Erik Brynjolfsson has been saying the same thing for at least four years. The Stanford economist, who directs the university's Digital Economy Lab and co-founded the AI workforce analytics firm Workhelix, told Brookings in 2022 that the next ten years could be the best in human history, or one of the worst. He told the Financial Times the same thing in 2024. In a podcast interview this past July, he said it again, in more detail than before.
Brynjolfsson has spent three decades studying what new technology actually does to jobs, dating back to a 1991 doctoral thesis on information technology and the reorganization of work, well before predicting AI outcomes became a crowded field.
What has changed since 2022 is not the claim. It is how much evidence now sits behind each half of it.
Both halves of the prediction are still standing. Only one of them has picked up hard data to back it.
What does the "worst" half actually look like in data?
On the optimistic side, Brynjolfsson pointed to unprecedented wealth creation and, citing a conversation with DeepMind's Demis Hassabis, the possibility that most diseases could be curable within a decade. On the pessimistic side, he named the same category of risk he has named for years: engineered pandemics, AI-driven manipulation of public discourse, and a dangerous concentration of power.
That data comes from Brynjolfsson's own research group. Along with Stanford colleagues Bharat Chandar and Ruyu Chen, he has spent the past year tracking a specific, narrow claim using payroll records from ADP covering millions of American workers. The paper, called Canaries in the Coal Mine, first reported in August 2025 that workers aged 22 to 25 in the most AI-exposed occupations had fallen 13 percent behind where employment trends would otherwise predict. A November 2025 revision put the gap at 16 percent. The most recent update, published in August 2026, puts it at 19 percent, and rising.
For software developers specifically, one of the occupations the paper tracks most closely, headcount among 22-to-25-year-olds has fallen roughly 20 percent since late 2022, even as headcount for developers in their thirties and forties has kept climbing over the same stretch.
The paper is explicit about what this is not. There is no evidence, the authors write, of widespread, economy-wide job displacement. Overall employment keeps growing. The gap is narrow and specific: young workers, in occupations where AI substitutes for tasks rather than assisting with them, are simply not getting hired at the rate the data would otherwise predict. Layoffs are not driving the number. Reduced hiring is.
The authors have also tested whether something other than AI could explain it. They checked the pattern against interest rate exposure, since hiring in some industries is sensitive to borrowing costs, and found AI-exposed occupations are if anything less rate-sensitive than average, not more. They checked whether the decline predates ChatGPT's release, since a pre-existing trend would point to some other cause, and found the gap only becomes statistically significant starting in 2024, after generative AI tools were already in widespread use.
Why is it young workers specifically?
The authors offer a specific explanation for why it is young workers and not everyone. Entry-level roles tend to rely on what the researchers call codified knowledge, the kind of structured, learnable skill that shows up in training data and that a language model can reproduce reasonably well. Senior roles lean more on tacit knowledge, judgment built through years of experience that is harder to write down and harder for a model to imitate. The 22-to-25-year-olds losing ground are, roughly speaking, the workers whose entire value proposition an AI system can currently approximate.
Workhelix, the company Brynjolfsson co-founded in 2021, sells exactly this kind of occupational AI-exposure analysis to businesses, giving him a research agenda and a commercial one that point in the same direction.
We have written before about a related mechanism: institutions, whether governments or companies, invest in ordinary people because those people are useful to them, and that incentive weakens once the usefulness disappears. Canaries in the Coal Mine reads like an early, narrow, measurable instance of exactly that dynamic. Nobody had to decide to stop hiring young software developers and customer service reps as a matter of policy. The incentive to hire them simply got smaller, one role at a time, and the aggregate hiring data picked up the change before anyone had to say it out loud.
What does the "best" half rest on?
That is the part of Brynjolfsson's prediction that now has a number attached to it, one that has grown at every revision of the paper rather than shrinking. The optimistic half, the disease cures, the wealth creation, remains real and worth taking seriously, and Brynjolfsson has spent much of his career arguing that AI can augment human capability rather than simply replace it, pointing to occupations where AI use is already correlated with rising employment across every age group. But that half of the forecast is still mostly a projection resting on what capable people expect advanced AI to eventually do. The pessimistic half has started showing up in payroll records.
None of this settles which decade Brynjolfsson is actually describing. His own point, repeated across four years and several interviews, is that both outcomes are still genuinely live, and that which one dominates may depend less on the technology itself than on what gets built around it in the meantime. What has changed since 2022 is only that one half of his forecast has started leaving a paper trail, and the other one, for now, has not.