No more radiologists!

Geoffrey Hinton, the "Godfather of AI", told medical schools to stop training radiologists in 2016. Deep learning would outperform them within a decade, he said. It was obvious. A decade later, US radiologists still earn well north of half a million dollars a year, and the market still cannot fill the roles fast enough.

The technology worked, but the prediction failed anyway, because it confused interpreting a scan with practising radiology. The product was the licence to practise medicine, not the scan itself. The signature and acceptance of liability were part of that. The named professional accountable when something went wrong was the product all along.

Tech is now making the same mistake with code. Meta cut 8,000 people during a record quarter and moved 7,000 more onto AI teams. Amazon shed 16,000 corporate roles. Oracle and Microsoft are restructuring on the same basic premise: AI makes output cheaper, so fewer hands are needed.

Output is easy to measure, but judgement is not.

So companies optimise what they can count and quietly retire what they cannot. The result is debt, just not the kind that appears cleanly on an income statement. Capability debt, as the junior pipeline thins and nobody is left to train. Judgement debt, as the seniors who stop building lose the calibration that made them valuable in the first place.

Both compound over time. Neither shows up until something has to be secured, scaled, debugged, explained, or defended, and nobody can tell whether the machine got it right. What should worry smaller firms more than the giants? Meta can afford this mistake. It has the bench, the cash, and the ability to buy judgement back when judgement turns scarce.

A twelve-person company that lets AI replace one of its juniors has no bench. And it cannot outbid the market for the experience it chose not to grow.

Production was never the constraint in tech development. Discernment was.

Who in your business will still be able to tell good from bad in three years?