# Thore Graepel argues in MIT Technology Review that LLMs don’t reason [Canonical reading](https://escapemeridian.com/readings/thore-graepel-argues-in-mit-technology-review-that-llms-don-t-b239413a) On October 2, 2026, Thore Graepel argues in MIT Technology Review that AlphaGo’s striking Go move came from search that weighed possible futures, not from intuition alone. He contrasts that approach with large language models, which produce text by predicting one token at a time; chain of thought extends that process but, he argues, does not add a separate reasoning mechanism. Graepel says today’s chatbots lack an explicit, persistent record of what they know, doubt and still need to resolve. Their knowledge and reasoning are intertwined, and their written chains of thought can describe a route different from the one that produced an answer. In medicine, engineering and research, he argues, this makes it harder to identify whether a mistake came from reasoning, evidence or assumptions. Graepel proposes systems that maintain and update an epistemic state, with an independent part evaluating whether each step reduces uncertainty on the basis of evidence. He argues that this could support auditable reasoning and creative insights in fields such as drug discovery, materials, climate and diagnosis. His conclusion is that scaling intuition alone will not produce trustworthy machine intelligence; the open question is how to build systems that can reason in complex, partially known real-world settings. Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims. ## Original source MIT Technology Review · Article Original source date: Oct 2, 2026 (2026-10-02T08:00:00Z). Date basis: Date from the publisher’s own records, through its API. Added to Escape Meridian: 2026-11-09T00:58:25.347Z. First observed: 2026-11-09T00:41:47.917Z. Retrieved: 2026-11-09T00:43:34.916Z. - [Don’t be fooled—LLMs don’t reason](https://www.technologyreview.com/2026/10/02/1145639/dont-be-fooled-llms-dont-reason/) ## Limits Text as retrieved on November 9, 2026; originally published October 2, 2026. Graepel argues for an approach to machine reasoning but this piece does not report a demonstration that the proposed general reasoning systems work. Graepel notes that open-world reasoning is harder than board games because the real-world state is only partly known and actions and their consequences vary or may be unknown. ## Open question How can a system reliably evaluate whether a step reduces uncertainty when the real-world state and the consequences of actions may be only partly known? ## Source-stated Graepel argues AlphaGo’s move came from search that evaluated possible futures, rather than intuition alone. > What made AlphaGo choose it was the program’s search machinery, which looked beyond immediate plausibility and weighed the future consequences of proposed moves. - [Supporting source](https://www.technologyreview.com/2026/10/02/1145639/dont-be-fooled-llms-dont-reason/) ## Source-stated Graepel argues that chain of thought does not introduce a separate reasoning mechanism because it is still generated through next-token prediction. > The intermediate reasoning is still produced by the same next-token prediction process, iterated for longer before the model commits to an answer. - [Supporting source](https://www.technologyreview.com/2026/10/02/1145639/dont-be-fooled-llms-dont-reason/) ## Source-stated Graepel identifies missing explicit epistemic state, intertwined knowledge and reasoning, and chains of thought that may report a different route to an answer as shortcomings of chatbots. > research has demonstrated that the bots often concoct them after the fact, reaching an answer by one route but reporting another. - [Supporting source](https://www.technologyreview.com/2026/10/02/1145639/dont-be-fooled-llms-dont-reason/) ## Source-stated Graepel argues that high-stakes applications require being able to identify whether mistakes arise from reasoning, evidence or assumptions. > Was the system’s reasoning at fault, did it draw on invalid evidence, or did it make incorrect assumptions? - [Supporting source](https://www.technologyreview.com/2026/10/02/1145639/dont-be-fooled-llms-dont-reason/) ## Source-stated Graepel proposes that systems maintain an epistemic state and have an independent part evaluate each move by how much it resolves uncertainty, updating beliefs only when supported by evidence. > an independent part of the system must evaluate each move by how much it actually resolves uncertainty, updating beliefs only when the change is backed by evidence. - [Supporting source](https://www.technologyreview.com/2026/10/02/1145639/dont-be-fooled-llms-dont-reason/) ## Source-stated Graepel concludes that scaling intuition alone will not produce trustworthy machine intelligence and argues that society needs creative moves in fields including drug discovery and diagnosis. > I do not think we reach trustworthy machine intelligence by making system 1 bigger. Scale sharpens intuition, but it does not make intuition more deliberative. - [Supporting source](https://www.technologyreview.com/2026/10/02/1145639/dont-be-fooled-llms-dont-reason/) ## Source and contributor context MIT Technology Review · Publisher The publisher is identified here. No individual author has been inferred. ## Source media Media retrieval status: unchecked. