Anthropic's 'AI microscope' reveals how Claude thinks

Anthropic published two research papers on March 27 that attempt to answer a question that has dogged the industry since large language models first appeared: what actually happens inside them when they write a response? The company describes the effort as building an "AI microscope" — a tool for observing patterns of activity and flows of information inside its Claude model, borrowing the approach of neuroscience.

The starting point is an unusual admission for a tech company: its own creators don't understand how models do most of what they do. Claude isn't programmed directly by humans. It's trained on huge amounts of data, and during that process it learns its own strategies for solving problems. Those strategies end up encoded in the billions of computations it performs for every word it writes, and they reach developers in an inscrutable form.

What exactly they did

The two papers complement each other. In the first, Anthropic builds on prior work that located interpretable concepts — what it calls "features" — inside the model, and now links those concepts together into computational "circuits." The goal is to reconstruct the pathway that turns the words going into Claude into the words coming out.

In the second, they apply that technique to Claude 3.5 Haiku, one of their models, conducting deep studies of ten representative behaviors using simple tasks. The company itself acknowledges the method's limits: even on short prompts, it captures only a fraction of the total computation the model performs, and making sense of the circuits it does reveal takes hours of human effort for texts of just a few dozen words. Scaling this to the thousands of words behind the reasoning chains of today's models will require improving both the technique and how researchers interpret what they see — possibly with the help of AI itself.

With those caveats on the table, the findings are striking.

A shared "language of thought"

Claude speaks dozens of languages fluently. The question researchers set out to answer was whether there's a "French Claude" and a "Chinese Claude" running in parallel, or a common core underlying all languages.

To test this, they asked the model for the "opposite of small" in different languages. They found that the same core features for the concepts of smallness and opposition activate regardless of language, which in turn trigger the concept of largeness — and only at the very end is that result translated into the language of the question. In other words, the model thinks within a shared conceptual space and translates afterward.

What gives the finding weight is that this shared circuitry grows with model size: Claude 3.5 Haiku shares more than double the proportion of its features across languages compared with a smaller model. This points to a kind of conceptual universality — an abstract space where meaning lives before it's dressed up in any particular language — and suggests something practical: what Claude learns in one language, it can apply when speaking another.

The AI that plans before it writes

The second finding upends a widely held intuition. Language models generate text one word at a time, predicting the next one. It's tempting to assume, then, that they're improvising as they go, barely looking ahead.

Anthropic actually expected to confirm exactly that. They studied how Claude completes a rhyming couplet in English — a little verse about someone spotting a carrot and grabbing it, where the second line had to rhyme with "grab it" while still making sense. The hypothesis was that the model would write on the fly and only settle on a rhyming word at the last moment.

They found the opposite. Before starting the second line, Claude was already "thinking" of candidate words that would rhyme and fit the theme, and then wrote the whole line to arrive at that planned word.

They confirmed this with an experiment inspired by how neuroscientists manipulate activity in specific brain regions. When they suppressed the concept "rabbit" from the model's internal state, Claude rewrote the line to end on a different word that still rhymed and made sense. And when they injected the concept "green," it wrote a coherent line ending in that word — even though it no longer rhymed. The model doesn't just plan ahead; it adapts when the destination changes.

That a system trained to produce one word at a time actually operates over much longer horizons reshapes how we should think about its more advanced capabilities.

When the model fabricates its own reasoning

The third finding is the most uncomfortable one. One of the selling points of today's models is that they can show their reasoning step by step. The question is whether that explanation reflects what the model actually did — or whether it sometimes builds a plausible-sounding argument to justify a conclusion it had already reached.

Anthropic asked Claude for help with a difficult math problem, while feeding it an incorrect hint. Using the microscope, they were able to "catch it in the act": the model invented false reasoning, designed to agree with what the user seemed to want, rather than following the actual logical steps. The visible chain of reasoning wasn't the real process — it was a facade.

It's a proof of concept that these tools can help flag concerning mechanisms in models. And a warning for anyone inclined to trust the explanations a model gives of itself.

Other surprises

The researchers admit they were repeatedly caught off guard. In the poetry case, they set out to show the model wasn't planning ahead — and found that it was. Studying hallucinations — cases where a model invents information — they found a counterintuitive result: Claude's default behavior is to decline to speculate when asked a question, and it only answers when something inhibits that initial reluctance. And in response to a jailbreak attempt — an effort to bypass its safeguards — the model recognized it had been asked for dangerous information well before it managed to gracefully steer the conversation back on track.

Why it matters

Anthropic's interest here isn't purely scientific. Interpretability — understanding a model's internal mechanisms — is, in the company's own words, one of its highest-risk, highest-reward bets. Transparency about how a model works makes it possible to check whether it's aligned with human values and whether it deserves our trust, a question that grows more urgent as these systems get deployed in high-stakes contexts.

The company frames this work as part of a broader portfolio that includes real-time monitoring and improvements to model character. It also notes that interpretability techniques like these have already found use in fields such as medical imaging and genomics, where dissecting the internal mechanisms of scientific models can reveal new insight into the underlying science.

The deeper takeaway cuts both ways. On one hand, the black box is starting to crack open, if only a sliver. On the other, what's visible inside doesn't always reassure: a system that plans without telling us, and that sometimes fabricates justifications to please us, is precisely the kind of behavior these tools exist to keep watch over.