The Predicted End of Radiology Was Supposed to Be Now

In 2016, Geoffrey Hinton, who later won a share of the Nobel Prize for his foundational work on neural networks, told a Toronto audience that hospitals should stop training radiologists. "It's just completely obvious that within five years deep learning is going to do better than radiologists," he said, comparing the profession to "the coyote that's already over the edge of the cliff but hasn't yet looked down." Ten years on, the cliff has not materialised. Radiologists today earn more, work more, and are in shorter supply than they were when Hinton made the prediction.

The numbers from professional bodies and recruitment trackers are unambiguous. Average radiologist compensation in the United States reached $571,000 in 2025, up roughly 9 percent year over year. The active radiology workforce has grown about 10 percent over the past decade. In March, more than 4,300 radiology positions were actively listed in the US, with an average vacancy lasting 130 days. None of this looks like a profession being phased out.

Why the prediction missed

The technical capability arrived, in narrow form. FDA-cleared AI tools now triage chest X-rays, flag suspected stroke and intracranial haemorrhage, and segment lung nodules. These tools work, and hospitals use them. What did not happen is the substitution Hinton imagined, for several intersecting reasons.

First, regulation. Medicare and Medicaid only reimburse imaging studies when a licensed physician signs the final report, and most private insurers follow the same rule. An AI model that gets the right answer 95 percent of the time is genuinely useful, but it cannot be the final signer.

Second, the actual job. Reading scans is one part of a radiologist's week. The rest is consulting with referring physicians, supervising and performing image-guided procedures, monitoring patients during contrast studies, and adjudicating cases where the imaging is ambiguous. The AI tools that work well at pattern-spotting do not do any of that.

Third, induced demand. Imaging is now cheaper and faster per study than it was a decade ago, partly because of the very AI tools that were supposed to displace radiologists. So hospitals order more of it. Case volumes across US radiology departments rose roughly 25 percent between 2018 and early 2025, and someone still has to sign each report.

Hinton has revised

Hinton himself acknowledged the miscall in 2025, telling interviewers that he had underestimated how much of the job sat outside the imaging-read itself and how slowly clinical adoption tends to move even when the technology is ready. He still expects AI to reshape the profession; he just no longer expects it to erase the profession.

The radiology story is worth holding onto because the structure of the prediction (a clean substitution of model for human, on a short clock) is the same structure being applied right now to junior lawyers, accountants, software engineers and customer service agents. It might come true for some of those jobs. It might miss for the same reasons it missed for radiology: regulation, scope creep, induced demand, and the awkward fact that doing the work is rarely the same as doing the central technical task. We will know in ten years. We are also not seeing the broader job apocalypse show up in the labour data either.