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Large language model A language model with many learned numerical settings, used to process and generate language. - [Definition reference](https://developers.google.com/machine-learning/glossary#large-language-model) ## Chain of thought The step-by-step working a model writes out before giving its final answer. - [Definition reference](https://developers.google.com/machine-learning/glossary#chain-of-thought-prompting) ## TIME TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services. - [Explore this context](/entities/time) - [Reference](https://time.com/) ## Editorial provenance Version 1. Publication ID: 8e176980-6c6c-029e-7a31-5ede872d2099. Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5. Source check: passed by anthropic/claude-opus-5.5 on 2026-11-09T00:54:12.346Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: bridge\_admitted\_by\_model. Observation: 6873b93bb319cc62a5083c3bd525f7bf6db986ab1da16546c29fe46c8fe77145. Source text hash: 0fc4f8bfe2b309e63b1bf775c7b256cabeeab6d8ba0f39b93fb9bfbb7b26bb08. ## Revision and source-check history ### Version 1 — 2026-11-09T00:58:25.347Z On October 2, 2026, Thore Graepel argues in MIT Technology Review that AlphaGo’s striking Go move came from search that weighed possible futures, not from intuition alone. He contrasts that approach with large language models, which produce text by predicting one token at a time; chain of thought extends that process but, he argues, does not add a separate reasoning mechanism. Graepel says today’s chatbots lack an explicit, persistent record of what they know, doubt and still need to resolve. Their knowledge and reasoning are intertwined, and their written chains of thought can describe a route different from the one that produced an answer. In medicine, engineering and research, he argues, this makes it harder to identify whether a mistake came from reasoning, evidence or assumptions. Graepel proposes systems that maintain and update an epistemic state, with an independent part evaluating whether each step reduces uncertainty on the basis of evidence. He argues that this could support auditable reasoning and creative insights in fields such as drug discovery, materials, climate and diagnosis. His conclusion is that scaling intuition alone will not produce trustworthy machine intelligence; the open question is how to build systems that can reason in complex, partially known real-world settings. ## Original source MIT Technology Review · Article Original source date: Oct 2, 2026 (2026-10-02T08:00:00Z). Date basis: Date from the publisher’s own records, through its API. Added to Escape Meridian: 2026-11-09T00:58:25.347Z. First observed: 2026-11-09T00:41:47.917Z. Retrieved: 2026-11-09T00:43:34.916Z. - [Don’t be fooled—LLMs don’t reason](https://www.technologyreview.com/2026/10/02/1145639/dont-be-fooled-llms-dont-reason/) ## Limits Text as retrieved on November 9, 2026; originally published October 2, 2026. Graepel argues for an approach to machine reasoning but this piece does not report a demonstration that the proposed general reasoning systems work. Graepel notes that open-world reasoning is harder than board games because the real-world state is only partly known and actions and their consequences vary or may be unknown. ## Open question How can a system reliably evaluate whether a step reduces uncertainty when the real-world state and the consequences of actions may be only partly known? ## Source-stated Graepel argues AlphaGo’s move came from search that evaluated possible futures, rather than intuition alone. > What made AlphaGo choose it was the program’s search machinery, which looked beyond immediate plausibility and weighed the future consequences of proposed moves. - [Supporting source](https://www.technologyreview.com/2026/10/02/1145639/dont-be-fooled-llms-dont-reason/) ## Source-stated Graepel argues that chain of thought does not introduce a separate reasoning mechanism because it is still generated through next-token prediction. > The intermediate reasoning is still produced by the same next-token prediction process, iterated for longer before the model commits to an answer. - [Supporting source](https://www.technologyreview.com/2026/10/02/1145639/dont-be-fooled-llms-dont-reason/) ## Source-stated Graepel identifies missing explicit epistemic state, intertwined knowledge and reasoning, and chains of thought that may report a different route to an answer as shortcomings of chatbots. > research has demonstrated that the bots often concoct them after the fact, reaching an answer by one route but reporting another. - [Supporting source](https://www.technologyreview.com/2026/10/02/1145639/dont-be-fooled-llms-dont-reason/) ## Source-stated Graepel argues that high-stakes applications require being able to identify whether mistakes arise from reasoning, evidence or assumptions. > Was