Former DeepMind researcher Thore Graepel pitches AlphaGo-style search to curb AI hallucinations
Thore Graepel, a former DeepMind researcher, is building Metis Reasoning to apply AlphaGo's search principle to general AI reasoning. The company proposes an explicit knowledge ledger that would let every conclusion be checked against evidence, though no working system or benchmarks have yet been published.
Former DeepMind researcher Thore Graepel is pursuing an approach to one of the most persistent problems in artificial intelligence: the tendency of language models to produce confident but false statements, commonly called hallucinations. Through his startup Metis Reasoning, Graepel wants to carry the search principle that powered DeepMind's AlphaGo into general reasoning tasks.
The core idea is to stop storing knowledge solely inside a model's weights, where it is difficult to inspect. Instead, Metis proposes an explicit knowledge ledger, a structured record against which each conclusion an AI system reaches could be checked against supporting evidence. In this design, a claim would not stand on the model's internal confidence alone but on traceable sources. The approach shifts the burden of proof from the system's fluency to the evidence it can point to, which is the property that users of current chatbots most often lack.
The comparison to AlphaGo is deliberate. The Go-playing system, developed at DeepMind, combined learned intuition with a systematic search through possible moves, allowing it to evaluate options rather than simply reacting to patterns. Graepel's argument is that a similar combination of learned components and explicit search could make AI reasoning more reliable, because the system would have to justify its answers rather than generate them in a single pass. The analogy rests on the idea that search lets a system test alternatives before committing to one, a step that standard language-model generation does not perform.
It is important to note the current state of the project. According to the available reporting, there is no working system yet, and no benchmarks have been published to measure how the approach performs against existing models or against established tests of factual accuracy and reasoning. The proposal therefore remains a research direction and a company thesis rather than a demonstrated result. No external evaluations, customer deployments or peer-reviewed results have been cited so far.
Hallucinations have become a central concern for developers and users of large language models, particularly in fields where errors carry real costs, such as medicine, law, and scientific research. Efforts to reduce them have included retrieval systems that pull in external documents, and training methods that reward accurate answers. Graepel's approach would place verification at the center of the architecture, rather than treating it as an add-on. That difference matters: retrieval can supply sources without guaranteeing that the final conclusion actually follows from them, which is the gap a checkable ledger would aim to close.
Whether an explicit ledger can scale to the breadth of human knowledge, and how it would handle contested or ambiguous claims, are questions the project has yet to answer publicly. Until Metis releases a working prototype and measurable results, its promise of checkable AI conclusions will remain a hypothesis to be tested.
8
