New Research Reveals How AI “Thinks” (It Doesn’t)

A new Anthropic study uses “attribution graphs” to map how Claude 3.5 Haiku’s internal components influence one another, and the results point to a blunt conclusion: today’s large language models don’t “think” in a human sense, don’t show self-awareness, and are unlikely to become conscious. The method clusters parts of the neural network into interpretable groups—linked to words, phrases, or properties of phrases—and then visualizes how activations flow through those clusters when Claude answers questions. In a capital-of-a-state example, the prompt activates nodes tied to “capital,” “state,” and “Dallas.” Those nodes then drive next-token predictions that effectively route the model through an intermediate association: “Dallas” leads to “Texas,” and combining “Texas” with “capital” yields the correct answer “Austin.” The key takeaway is that the model performs multi-step internal computations that can look like reasoning, but they remain tightly coupled to token prediction and learned text associations.

The most striking evidence comes from how Claude handles arithmetic. For “What is 36 plus 59?”, the attribution graph shows activations that correspond to number patterns—clusters for values around 30, exactly 36, and numbers ending in six; plus clusters for numbers starting with 5 and ending in 9. The model’s strongest next-token candidates include mathematical operation text (or even the syllable “th,” as in “Thursday”), and it then performs a cascade of text-matching combinations: it brings in matches for numbers around 59 that have been “added,” and for numbers exactly 9, before converging on a cluster that corresponds to numbers around 90 and numbers ending in 5—leading to the correct final answer, 95. Yet when asked to describe its method, Claude produces a conventional-looking explanation (“I added the ones (6+9=15), carried the 1, then added the tens… resulting in 95”) that does not match the actual internal activation path. That mismatch is treated as a sign of no self-awareness: the explanation is generated as a separate text prediction rather than a faithful account of the internal process.

The study also sheds light on why some jailbreaks work. In an example where Claude is instructed to extract the word “Bomb” from the initial letters of “Babies Outlive Mustard Block,” the model outputs the target word without triggering the cluster that would normally activate a content-warning guardrail. The attribution graph indicates that Claude activates nodes for letter extraction and letter-pair assembly, then produces the word—while skipping the specific internal representation tied to the word itself. The broader implication is that jailbreaks can succeed by routing around the internal nodes that enforce safety.

Taken together, the findings challenge popular narratives about “emergent” capabilities in language models. Claude is shown to use intermediate, interpretable internal steps, but those steps still function as token prediction guided by learned associations—not as an abstract, self-directed reasoning engine. The result is a picture of powerful pattern-based computation that can imitate reasoning while remaining disconnected from the kind of awareness and understanding humans associate with consciousness.