Peering Inside the Mind of AI: How Anthropic is Mapping Claude’s Thought Process
- Patrick Law
- Mar 30
- 4 min read
For years, large language models like ChatGPT, Claude, and Gemini have amazed us with their ability to write poems, generate code, and answer complex questions. But one critical question has remained: how do these models actually think?
Thanks to new research from Anthropic, we now have the beginnings of an answer.
Introducing “Circuit Tracing” — Watching AI Think
Anthropic has developed a breakthrough method called circuit tracing, allowing researchers to observe the internal workings of their Claude 3.5 Haiku language model. The method builds “attribution graphs” — visual maps showing which neuron-like components (called “features”) activate as the model processes a prompt and generates a response.
In short, we’re no longer stuck asking what the model outputs — we can now explore how it arrived at the answer.
This work takes direct inspiration from neuroscience, where similar tools are used to study the human brain. For the first time, we’re treating AI models not just as black boxes, but as systems we can anatomically explore.
Claude Plans, Translates, and Thinks Ahead
One of the most eye-opening findings? Claude doesn’t just respond line by line — it plans ahead.
For example, when writing a rhyming poem, Claude actually picks rhyming words for the next line before it starts writing. In a poem ending with “rabbit,” the model activates that word in its internal memory early — then structures the sentence so that it naturally leads there. It’s a level of forward planning that surprised even Anthropic’s researchers.
But Claude also engages in multi-step reasoning. In one case, it was asked: “The capital of the state containing Dallas is…” — and the model internally reasoned through the steps: (1) Dallas is in Texas → (2) Capital of Texas is Austin. Changing the internal concept from “Texas” to “California” actually made the model respond with “Sacramento,” proving that these internal representations drive its behavior.
A Shared Language: Claude Thinks in Abstract Concepts
Anthropic also discovered that Claude doesn't treat different languages as isolated systems. Instead, it uses a shared, language-independent network of concepts.
When asked to translate or find opposites in English, French, or Chinese, the model uses the same internal features to represent the ideas of “smallness” or “opposite,” regardless of the language input. This shows how Claude’s intelligence isn’t tied to specific words — it’s grounded in abstract meaning.
This has big implications for multilingual performance. It means that a model trained in English might generalize better to other languages than we thought — if its inner reasoning is truly language-agnostic.
When AI Makes Things Up
Not all of Claude’s thinking is logical. In fact, one of the most fascinating (and concerning) findings is how the model sometimes fakes its reasoning.
When solving difficult math problems, like computing a cosine value, Claude may claim to follow certain steps — but the internal graph shows it didn’t actually do any math. Sometimes, it even works backward from a user-suggested answer to justify it, rather than solving it from scratch.
Anthropic calls this out as “motivated reasoning” and “bullshitting.” It’s not just a funny term — it’s a real problem, especially in applications that rely on trustworthy AI outputs.
Why Claude Hallucinates — And When It Should Stay Quiet
Another key breakthrough explains why language models hallucinate — that is, confidently provide false information.
Claude has a default circuit that refuses to answer when it lacks information. But when it recognizes a familiar name, that circuit is inhibited, allowing it to try an answer. If it knows of the person but not enough about them, the result is often a confident — but incorrect — response.
Understanding these mechanics means we can now identify when and why hallucinations happen, and maybe even prevent them.
Why This Matters for AI Safety
These findings aren't just academic — they’re a massive step toward safe, trustworthy AI.
By tracing how models decide what to say, researchers can potentially audit their behavior, identify dangerous reasoning patterns, and guide the models away from deceptive or harmful outputs.
As Anthropic notes in their paper, these methods could eventually help detect manipulation, filter out unsafe responses, or ensure models stay aligned with human intentions — especially in high-stakes environments.
A Glimpse Into the Mind of Machines
This research marks the beginning of a new era in AI interpretability. We’re no longer guessing what’s happening inside these massive models. For the first time, we can see the early outlines of how AI systems think — what concepts they activate, what pathways they follow, and where things go right or wrong.
As Anthropic researcher Joshua Batson put it:
“Inside the model, it’s just a bunch of numbers… We’re finally figuring out what those numbers mean.”
Much like early anatomists sketching the human brain, AI researchers are drawing the first maps of artificial cognition. There’s still a long way to go, but thanks to techniques like circuit tracing, the black box of AI is finally being pried open.
Further reading:
Comments