The Canaries Stopped Singing

For two years, every essay in this series has leaned on trendlines and testimony. This month, the microdata arrived.

Stanford's Digital Economy Lab — Erik Brynjolfsson's shop — published "Canaries in the Coal Mine?", analyzing payroll records from millions of workers via ADP. The finding: since generative AI's arrival, workers aged 22–25 in the most AI-exposed occupations — software development, customer service among them — have seen a roughly 13% relative decline in employment, while older workers in the same occupations and same firms held steady or grew. It's the pattern a capability-driven story predicts: the machine substitutes first for codified, book-learned skill (what new graduates sell) and complements experience-driven judgment (what veterans sell).

The same month, an MIT-affiliated report went viral for finding that ~95% of enterprise generative-AI pilots produce no measurable P&L impact — seized on by skeptics as proof the whole thing is hype. And GPT-5 launched to all of ChatGPT's hundreds of millions of users, to grumbling about tone and a bumpy rollout.

Hold all three honestly

Here's how these reconcile. Displacement is real and measurable — at the entry level, in exposed occupations, exactly where theory predicted first impact. Enterprise transformation is slower and harder than the vendor decks claim — integration, not intelligence, is the bottleneck, which is why the failed pilots cluster where firms sprinkled chatbots on unchanged workflows. And the frontier keeps advancing regardless of either fact. Anyone selling you a single-note story — "it's all hype" or "it's all over" — is skipping at least one of this month's datasets.

Actionable Steps

Policymakers: you now have peer-quality evidence of differential harm to young workers. That's the trigger condition for the transition programs I've argued for since 2024 — apprenticeship subsidies, hiring-data transparency, income floors piloted at scale. Executives: the 95% failure number is a checklist, not a comfort — the 5% that worked redesigned the workflow around the machine and kept humans on judgment and exceptions. Copy them. Young workers: the data says codified skill is the exposed asset. Accumulate the other kind — ship real things, own small outcomes, collect evidence of judgment.

I take no pleasure in the canaries. But the whole point of canaries is that you act when they go quiet.