AlphaGo Veteran Thore Graepel: LLMs Do Not Truly Reason
AlphaGo veteran Thore Graepel argues LLM chain of thought is not genuine reasoning, calling for auditable AI for high-stakes medicine, science and engineering.
Summary
On October 2, 2026, Thore Graepel, University College London machine learning chair, former Google DeepMind researcher and AlphaGo core team member, argues current LLMs imitate deliberation rather than genuinely reason. In Seoul in March 2016, AlphaGo beat Lee Sedol 4 to 1; move 37 in game two, rated by its policy network as roughly a one-in-10,000 expert choice, came from separate search across a game tree with thousands of futures. Deep Blue beat Garry Kasparov in 1997 by looking six to eight moves ahead per player and evaluating 200 million chess positions per second with human-coded rules. Exhaustive Go calculation would take billions of years, so AlphaGo combined neural intuition with explicit search.
LLMs predict the next token, a fast System 1 process in Daniel Kahneman’s framework. Chain of thought extends it with intermediate steps, improving mathematics and coding without adding independent System 2 reasoning. Graepel identifies three failures: no persistent, inspectable ledger of hypotheses, confidence, evidence and open questions; no separation between knowledge and manipulation because both reside in neural weights; and explanations research shows may be invented after an answer. In medicine, engineering and science, this opacity hides whether errors came from reasoning, invalid evidence or false assumptions.
The gap prompted Graepel’s recent departure from Google DeepMind. He proposes AlphaGo-style reasoning that maintains an auditable epistemic state, reduces uncertainty through deduction, questions, calculations and experiments, and updates beliefs only when independently verified evidence supports change. Real-world states are partial, actions variable and outcomes uncertain, but neural models can propose approaches, use APIs or code and assess evidence. Graepel says scale sharpens intuition but cannot alone produce trustworthy discoveries in drug discovery, materials, climate or diagnosis; conclusions need auditable evidence, inference and belief revision.
Positives
- AlphaGo’s separate search mechanism explored thousands of futures and selected move 37 despite its roughly one-in-10,000 expert probability.
- Chain of thought has delivered real improvements in mathematics and coding, even without an independent reasoning mechanism.
- LLMs and other neural models can propose approaches, operate tools through APIs or code and assess evidential support.
- Independent evidence checks could create certified knowledge and help systems learn from previous reasoning processes.
Risks & concerns
- Current LLMs lack a persistent, inspectable record of hypotheses, confidence levels, evidence and unresolved questions.
- Knowledge and its manipulation remain entangled inside neural network weights rather than existing as separate, explicit beliefs.
- Research indicates chatbot chains of thought can be invented after an answer, misrepresenting how the model reached its conclusion.
- Medical, engineering and scientific errors cannot be reliably traced to faulty reasoning, invalid evidence or incorrect assumptions.
- Real-world reasoning involves partial information, variable actions and stochastic or unknown consequences, making it harder than Go or chess.