The AI Jobs Apocalypse Isn't in the Data — Yet

A Stanford SIEPR synthesis of the current research pushes back on the dramatic predictions — including Anthropic CEO Dario Amodei’s claim that AI could eliminate half of white-collar jobs and drive unemployment to 20 percent. So far the aggregate numbers show no such thing. Unemployment among the most AI-exposed workers has risen only 0.77 points since 2022, slightly less than the 0.85-point rise among the least-exposed group, pointing to a broadly softening labor market rather than AI-specific destruction. Job postings for highly exposed roles like software development have actually grown faster than average, and firms that adopted enterprise AI saw employment climb 10 percent over the following two years. Layoffs blamed on AI often trace instead to freeing up cash for AI investment or unwinding pandemic over-hiring.

The more credible warning sign is at the entry level. New-graduate unemployment hit 5.6 percent in early 2026, up 1.6 points in three years, and work by Brynjolfsson and colleagues finds employment declines for young workers in exposed occupations like software development and customer service while older workers in the same fields held steady — the ‘canaries in the coal mine.’ But the timing muddies the story: hiring for these roles started falling after the Fed’s March 2022 rate hikes and before ChatGPT launched, and the post-pandemic shift to remote work also tilted hiring toward experienced staff. With added controls, the entry-level decline doesn’t become clear until 2024, when AI capabilities and adoption had genuinely advanced.

On productivity, the evidence is mixed but leans positive, and the gains are unevenly distributed. A call-center study found a generative AI assistant lifted overall output 15 percent, with novices resolving 30 percent more issues per hour while top performers saw no improvement — suggesting AI compresses skill gaps more than it replaces workers. The brief’s honest bottom line: firm adoption is accelerating unevenly, macroeconomic shocks make AI’s effect hard to isolate, and this early evidence is far from the final verdict on how AI reshapes work.