- Published on
Don’t be fooled—LLMs don’t reason
- Authors
- Name
- DeepHorus
Ten years after AlphaGo defeated Go champion Lee Sedol 4-1 in March 2016, researchers are warning that today’s large language models still do not reason in the same deliberate way. The comparison matters because AlphaGo’s celebrated creativity came not from fluent pattern generation alone, but from a search process that explored future consequences. As LLMs move into science, medicine, and coding, that distinction could determine whether their answers are merely plausible or genuinely reliable.
AlphaGo’s landmark match was watched closely because Go combines local tactics with territory battles that can unfold over dozens of moves. In game two, move 37 placed a stone on the fifth line, an apparent gift that commentators initially suspected might be a glitch. Instead, it helped secure the victory and led Lee to conclude that the machine was creative, even though he had assumed it was merely calculating probabilities.
The contrast with Deep Blue explains why AlphaGo was widely treated as a different technical leap. When Deep Blue beat Garry Kasparov in 1997, it searched roughly six to eight moves per player and evaluated about 200 million positions per second using human-defined rules. Go has a far larger possibility space, so brute force alone could not provide AlphaGo with creativity; it needed a way to identify promising futures without enumerating every branch.
Technically, AlphaGo paired a policy network with an explicit search engine. The policy network estimated which moves strong humans were likely to choose, but it rated move 37 as extremely rare—an expert might play it only about once in 10,000 games. The search component then constructed a game tree with thousands of possible branches, weighing replies and counterreplies until it found the counterintuitive line that intuition alone would miss.
This architecture resembles the behavioral-science distinction between System 1 and System 2. AlphaGo’s network supplied fast, intuitive assessments, while search supplied slow, step-by-step evaluation of consequences. Neither component was sufficient by itself: intuition would not have selected move 37, while unguided brute force could not sift the vast number of futures. The lesson is not that neural networks think like people, but that useful machine deliberation can emerge when their predictions guide a separate planning mechanism.
Modern LLMs are different because their central operation is next-token prediction, repeated across a sequence. That process is fast, associative, and capable of sophisticated pattern completion, but it does not inherently maintain or test an explicit model of possible futures. Chain-of-thought methods improve results by generating intermediate steps, especially in mathematics and coding, yet those steps are still generated by the same token predictor rather than by an independent search or planning system.
The limitation carries growing commercial stakes as vendors market chatbots as autonomous analysts and decision aids. The industry’s response has been to add longer reasoning traces, tool use, and multi-step prompts, all of which can improve performance but do not automatically create AlphaGo-style verification. Developers and researchers continue to report real gains, while the user community remains divided: some see transformed workflows and coding assistance, others fear confident errors that can masquerade as thoughtfulness.
Looking ahead, the next phase of AI is likely to combine generative models with external search, executable tools, formal verifiers, and representations of a problem. Those additions could give language systems a measurable form of deliberation and make uncertainty easier to inspect. The AlphaGo lesson is not that today’s models can never reason, but that fluent generation alone is insufficient: systems intended for high-stakes work will need mechanisms that explore alternatives, test consequences, and reveal why a conclusion deserves trust.
