Doctor's thread: AI beats doctors who use AI on medical reasoning, human-in-loop can hurt
jonc101x · x · 2026-09-27
Physician jonc101x shares a 9-part thread of takes he "wouldn't have said out loud three years ago":
- AI likely knows more than most doctors, but unlimited knowledge plus zero accountability doesn't make it trustworthy
- Multiple studies show chatbots outperform doctors who are using those same systems on medical reasoning assessments — contradicting informatics' fundamental theorem that human+computer beats either alone
- Non-medically trained people using AI for triage questions did worse than those without AI
- Medical training is largely about framing the question; without it, AI sycophantically agrees with you while leading you down the wrong rabbit hole. Tip: don't tell AI what you suspect — it's inclined to agree
- Hallucinations are less of an issue than a few years ago; harms of omission are the real danger — in his NOHARM study, an incidental lung nodule missing from the AI summary could leave cancer undiagnosed
- "Chart lore": mislabeled diagnoses get copy-pasted forward forever, and AI summaries confidently repeat them
More from AGI Musings
- François Fleuret: Can a cognitive entity exist with full control over its information feed? — francoisfleuret · 2026-09-27
- Musk: AI will exceed the intelligence of all humans combined by 2030 — XFreeze · 2026-09-27
- Debate: scaling today's 150 IQ AIs beats racing toward a 3000 IQ superintelligence — JOBhakdi · 2026-09-27
- Startup mantra: build something so good it gets regulated — next up, 'make something agents want' — seanwbren · 2026-09-27
- Researcher: within months, submitting proofs without AI verification will be malpractice — RexDouglass · 2026-09-27
- 'Stochastic Parrot' fight reignites as Gebru hits back at critics — mjdramstead · 2026-09-27