AI will not automate radiologists like you think — Jason Wei
Short summary
Jason Wei argues that radiologists will stay in demand despite AI, because cheaper scans will drive elastic demand, rare diseases will keep AI in a triage role for years, and healthcare's slow adoption and regulation will buy time. He uses Geoff Hinton's 2016 prediction that radiology training should stop as the foil.
- AI will handle common, easy-to-see conditions first but lag on rare diseases, since there is less data for them, so expect 3-5 years of triage use before full automation.
- Jevons paradox: if reading a scan becomes nearly free, radiologists might read only 20% of scans while five times as many arrive, and many developing countries with under one radiologist per million people would add demand.
- Cheap interpretation could push device makers toward cheaper, more frequent scanners, and medical regulation, legacy infrastructure, and data access limits will slow adoption.
Full summary
Jason Wei's core claim is that radiologists won't be automated away soon, and he opens by mocking Geoff Hinton's 2016 prediction that deep learning would beat radiologists within five years and that training radiologists should stop. Demand for radiologists is now higher than ever, which he takes as evidence the forecast missed. His argument rests on three pillars. First, AI is "jagged": it does well on common, easy-to-see conditions with lots of data but struggles with rare ones. Radiology is far more multi-task than people assume, with around a hundred conditions diagnosable from x-rays alone and thousands from CT and MRI. Wei expects AI to serve as a triage tool first, with full automation taking perhaps 3 to 5 years in the best case, held back by rare-disease data collection and the fact that frontier labs aren't focused on radiology.
Second, he leans on Jevons paradox. When something gets cheaper, people use more of it. He points to his own experience using Claude Code to build a custom 3,000-line HTML viewer for visualizing RL runs, which he says he would never have built before. Applied to radiology, if reading an MRI costs almost nothing and takes no time, physicians will order scans preemptively. A radiologist might read only 20% of incoming scans but face five times the volume. He also notes that many developing countries have fewer than one radiologist per million people, so demand there could surge. Longer term, cheap interpretation should push manufacturers to build cheaper, more frequent-use scanners, possibly even intraoperative MRI. He imagines people getting 10 to 100 times as many scans, mostly for prevention; a $100 full-body MRI, he says, would mean monthly scans instead of biennial ones.
Third, he argues that even a model better than today's top radiologists wouldn't end the work. Treatment protocols will change, new scanners will reveal new diseases, and some radiologists will shift into research. Future models might predict disease months earlier than current detection. And institutional friction matters: hospitals are slow to adopt technology, legacy infrastructure is hard to replace, and HIPAA, de-identification rules, and IRB requirements limit data access, especially in developing countries. Wei closes by noting that Hinton acknowledged being wrong in a recent podcast and now agrees with the elastic-demand argument. He suggests the same logic applies to any field where demand is elastic and quality scales with intelligence, programming included.