated (July 2026)