A strong AI answer is not the same as a checked reason
Thore Graepel contrasts AI text that sounds thoughtful with a system that tracks evidence and open questions. This is his argument about how to check an answer, not a test proving all AI systems lack reason.
What the source reports
Thore Graepel writes in MIT Technology Review about how AI shows its work. He compares AI text with the search that game systems such as AlphaGo use. To him, smooth words do not prove that a system kept track of facts. The words may not show what would make it change its answer. He wants a record of facts and open questions that people can check. This is an expert view, not a test of every AI system. People may also disagree on what counts as reasoning. Before a high-stakes choice, ask what source backs the answer. Then name a fact that would make you change course.
Original source
- Title
- Don’t be fooled-LLMs don’t reason
- Author
- Thore Graepel
- Publication
- MIT Technology Review
- Date
- Friday, October 2, 2026
Also covered by the other newsletter.