The wrong question every teenager is asking about AI

In 2016, Geoffrey Hinton — one of the godfathers of modern AI — made a famous prediction. 'People should stop training radiologists now,' he said. 'It's just completely obvious that within five years, deep learning is going to do better than radiologists.' He was half right. Today, the FDA has approved more than 1,000 AI radiology tools. Some can detect injuries and diseases more accurately than human specialists. Yet the number of human radiologists has risen by 17 percent since 2016. They are in more demand than ever.

This is the puzzle at the heart of a brilliant new Atlantic article by Rogé Karma. The question on everyone's mind — 'Can AI do my job?' — turns out to be the wrong question entirely. The better question is: what kind of bundle is my job?

Karma's argument is simple and powerful. Every job is a bundle of tasks. Some are 'clean' — repetitive, predictable, easy to automate. Others are 'messy' — ambiguous, social, requiring judgment and context. A 'weak bundle' is one where the clean tasks can be peeled away from the messy ones without much trouble. A 'strong bundle' is one where the clean and messy tasks are tightly intertwined.

Radiology is a strong bundle. Yes, AI can read a scan faster than a human. But the radiologist's job was never just reading scans. It is talking to the referring doctor about what the finding means. It is calming an anxious patient. It is deciding whether a subtle shadow warrants a biopsy or just a follow-up scan. The clean task (image analysis) and the messy task (human judgment) are wrapped around each other so tightly that removing one destroys the other.

This matters enormously for how we think about our children's future. When a fourteen-year-old asks 'Will AI replace lawyers?' or 'Will AI replace doctors?' they are asking the wrong question. The right question is: will they understand how the clean tasks and messy tasks fit together? Will they know when to trust the AI and when to override it? Will they be the person who can explain the AI's output to a frightened client, a confused patient or a sceptical judge?

Karma offers three better questions to ask instead of 'Can AI do my job?' One: how much of my job is clean versus messy? Two: how tightly are those tasks bundled together? Three: if the clean tasks are automated, what new messy tasks emerge that only a human can handle?

For parents, this is a much more useful framework than the usual panic headlines. It means we do not need to guess which professions will survive. We need to teach our children to recognise the messy parts of any job — the parts that require empathy, negotiation, ethical judgment and creative problem-solving — and to become indispensable there.

It also means AI literacy is not about learning to code. It is about learning to manage the boundary between machine capability and human judgment. The child who grows up asking 'Is this the clean part or the messy part?' will have a professional instinct that no purely technical training can give them.

Hinton was wrong about radiologists. Not because AI failed — it succeeded brilliantly at the clean task. He was wrong because he underestimated how much of the job was messy. Let us make sure our children do not make the same mistake about their own future careers.