Will AI take my job? The measured effect is real, narrow, and concentrated on people under 26

Most answers to this question are forecasts. The most useful evidence is not: it comes from payroll records of people who already have jobs, or recently did not get one.

The Stanford Digital Economy Lab, working with data from ADP — the largest payroll processor in the United States, covering millions of workers across tens of thousands of firms — published an analysis titled *Canaries in the Coal Mine?* examining employment changes since generative AI came into wide use. The headline finding is specific rather than general.

**Workers aged 22 to 25 in the most AI-exposed occupations have seen a roughly 16 per cent relative decline in employment.** Not the whole workforce. Not every industry. Early-career workers, in the occupations most exposed, measured against comparable workers who are not.

The second finding matters as much and travels less. **Employment is growing in occupations where AI is used to augment workers rather than automate their tasks.** The same technology, applied differently, produces the opposite sign. That distinction — automate versus augment — predicts the direction better than industry does.

So the honest answer to the question in the headline has three parts. If you are early in your career in an exposed occupation, the effect is real and already measurable. If you are mid-career, the evidence for displacement is much weaker. And if your work is being augmented rather than substituted, the measured direction is currently upward.

There is a mechanism worth understanding underneath the numbers. What the data mostly shows is **suppressed hiring rather than mass firing** — positions that would have been opened for junior staff not being opened. That is a quieter process than a layoff round and it produces no announcement, which is part of why it took payroll data rather than headlines to detect.

Now the caution, because this is a field where numbers are quoted loosely. Reported figures for AI job displacement vary enormously, and the main reason is that studies measure different objects. Some count tasks that could theoretically be automated. Some count hours. Some count actual headcount. A study finding that 40 per cent of tasks are exposed is not saying 40 per cent of jobs will go, and the two get conflated constantly. The Stanford figure above counts employment, which is why it is the one used here.

Projections should be read separately from measurements. The World Economic Forum's Future of Jobs Report projects 92 million roles displaced globally by 2030 against 170 million created — a net gain of 78 million. That is a projection about the end of this decade, produced by a model, and it belongs in a different category from a payroll record of what already happened. Both are informative. Only one is evidence.

What none of this establishes is whether the early-career effect spreads, stays contained, or reverses as firms work out where the technology actually helps. The measurement is a few years old at most. It describes a beginning, and the shape of the rest is not in the data yet.