Why LLMs are not actually reasoning

A former DeepMind researcher argues that large language models should not be mistaken for systems that truly reason. In his view, models like ChatGPT are strong at pattern completion and can generate useful stepbystep outputs, but that is not the same as maintaining an explicit, auditable reasoning process. The article contrasts LLMs with AlphaGo, which combined intuition with a separate search process that evaluated future possibilities. By comparison, chainofthought prompting still relies on nexttoken prediction and does not create a distinct reasoning mechanism, the author says. He also argues that these systems often lack a persistent record of beliefs, cannot clearly separate knowledge from manipulation of that knowledge, and may produce explanations after reaching an answer rather than showing the real path they took. The piece says these limits matter most in highstakes areas such as medicine, engineering, and scientific research, where users need to know not just what a system concluded but how it reached that conclusion. The author calls for new AI designs that keep an inspectable epistemic state and update beliefs only when evidence supports doing so.