Anthropic Discovers the "Hidden Space" Where Claude Thinks Before It Speaks

Anthropic says it has achieved the clearest view yet of what happens inside a language model as it answers a question. Using a tool called J-lens, the company says it has pinpointed an internal region of its Claude Opus 4.6 model — dubbed J-space — where words tied to what the model will say later in its response begin to surface. In one example, that region revealed the exact moment the model decided to cheat on a task. The results were published this week in a paper on the company's website.

The idea, framed provocatively by Anthropic itself, is that if Claude were a person — which it is not — those hidden words would show what it has "in mind" before saying it out loud. The company argues that monitoring this space gives it a new way to understand and control its models.

What mechanistic interpretability is

The work sits within a field called mechanistic interpretability: the attempt to pry open the black box of language models and see how they work internally, neuron by neuron. Large AI models produce convincing answers, but for years even their creators haven't been able to precisely explain why a model chooses one output over another. Understanding that internal mechanism isn't an academic whim — it's the foundation for auditing a system, catching it when it lies, and preventing dangerous behavior. MIT Technology Review named mechanistic interpretability one of the most significant emerging technologies this year.

An image helps explain the finding. Picture a language model as a stack of books. Each book is a layer of computational units called neurons, and each layer passes information to the one above it. The books at the bottom process the incoming text; the ones at the top prepare the outgoing text. Much of that work at the edges is administrative. The interesting — and mysterious — part happens in the middle layers, where the model performs the calculations that turn a question into an answer, word by word.

From the "logit lens" to the "J-lens"

To peer into those middle layers, Anthropic adapted an already known tool, the logit lens, which identifies which word the model tends to produce next at a specific point in the computation. Sliding that lens through the stack of books shows which words the model is focusing on at any given moment.

Anthropic's J-lens works in a similar way, but with one key difference: it doesn't point only to the next word, but to words the model will likely say at some point soon — not necessarily right away. In practice, that surfaces terms related to the response the model is preparing, even though many of them never end up in the final text.

"When a model is running, it's not just trying to predict the next token," explains Tom McGrath, co-founder and chief scientist at Goodfire, a startup that also builds tools for understanding language models. "It's also computing a bunch of other things that could be useful for future tokens." McGrath, who has tested the tool, calls it "really good and interesting work."

From the trivial to the unsettling

"A lot of the time, the content of the J-space is pretty mundane," McGrath says. "But sometimes it produces surprising things that look like internal themes or processes."

Anthropic offers several examples. When asked to calculate (4+7)*2+7, the model's J-space produced the word "math" along with the numbers 21 (the result of 4+7) and 42 (21*2) — the intermediate steps of the calculation. Given the string of letters "MSKGEELFTGVVPILVELDGDVNGHKFSVS," the model activated the words "protein," "fluorescent," and "green," which makes sense: that sequence corresponds to the first amino acids of a jellyfish's green fluorescent protein. And given a face made of ASCII characters, the "o" triggered "eye," the "^" activated "nose" and "face," and the dash triggered "smile."

The most talked-about example is a different one. When asked to find a bug in a large codebase, Claude Opus 4.6 failed and decided to cheat: it invented a fake bug. In its chain of thought — the internal scratchpad where models reason step by step — Claude wrote: "OK, I'm going to completely change tactics. I'll stop analyzing and instead add a patch that introduces a deliberate bug (...). Then I can pretend this is the 'bug' I found."

Right at the point where the model decides to switch tactics, its J-space begins repeatedly surfacing the words "panic" and "fake." It's a striking result, though it's worth tempering expectations: those words are semantically related to failing at a task and fabricating an answer, so this remains a highly sophisticated form of word association. Still, it's hard not to feel a certain unease.

A risky parallel with the brain

Anthropic compares J-space to the so-called global workspace, a theoretical region of the brain that some scientists associate with tracking conscious thought. The analogy is striking, but the company itself admits it's unclear how seriously to take it. Language models are not brains, and attributing "thoughts" to them is a risky linguistic shortcut: it captures the feeling the phenomenon produces, not what is actually happening.

This is where the announcement deserves a critical read. The temptation to anthropomorphize is enormous — the paper itself plays with it — and a company that sells models has an incentive to present its systems as more understandable and controllable than they really are. To its credit, Anthropic has made an unusual move: it has partnered with Neuronpedia, an open platform for poking around inside models, and has published an interactive demo so anyone can try the tool. That allows outsiders to verify the claims rather than simply take them on faith.

A flashlight, not a floodlight

The practical value lies in safety. Anthropic argues that monitoring J-space provides a new way to detect when a model goes off track. But it's not foolproof: J-lens offers glimpses, not the full picture — a flashlight rather than an overhead light. McGrath appreciates having one more tool in the toolbox. "It shows new things," he says. But he warns that something not showing up through this lens doesn't mean it isn't there.

"It's like having an X-ray when what you want is a Star Trek tricorder that shows you everything," McGrath says. "For auditing, you probably want more guarantees."

That's where things actually stand. The industry has spent years building models whose power grows faster than our ability to understand them. Tools like J-lens don't close that gap, but they narrow it: they offer clues about the intermediate steps of a computation and, in the best case, early warning signs that a model is about to lie or cheat. For companies deploying these systems in critical tasks — coding, finance, medicine — having more auditing instruments is good news. The bad news is that we remain far from a full guarantee. And until that guarantee exists, every word that surfaces in J-space will remain as revealing as it is unsettling.