Real payroll records for millions of American workers now show what surveys and CEO opinion polls could only guess at: entry-level jobs in AI-exposed fields are falling behind, and the gap keeps widening.
Stanford economists Erik Brynjolfsson, Bharat Chandar and Ruyu Chen have an answer to the question every founder keeps asking about AI and hiring. It isn't comforting. Not if you're 23 and looking for work. Their study, "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence," uses ADP payroll data across more than 730 occupations to track what's happened to employment since ChatGPT launched in late 2022. In a revision published on August 12, 2026, the Stanford Digital Economy Lab said employment among workers aged 22 to 25 in highly AI-exposed jobs now sits about 19% below where it would be if it had kept pace with less-exposed peers. That's up from 15% in the July 2025 data vintage. The direction is plain.
This isn't survey data. It isn't a CEO checking a box on what he expects AI to do next year. The Canaries Dashboard, built by Stanford Digital Economy Lab and ADP Research, said its launch sample covered 25,000 firms and 4.6 million workers in November 2022 who were matched to occupation codes. That distinction matters. An NBER working paper published in February 2026 surveyed nearly 6,000 CEOs, CFOs and senior executives across the US, UK, Germany and Australia and found that 89% reported no productivity impact from AI use over the prior three years. That paper measured belief and company self-reporting. This one measures paychecks.
The hiring cut is the story #
Look at the mechanism, because it's the part every founder weighing an AI hiring decision should sit with. Brynjolfsson's team finds the decline runs primarily through reduced hiring, not increased separations. Companies aren't cutting junior software engineers loose in waves. They're simply not adding the junior roles in the first place, routing the summarizing, formatting and boilerplate coding work that used to train new hires straight into a model instead. That's the shift. Between November 2022 and June 2026, employment for 22 to 25 year olds in the two most AI-exposed occupation quintiles fell about 11%, while the same age group in the three least-exposed quintiles grew about 10%, according to the Stanford Digital Economy Lab.
The age gradient inside AI-exposed jobs is just as telling. The Stanford dashboard says early-career software developers and customer service representatives show large declines, while the pattern fades for older workers. Experience is protecting the paycheck right now, not just talent or effort. A junior engineer with two years on the job is competing against a senior one who already knows what to ask the model and which answer to throw away. No six months of hand-holding required.
Computer science enrollment falls for the first time in 20 years as AI reshapes who gets hired
Computer science enrollment fell for the first time in roughly 20 years in 2025-2026, with CS and programming majors each down more than 10%. Stanford research shows employment for developers aged 22-25 has dropped nearly 20% since ChatGPT launched. The entry-level talent pipeline is thinning just as AI-native companies prepare to scale. - computer science enrollment declining 2026 - junior developer job market harder
There is a caution here, and you shouldn't skip it. Stanford says the patterns are descriptive, not causal estimates, and the authors note that the gaps shrink under some education controls and are larger in the ADP sample than in national survey benchmarks. Fair enough. But the same update also says the divergence has widened through mid-2026, remains when technology firms and computer occupations are excluded, and shows up most clearly where AI is used to automate tasks rather than help workers do them. That's not proof of everything. It is enough to stop pretending nothing is happening.
"Whatever it is," Brynjolfsson told Fortune in June, "it's not going away." The newest Stanford revision makes that line harder to brush off. The gap has widened at each major data vintage the lab has published since it first surfaced the pattern in August 2025.
The founder's uncomfortable choice #
For founders, this isn't an abstract labor economics debate. It's the calculation behind every headcount plan written this quarter. If a model can now handle the summarizing and first-draft coding that used to fill a junior hire's first year on a codebase, the honest question isn't whether to use AI. That question's already answered. It's whether your startup still needs the entry-level rung on the ladder at all, and if you cut it, who trains your senior engineers five years from now? Here's the thing. New jobs, if they arrive, aren't obligated to be entry-level jobs. Nobody's promised that. A labor market that stops training 23 year olds today is one short on experienced 33 year olds a decade from now. For a startup trying to extend runway, skipping junior hiring can look rational on a spreadsheet. For the economy, it leaves a hole where apprenticeship used to be.
Also read: Sam Altman Fears a Handful of Companies Will Control AI, SEC Subpoenas Wall Street Banks Over AI Hedge Fund Situational Awareness, Chinese State Hackers Doubled Their Attack Volume Using DeepSeek AI