What was said
In 2016, Geoffrey Hinton told an audience in Toronto that people should stop training radiologists now — within five years, deep learning would read scans better than humans; "it's just completely obvious." The line was quoted for a decade as the canonical death sentence for an expert profession. The argument held up, too: image recognition was exactly where deep learning matured first, and benchmark results were genuinely closing on, then passing, human performance.
What actually happened
Ten years on, radiology has not disappeared. Radiologists remain in shortage in the US, residency spots are as contested as ever, and imaging departments are busier, not quieter — scan volumes grew faster than the workforce. AI reading tools did enter the clinic, but in the second chair: triage, flagging suspicious regions, drafting reports. The signature on the report is still a physician's.
Why it didn't come true
Three things were left out of the calculation. First, a job is a bundle of tasks: reading scans is only part of radiology — there is also protocol design, clinical consultation, procedures, and answering for the conclusion. Second, liability cannot be outsourced: a missed diagnosis sues the signer, not the model, and neither hospitals nor regulators will accept "the model said so." Third, demand grows: as reading got faster and cheaper, physicians ordered more imaging — the efficiency gain was eaten by volume, Jevons' paradox in a white coat.
The unexpected part
Hinton himself later became the most famous warner about AI risk. In 2016 he worried about radiologists' jobs; after 2023 he worried about something else entirely — two predictions from the same mouth, in nearly opposite directions. The drift of the predictor belongs in the ledger too.