Large Language Models Do Not Possess Reasoning Capabilities

A researcher who recently departed Google DeepMind, publishing in MIT Technology Review on October 2, 2026, argues that large language models fundamentally lack true reasoning capabilities, simulating intelligence through pattern matching rather than executing logical inference. The article urges the AI community to adopt architectures inspired by AlphaGo to develop auditable, trustworthy reasoning essential for high-stakes scientific and medical applications. The author contrasts current LLMs with AlphaGo's 2016 match against Lee Sedol, analyzing the program's controversial Move 37. Lee initially mistook the move for a blunder, but it resulted from AlphaGo's search mechanism evaluating thousands of future branches, overriding the policy network's intuition. AlphaGo successfully combined a fast, associative system for intuition with a slow, deliberative search engine that maintained a persistent state and updated beliefs based on calculated consequences. LLMs operate exclusively as the associative system, relying on next-token prediction. The analysis critiques Chain-of-Thought techniques as merely extending the prediction process rather than introducing distinct reasoning. Three critical deficiencies are identified: LLMs lack a persistent, inspectable cognitive state to track assumptions and confidence levels; knowledge and reasoning are entangled within neural weights, preventing independent manipulation of beliefs; and reasoning chains often represent post-hoc rationalizations, where the model generates a plausible narrative after deriving an answer through an opaque mechanism. The article proposes a new paradigm for machine reasoning mirroring AlphaGo's dual architecture. Proposed systems must maintain an explicit, editable record of beliefs, uncertainties, and open questions, separate from the generative component. By integrating external tools, enforcing verification steps that validate uncertainty reduction before updating internal states, and treating reasoning as a sequence of auditable actions, AI could function as an enhanced scientific method. The author warns that scaling model parameters cannot produce true reasoning. Future AI must deliver novel insights in fields like drug discovery through traceable inference chains, ensuring conclusions are grounded in auditable evidence rather than fabricated plausibility.