Atlantic: three questions to judge AI's job-displacement risk

Ten years after Geoffrey Hinton told an audience in 2016 that 'people should stop training radiologists now,' the number of U.S. radiologists is up roughly 17 percent and average pay has climbed from about $350,000 to around $570,000. Imaging case volume rose about 25 percent between 2018 and early 2025.

That is the opening Rogé Karma uses in The Atlantic to argue against reasoning straight from an automatable task to the end of the job containing it. Dario Amodei's claim that AI would 'wipe out half of all entry-level white-collar jobs' is the kind of forecast Karma thinks needs sharper tests before anyone acts on it.

He offers three. The first asks whether the job is a strong bundle or a weak bundle. A trial lawyer's automatable work is tangled with messy courtroom judgment that is not; a recruiter's résumé-sorting can be lifted out cleanly, leaving sourcing, interviewing, and closing behind.

The second test is the Jevons paradox: when the automatable part gets cheap, demand for the whole service can rise rather than fall. ATMs were supposed to end bank tellers. Teller headcount instead grew from roughly 500,000 to 600,000 as cheaper branches multiplied. Rising imaging volumes have played a similar role in radiology.

The third question asks whether a technology enhances expertise or commodifies it. Accounting clerks consolidated into fewer, better-paid roles once software did the arithmetic; inventory clerks were deskilled and multiplied at lower wages. Radiology, in Karma's reading, sits with the accountants.

'The real lesson from radiology is not that Hinton's forecast was wrong about the technology,' Karma writes. The market, in his telling, buys a bundled service, not a single task. Hinton himself now calls the 2016 prediction early rather than wrong.