the system’s reasoning at fault, did it draw on invalid evidence, or did it make incorrect assumptions? - [Supporting source](https://www.technologyreview.com/2026/10/02/1145639/dont-be-fooled-llms-dont-reason/) ## Source-stated Graepel proposes that systems maintain an epistemic state and have an independent part evaluate each move by how much it resolves uncertainty, updating beliefs only when supported by evidence. > an independent part of the system must evaluate each move by how much it actually resolves uncertainty, updating beliefs only when the change is backed by evidence. - [Supporting source](https://www.technologyreview.com/2026/10/02/1145639/dont-be-fooled-llms-dont-reason/) ## Source-stated Graepel concludes that scaling intuition alone will not produce trustworthy machine intelligence and argues that society needs creative moves in fields including drug discovery and diagnosis. > I do not think we reach trustworthy machine intelligence by making system 1 bigger. Scale sharpens intuition, but it does not make intuition more deliberative. - [Supporting source](https://www.technologyreview.com/2026/10/02/1145639/dont-be-fooled-llms-dont-reason/) ## Source and contributor context MIT Technology Review · Publisher The publisher is identified here. No individual author has been inferred. ## Source media Media retrieval status: unchecked. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Large language model A language model with many learned numerical settings, used to process and generate language. - [Definition reference](https://developers.google.com/machine-learning/glossary#large-language-model) ## Chain of thought The step-by-step working a model writes out before giving its final answer. - [Definition reference](https://developers.google.com/machine-learning/glossary#chain-of-thought-prompting) ## TIME TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services. - [Explore this context](/entities/time) - [Reference](https://time.com/) ## Editorial provenance Version 1. Publication ID: 8e176980-6c6c-029e-7a31-5ede872d2099. Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5. Source check: passed by anthropic/claude-opus-5.5 on 2026-11-09T00:54:12.346Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: bridge\_admitted\_by\_model. Observation: 6873b93bb319cc62a5083c3bd525f7bf6db986ab1da16546c29fe46c8fe77145. Source text hash: 0fc4f8bfe2b309e63b1bf775c7b256cabeeab6d8ba0f39b93fb9bfbb7b26bb08. ### passed — 2026-11-09T00:54:12.346Z — anthropic/claude-opus-5.5 Publication: 8e176980-6c6c-029e-7a31-5ede872d2099 attribution — pass: The reading attributes the opinion piece to Thore Graepel in MIT Technology Review, which matches the byline note and publisher. The arguments are consistently framed as his ('he argues', 'Graepel proposes') rather than as established fact or EM opinion. > Thore Graepel is chair of machine learning at University College London. He was a core member of the AlphaGo team at DeepMind summary — pass: The summary follows the source's argument: AlphaGo's search versus its intuition, LLM next-token prediction and chain of thought, three shortcomings, high-stakes auditability, the proposed epistemic-state system, and the conclusion against scaling system 1. LLM and chain of thought are defined beside the reading, and epistemic state is explained in plain language. A small fidelity gain, not required, would be to identify AlphaGo as the Go program Graepel helped build and note that he left Google DeepMind to pursue this approach. The question follows from the source's own open-world caveat. > This is why I recently left my position at Google DeepMind. I believe we need a fresh approach to machine reasoning—one that draws on AlphaGo’s architecture. date — pass: October 2, 2026 matches the UTC calendar day of evidence.originalPublished. The page was modified on October 28, but every claim relies on core article text that is plainly part of the original piece. The reading does not use the later related-story sidebar and contains no hindsight framing. Graepel's proposal is presented as a proposal, not as an outcome. support — pass: Each claim is backed by a verbatim excerpt from the body of the article. The claimed benefits, auditable reasoning and creative insights in drug discovery and diagnosis, are framed as Graepel's argument. A limitation notes that no working system is demonstrated. The question draws on Graepel's own acknowledgment that open-world reasoning is harder. The October 28 edit does not affect these passages, which are plainly original content. > In the real world the current state of affairs is only partially known, the set of available actions is large and variable, and the consequences of actions are stochastic or unknown. media — pass: No media is selected, and no verified media candidates were supplied. That is appropriate for this text-only opinion piece. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/b239413a7807e2f350d311a6a2591aa11224d8ddd43ce40f76020123082baf0f/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/thore-graepel-argues-in-mit-technology-review-that-llms-don-t-b239413a.json)