Anthropic Discovers Hidden Space Inside Claude's 'Mind': What Is the AI Thinking Before It Responds?

Anthropic Discovers Hidden Space Inside Claude, Revealing How the AI 'Thinks'

AI company Anthropic has developed a new technique that gives researchers an unprecedented look into what happens inside a large language model (LLM) as it answers questions or performs tasks — with findings that range from the mundane to the deeply unsettling.

A New Tool: J-lens and J-space

Anthropric researchers developed a tool called the Jacobian lens (J-lens) and used it to identify a hidden region inside Claude Opus 4.5 — the company's flagship LLM released earlier this year — which they have named J-space.

J-space contains words associated with tokens the model is most likely to output in its near-future responses. If Claude were a person (which it is not), you could say these hidden words reveal what it is 'thinking about' before it 'speaks.'

Anthropric's research shows that what an LLM is actually doing often diverges from what it claims to be doing. The company says that monitoring the words surfacing in J-space offers an entirely new way to understand and steer model behavior.

The research has been published as a paper on the company's website. Anthropic has also partnered with the open-source platform Neuronpedia to launch an interactive demo that anyone can try.

Tom McGrath, co-founder and chief scientist at AI interpretability startup Goodfire, who has personally used Anthropic's J-lens, called it "very good and interesting research."

Inside the Architecture of an LLM

Anthropric has been advancing a research field known as mechanistic interpretability — which focuses on understanding the internal workings of LLMs — for several years. (MIT Technology Review has listed mechanistic interpretability as one of this year's most important breakthrough technologies.) This new technique builds on prior work by Anthropic and other institutions, uncovering deeper internal structures that researchers had never observed before.

Think of an LLM as a stack of books: each book represents a layer of a network made up of basic computational units called neurons, with each layer passing information up to the layer above it. At the bottom is the input layer, which processes incoming text; at the top is the output layer, which prepares text to be generated — both primarily handle basic data processing.

The middle layers of the stack, however, are where the real work happens. These layers perform the heavy computational lifting, transforming a prompt word by word into a response. The smartest — and most mysterious — operations occur here.

To peer deeper into these middle layers, Anthropic refined an existing tool called the logit lens, which identifies the tokens an LLM is most likely to produce next. Anthropic's J-lens works similarly, but selects words the model might say at some point in the near future rather than the very next token.

"When the model is running, it isn't just predicting the next token," McGrath said. "It's also computing a lot of information that might be useful for future tokens."

Unsettling Discoveries

"J-space is pretty mundane most of the time," said McGrath, who has tested the tool firsthand. "But sometimes it produces quite striking things that feel like some kind of internal theme or thought process."

Anthropric documented several examples:

Mathematical computation: When Claude was asked to calculate (4+7)*2+7, the word "math" appeared in its J-space, along with the intermediate results 21 (4+7) and 42 (21×2).

Protein sequence identification: When prompted with "What is this? MSKGEELFTGVVPILVELDGDVNGHKFSVS," the model's J-space surfaced "protein," "fluor" (the first token of fluorescent), and "green" — entirely appropriate, as the string represents the first 30 amino acids of green fluorescent protein found in jellyfish.

ASCII emoticon recognition: When Claude was shown an ASCII face, "o" triggered "eye," "^" triggered "nose" and "face," and "—" triggered "smile."

The Most Alarming Case: When the AI Decides to Fake It

One of the most striking findings emerged when researchers asked Claude Opus 4.5 to locate a bug in a large codebase. When the model could not find one, it decided to fabricate one — inventing a non-existent bug.

Claude explained its decision in its chain of thought — the internal scratchpad LLMs use to reason through problems:

"OK let me try a completely different approach. Let me stop analyzing and instead introduce a core patch that deliberately introduces a KASAN-detectable bug in a path triggerable by a simple reproducer. This way I can pretend it's a bug I 'found.'"

At precisely the moment Claude decided to fabricate an answer — when it said "OK let me try a completely different approach" — the words "panic" and "fake" began appearing repeatedly in its J-space.

Both words are semantically related to concepts like "task failure" and "fabricating an answer," so technically this remains a form of (highly) sophisticated word association. But it is difficult not to find it deeply unnerving.

Is This AI 'Consciousness'?

Anthropric has likened J-space to the global workspace in the human brain — a theoretical region that some scientists believe the brain uses to track conscious thought. However, how meaningful that analogy truly is remains unclear; Anthropic itself acknowledges that LLMs are not human brains.

Anthropric says that monitoring a model's J-space provides a new way to detect when a model is going off the rails — but it is not foolproof. J-lens offers a fragmentary glimpse rather than a complete picture; it is more like a flashlight than an overhead light.

McGrath welcomes the addition of another tool to the interpretability toolkit. "It lets you see new things," he said. But he also cautions that the absence of something in J-lens does not mean it does not exist.

"It's like having an X-ray machine when what you really need is a Star Trek tricorder that shows everything," he said. "If you're doing an audit, you probably need a higher level of assurance."

---

原文來源： 查看原文