The Best Evidence on AI and Jobs Has a Blind Spot: It Ends Before Agents Got Good

⬅️ Back to Articles

If you want one serious answer to “is AI eating jobs yet,” read the July 2026 SIEPR policy brief by Neale Mahoney, Erika McEntarfer, and Karsen Wahal. It is the most careful synthesis I have seen. My one reservation: its evidence window largely closes before LLM agents and agent workflows got genuinely good after February 2026. Treat it as the before photo.

  1. The apocalypse is not in the data. Unemployment among the most AI-exposed quintile of workers rose 0.77 percentage points since 2022; the least-exposed quintile rose slightly more, 0.85 points. A broadly softening market, not an AI-driven one. Enterprise-AI adopters grew employment 10% over two years, and software developer postings grew faster than other occupations this year.

  2. The canaries are real though. Recent graduates face 5.6% unemployment, up 1.6 points in three years. Brynjolfsson, Chandar, and Chen found early-career employment declining in exposed occupations like software development and customer service while older workers held steady.

  3. But the timing refuses to cooperate. Hiring in exposed occupations began sliding around 2022, months before ChatGPT launched. The Fed started hiking that March, and remote work erodes on-the-job learning for juniors. With better controls, entry-level declines only become notable in 2024. Attribution here is genuinely hard, and they admit it.

  4. Productivity gains are real but jagged. A call-center assistant lifted productivity 15%, concentrated at +30% among novices with nothing for experts. GitHub Copilot sped coding tasks 56%. Yet in Kenya, less-skilled entrepreneurs who acted on generic AI advice posted lower profits than peers using nothing. Extracting value from AI is itself a skill.

  5. The quiet cost is sameness. AI-assisted stories converge toward each other; scientists who adopt AI publish more papers on fewer topics with less collaboration. Even a rising tide flattens things.

  6. Adoption is fast but shallow. The trustworthy floor: about 20% of US firms use AI per the Census Bureau. Concentrated in tech and finance, usually one or two functions deep, most firms still piloting. Only 5% of adopting firms report any employment impact at all, and a Danish study finds tasks restructuring without headcount moves.

  7. Why you should not relax anyway. Solow joked in 1987 that you could see computers everywhere but the productivity statistics; the returns showed up in the late 1990s after firms reorganized around PCs. Measured invisibility is what general-purpose technologies look like mid-adoption.

  8. Now my addition to their caveat. The brief’s series mostly end in early 2026, which is precisely when agentic workflows stopped being demos and started finishing multi-step work. If agents compound the way the last year suggests, the 2027 update to this brief will be measuring a different regime. Read this one for its methods and baselines; hold the conclusions loosely.

The takeaway: If you are early-career in an exposed occupation, the nuance beats either headline: the aggregate market is calm but junior demand is soft, so stack judgment-heavy skills now. Then re-check this research series quarterly; the interesting results arrive in the next edition.

Related TMFNK Content

Crepi il lupo! 🐺