The headline numbers on AI and jobs look like good news. They are incomplete.
New research from labor economist Andrew Johnston finds that industries most exposed to AI — measured by how much of their work could theoretically be automated — posted 10% higher productivity, 3.9% faster job growth, and 4.8% higher wages in 2024 compared to less-exposed sectors. The finding cuts against a popular narrative: that AI destroys jobs. Johnston's data covers nearly all US employers from 2017 through 2024, a large enough sample to matter.
But the aggregate masks a split that the headline number doesn't show. The gains are not evenly distributed across workers.
The data reveals two distinct deployment modes with opposite labor effects. In sectors where AI primarily complements human workers — handling routine tasks while humans make judgment calls — employment rose roughly 3.6% per standard deviation increase in AI exposure. In sectors where AI executes tasks more autonomously, with less human involvement, there was no significant employment change, but workers in those roles experienced slower wage growth. Johnston's team used administrative records, not surveys, which makes the numbers harder to dismiss.
The broader context reinforces the stakes. By late 2025, roughly 26% of workers across all occupations reported using AI frequently — up from about 12% in mid-2024, according to The Conversation. The dose-response relationship is measurable: each additional percentage point of frequent AI use in a state and industry is associated with about 0.1% to 0.2% higher real output and 0.2% to 0.4% higher employment. More AI, more output, more jobs — on average.
But the distribution story diverges from the average. Workers aged 22 to 25 in highly exposed occupations experienced employment declines of roughly 16% relative to trend following ChatGPT's release, according to the Law and Economics Center at George Mason University. Senior employment remained stable. The gap is not small, and it is not explained by remote work normalization or cohort size effects alone.
The finding requires scrutiny. The Yale Budget Lab argues that occupational mix changes since ChatGPT are only about one percentage point above the internet-era baseline and predate the generative AI wave in some sectors — meaning the dose-response gains may reflect ongoing automation trends rather than something fundamentally new about this technology cycle. If that reading holds, the Makridis-Johnston numbers describe a continuation of historical patterns, not an AI-specific transformation.
For founders and product managers, the practical question is not whether to adopt AI — the data suggests modest adoption correlates with modest gains — but how. The research points to a specific deployment distinction: complementary deployment (AI handles routine tasks, humans handle judgment calls) correlates with job growth. Autonomous execution (AI replaces the human task entirely) correlates with slower wage growth in those roles, at least through 2024. The difference is not academic. It is a product design choice with workforce consequences.
What to watch: whether the employment decline for younger workers stabilizes, reverses, or accelerates as AI capability continues expanding — Anthropic's own research suggests current AI covers roughly 33% of Computer & Math tasks that could theoretically be automated, leaving the rest still in human hands. That ratio is not fixed. If it moves toward 50% or higher, the generational inequality angle becomes the dominant story, not a footnote to an otherwise optimistic headline.
