AI & Labor Markets

Current landscape

The strongest empirical evidence available is a Stanford Digital Economy Lab study (Brynjolfsson, Chandar, and Chen, "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence," Nov. 13, 2025) using ADP administrative payroll microdata — not a survey, not a model, actual employment records — with an event-study design. Its headline finding: workers aged 22–25 in the most AI-exposed occupations saw a 16% relative decline in employment since late 2022, concentrated specifically where AI appears to automate tasks rather than augment a worker doing them. An earlier August 2025 draft of the same ongoing work cited roughly 13% for the same cohort. The number moved with revision, which is a normal feature of live empirical research, not a reason to distrust the finding, but it is a reason not to over-rely on the precise magnitude.

The same paper's companion finding is, if anything, more informative: workers 30 and older in the same high-exposure occupations saw employment grow 6–12% over the identical period. That's specifically inconsistent with the "this is just a soft macro economy, not AI" framing that's common in public debate. A generic economic headwind would be expected to hit both age cohorts in the same occupations similarly, not diverge this sharply by age within the same job category. A live "Canaries Dashboard" extending the series through April 2026 shows the entry-level effect still growing; the specific pace often quoted in press coverage ("roughly half a percentage point per month") comes from an interview with the authors, not a table in the paper itself, and is treated here as somewhat softer-sourced than the core statistics.

The strongest counterweight to alarm comes from Daron Acemoglu's (MIT) calibrated macroeconomic model, which implies AI is likely to raise US total factor productivity by no more than roughly 0.71% cumulatively over ten years. That's explicitly positioned by its own author as a conservative counterweight to much larger, more optimistic industry productivity forecasts. Acemoglu's model also suggests AI is unlikely to worsen inequality through the channels prior waves of automation used, but does predict a widening gap between returns to capital and returns to labor. On the other side of the same debate, David Autor (also MIT) argues AI's better framing is as a collaboration tool that extends expert judgment to more workers — potentially rebuilding rather than hollowing out middle-class work. That's a conditional, hopeful argument about how AI could be deployed, not a measurement of how it currently is being deployed, and it sits in direct tension with the Stanford paper's observed, automation-leaning entry-level effect.

One additional, widely circulated claim — that current AI could already perform tasks tied to roughly 11.7% of the US labor force, about 151 million workers and $1.2 trillion in pay — is deliberately not treated as fact in this issue. It traces to MIT-affiliated reporting but was not independently confirmed via a primary source during this project's research pass, and per this platform's own verification-status discipline (see issue-content-model.md), an unverified headline number doesn't get cited as settled evidence just because it's widely repeated. A single external methodological critique of the Stanford paper (Casilli) argues its causal framing may not hold up; no peer-reviewed rebuttal has been found, so this is noted as a live, contested question, not a refutation.

Trade-offs and opposing views

Acting on a still-developing empirical signal, versus waiting for economists to reach consensus (which could take years, during which displaced workers get no support), is a trade-off. The party's answer is that targeted, portable transition support is a reasonable hedge regardless of which side of the academic debate eventually wins: it helps people if the effect broadens, and it doesn't cost much if it stays concentrated in the population already showing up in the data, because it's targeted rather than universal.

Two objections deserve direct answers. Labor advocates who see the Stanford data as reason for much stronger intervention (hiring bans, mandated quotas, dedicated compensation funds) get a direct response: the observed data is specific — entry-level workers in highly exposed occupations — not (yet) a broad-based labor-market effect, so a blanket intervention is disproportionate to what's been measured so far. The party commits to revisiting this if the ongoing monitoring in Proposal 2 shows the effect broadening. Those who dismiss displacement concerns as overblown corporate excuse-making get the direct empirical answer above: the age-cohort divergence within the same occupations is specifically hard to explain as generic macro softness, which is exactly the kind of evidence a "just the economy" dismissal would need to explain away and doesn't.