AI and Jobs, Signal vs Noise

Narrative

Measured outcome data (a) point to a narrow, age-specific effect rather than aggregate displacement. The Stanford Digital Economy Lab's ADP payroll study (Brynjolfsson, Chandar, Chen) first reported a 13% relative employment decline for ages 22-25 in the most AI-exposed occupations. Its August 2026 revision puts the gap at about 19%, with experienced workers showing no comparable gap. The authors also state that they see no widespread, economy-wide displacement, that declines concentrate where AI substitutes for tasks, and that some differential trends pre-date generative AI. The Yale Budget Lab's occupational-mix and synthetic difference-in-differences work finds no clear economy-wide AI footprint. Humlum and Vestergaard's Danish administrative-data study rules out earnings and hours effects larger than 2%.

The causal reading of the entry-level result is contested. A 2026 CNS working paper on 368 million Lightcast postings finds a negative effect on posting volume in exposed occupations, but the junior-senior divergence opens in 2021-22 and is absorbed by remote-work exposure. The Stanford SIEPR policy brief also cites two papers dating the hiring decline in AI-exposed occupations to the Fed's policy shift, before ChatGPT. Hosseini and Lichtinger's adopter-versus-non-adopter design finds fewer junior hires at GenAI-adopting firms, but it is an unreviewed working paper and cannot rule out selection. Johnston and Makridis find employment rising 4% where AI requires collaboration and no significant effect where it acts independently. On the offset side, Tom Harris's October 2026 SSRN paper finds new-task duties in AI-intensive occupations rising from under 1 in 100 postings in 2022 to about 3.5 in 2026.

Company attribution (b) is the weakest category. Challenger, Gray & Christmas counted 120,136 AI-cited cuts through September 2026, about 21% of 573,195 announced cuts, against 54,836 (about 4.5%) for all of 2025. Total announced cuts were 39% lower than a year earlier, the AI share peaked at roughly 40% in May, and AI ranked only fifth among stated reasons in September. Challenger records employers' explanations, not verified causes. A Duke and Federal Reserve Banks CFO survey reported negligible 2025 AI employment impact, and Census data show only 5% of firms reporting any employment effect. Treat Microsoft, Amazon, Salesforce, IBM and Klarna attributions as unverified unless role-level evidence exists. Lab usage data are a separate signal. Anthropic's Economic Index shows augmentation (52%) ahead of automation (45%) on Claude.ai, but API traffic is heavily automation-oriented (77% against 12% augmentation in the September 2025 sample).

Forecasts (c) rest on capability extrapolation and are not outcome evidence. METR's time-horizon paper finds a roughly seven-month doubling in the length of tasks frontier agents complete at 50% reliability, with a 110-minute horizon for o3. If the trend holds, it implies month-long software tasks between 2027 and 2031. METR's randomised trial found a 19% slowdown for experienced developers in early 2025, but METR now says that result is outdated and that speedups seem likely, so the productivity baseline is moving. The measurable leading indicators are payroll and posting data by age and exposure, the automation-versus-augmentation split in usage data, and whether the Canaries gap spreads beyond ages 22-25.