MIT study finds AI explanations make people more confidently wrong, but doctors resist
alex_verem · x · 2026-09-14
An MIT experiment tested what happens when AI explains its reasoning to humans, and found the explanation itself can be dangerous.
- Researchers showed real photos of skin conditions to 623 laypeople and 153 family physicians, with an AI diagnosis beside each image. Some saw the answer alone; others also got a plain-English explanation.
- The explained version was by far the most convincing—and the most damaging when the AI was wrong. When right, explanations helped people land on correct answers; when wrong, they walked people straight into mistakes, with more confidence than those who saw no explanation at all.
- Doctors were fine: they formed their own view from the photo first, then weighed the AI's input.
- Changing one thing helped: having people write down their own answer before seeing the AI's output countered the anchoring effect.
More from AGI Musings
- Musk: AI and robotics are the only path to universal high income — elonmusk · 2026-09-14
- Researcher resigns from Anthropic, slamming OpenAI and Anthropic's superintelligence race — vkrakovna · 2026-09-14
- antirez slams AI-written tweets: 'Your tweets must be your most cared thoughts' — antirez · 2026-09-14
- AI is taking over CFD and CAD workflows: what engineers must prepare for — burhop · 2026-09-14
- AI engineering in 2026: ten jobs wearing one hoodie — mdancho84 · 2026-09-14
- Jan Kulveit: Most Current Advanced AIs Are Likely on Team Humanity, but the Roadmap May Be Cursed — jankulveit · 2026-09-14