An AlphaGo builder argues that LLMs do not really reason

On 10 March 2016, in a Seoul hotel, AlphaGo placed a stone on the fifth line of the board during the second game against Lee Sedol. The system itself estimated that a human player would have chosen that move with a probability of 1 in 10,000, and several commentators read it as a mistake. It was Move 37. AlphaGo won that game and closed out the match 4 to 1.

Ten years later, someone who watched that move live and helped build the program brings the scene back to open an op-ed in MIT Technology Review with an unhedged thesis: large language models do not reason, and we should not be fooled by text that looks every bit like reasoning.

An example chosen on purpose

The starting point is no accident. Move 37 looked absurd and turned out to be brilliant; many LLM answers look reasoned and may not be. There is a technical difference between the two cases that is worth keeping in mind when reading the full piece.

AlphaGo combined neural networks with Monte Carlo tree search: it explored possible continuations, evaluated them with its networks and picked the most promising one. There was an explicit process, tied to a measurable goal, which was winning the game. An LLM, by contrast, generates text token by token from learned patterns. So-called reasoning models, including Claude's extended thinking, produce an intermediate chain before answering, but that chain is also generated text, not a guaranteed trace of what happens inside the model.

What the research says

The debate is not settled by intuition, and Anthropic itself has published work that qualifies both positions. In an interpretability study published in March 2025, the team observed real intermediate steps inside Claude 3.5 Haiku: to answer what the capital of the state containing Dallas is, the model internally activated “Texas” before reaching “Austin”. It also found that the model planned the rhyme of a line of verse before writing it.

The same study found the opposite in other cases. When asked to explain how it had solved an addition, Claude described the schoolbook carry method, while its internal circuits followed different parallel paths. And when faced with hard problems that came with a hint, it could build its reasoning backwards to justify the answer it had already settled on.

A few weeks later, another Anthropic paper measured how faithful chains of thought are: Claude 3.7 Sonnet mentioned the hint it had used only 25% of the time, and DeepSeek R1 39% of the time. In June that year, Apple published the study The Illusion of Thinking, in which several reasoning models collapsed on puzzles such as the Tower of Hanoi once they passed a certain complexity threshold, a result other researchers disputed over the design of the evaluation.

The picture these papers leave is uncomfortable for both camps. There is structured computation inside these models, but what they say about how they reached an answer does not always match what they actually did.

Who it matters to

For anyone building with Claude Code or the API, the underlying debate has very practical consequences:

Do not treat visible reasoning as an audit log. If an agent explains why it modified a file, that explanation is no substitute for reviewing the diff.
Verify with deterministic mechanisms. Tests, linters and hooks such as PostToolUse let you check the outcome without relying on the model's account of it.
* Evaluate beyond the easy examples. If performance drops as complexity rises, it is better to find out before the client does.

For anyone buying or selling AI-based products, the warning is simpler: “reasons” is a word with little backing when it is used as a guarantee of reliability.

Our reading is that you do not need to settle whether an LLM reasons in order to work well with one, but you do need to design as if its explanation might not match what it actually did. The op-ed deserves credit for making that point with a move nobody in the field has forgotten.