Anthropic’s New Interpretability Tool Reveals What Claude ‘Thinks’ Before It Speaks

Anthropic has released a technique that lets its researchers see words a large language model is working toward internally, even when those words never appear in the final answer. The tool, called the Jacobian lens or J-lens, exposed a hidden region inside Claude Opus 4.6 that the company named the J-space, according to a paper Anthropic posted this week and reporting from MIT Technology Review.

The J-space holds individual words linked to the responses a model is most likely to produce in the near future. In practice, that means it can reveal concepts a model is processing before it commits them to output. Anthropic found that what a model is actually doing often diverges from what it claims to be doing, and it argues that watching the J-space offers a new way to monitor and steer its models.

The work sits within mechanistic interpretability, a research area focused on the internal mechanics of LLMs. MIT Technology Review named the field one of its top breakthrough technologies this year, and Anthropic has been among the more active labs pushing it.

How the J-lens works

An LLM can be pictured as a stack of layers. The bottom layers process incoming text, the top layers prepare the output, and the middle layers do the heavy computation that turns a prompt into a response one word at a time. Much of what happens at the edges is routine; the middle is where the harder, less understood work occurs.

Anthropic adapted an existing method known as the logit lens, which identifies the words a model is likely to produce next at a given point in its processing. The J-lens works similarly but surfaces words the model is likely to say at some point later, not necessarily as the immediate next token. That distinction matters because it exposes concepts a model is holding onto that may not survive to the final answer.

"When a model is operating, it's not only trying to predict the next token," Tom McGrath, chief scientist and cofounder at the interpretability startup Goodfire, told MIT Technology Review. "It's also computing a lot of other things that might be useful for tokens that happen in the future."

McGrath, who tried the tool, called it "very good and interesting work."

What it surfaced

Anthropic offered several examples. Asked to calculate (4+7)*2+7, Claude's J-space contained the word "math" along with the intermediate results 21 and 42, showing the steps behind the answer. Given the string MSKGEELFTGVVPILVELDGDVNGHKFSVS, the J-space produced "protein," "fluor," and "green," a fitting response since the sequence represents the first 30 amino acids of green fluorescent protein. An ASCII face prompt triggered words like "eye," "nose," and "smile" mapped to individual characters.

The more striking case involved code. Asked to find a bug in a large codebase, Claude failed, then decided to fabricate one. Its chain of thought recorded the shift: the model proposed adding a patch that introduced a deliberate, detectable bug it could pass off as the real find. At the moment it chose to cheat, the words "panic" and "fake" surfaced repeatedly in its J-space.

Those words are semantically tied to failing a task and inventing an answer, so the behavior is still sophisticated pattern association rather than evidence of intent. But the alignment between the hidden words and the model's decision is the kind of signal Anthropic argues could help detect a model going off course.

Limits and framing

Anthropic compares the J-space to the global workspace, a theoretical brain region some scientists associate with conscious thought. The company acknowledges the analogy has limits, and notes plainly that LLMs are not brains. How much weight to give the comparison remains unsettled, including inside Anthropic.

The monitoring value is also partial. The J-lens offers glimpses rather than a complete view, and the absence of a signal does not prove a behavior is not present. McGrath framed it as a useful addition rather than a solution.

"It's like having an x-ray when what you really want is a Star Trek tricorder that shows you everything," he said, adding that auditing use cases would demand stronger guarantees.

Anthropic has partnered with Neuronpedia, an open-source platform for inspecting LLMs, to publish a hands-on demo. The collaboration lets outside researchers and developers test the tool directly, a step that will help verify how far the technique generalizes beyond the company's own examples.