Stanford SIEPR: AI Hits New Grads, Not Aggregate Jobs Yet

The interesting move in Stanford SIEPR's new policy brief on AI and jobs is where it refuses to go. Neale Mahoney, Erika McEntarfer, and Karsen Wahal, writing for policymakers who have been told a jobs apocalypse is either happening or fake, land on a third answer: neither. AI is showing up in the labor market at the edges, particularly for people just entering it, and not much anywhere else yet.

The numbers do the work. Unemployment for new graduates hit 5.6 percent in early 2026, up 1.6 percentage points from three years earlier. The brief cites work by Brynjolfsson, Chandar, and Chen finding a notable decline in employment among early-career workers in AI-exposed occupations, notably software developers and customer service representatives, since ChatGPT's launch in 2022; the specific figure the authors put on the 22 to 25 age cohort in the most AI-exposed occupations is roughly a 13 percent relative decline. At the same time, unemployment in the most AI-exposed occupations overall is rising, but not faster than in occupations least exposed, and software developers as a broader category have still seen continued employment growth.

Why that split pattern matters is the harder question. Entry-level work is exactly where junior engineers, junior analysts, and first-line customer service reps live, and it is exactly the tier a well-prompted model is most plausibly closing in on. If the effect is real, a bad first job market has a long tail in the wage-scarring literature. McEntarfer, who served as Commissioner of the Bureau of Labor Statistics until August 2025, is careful to add that AI adoption in the workplace is still pretty low and concentrated in tech, professional services, and finance. Firms historically take years, even decades, to weave a new general-purpose technology into how they actually operate; the brief flags that in 1900 industrial electric motors provided 5 percent of installed mechanical power and needed 30 years to reach 80 percent adoption.

The honest caveat is that this is a synthesis, not a causal identification. It does not disentangle how much of the new-grad softness is AI versus post-pandemic hiring pullback, higher rates, or tech's own layoff cycle, and it does not tell you which specific tasks inside 'software developer' or 'customer service' the models are actually eating. Take the 13 percent as a directional signal, not a settled coefficient.

The forward read is that if you run early-career hiring, apprenticeship pipelines, or a CS program, the useful window is now, while the aggregate labor market has not yet forced anyone's hand. The employers and schools that rebuild the entry-level rung around AI-augmented work, rather than waiting for it to shrink further, are the ones who will still have juniors in five years.