Research question

If AI is going to displace work, who is it displacing first — and does the evidence support the sweeping occupation-wide predictions, or something narrower and stranger?

Abstract

Serious estimates of how many jobs are at risk from AI and automation range from 9% to 80% — a gap large enough that the disagreement is itself the story. The lower figures come from task-based analyses that treat most jobs as partially, not wholly, automatable. The higher ones come from long-horizon forecasts stretching to 2050. Meanwhile, the actual labour-market data collected so far tells a much narrower story than either extreme.

The Dallas Fed, Stanford, and the International Labour Organization all report little evidence of broad, economy-wide job losses attributable to AI as of 2025–2026. But underneath that reassuring aggregate sits one sharp, repeated, cross-country signal: workers aged 22 to 25 in AI-exposed roles are losing ground, while older workers in the identical roles are gaining. Occupation may turn out to be the wrong axis entirely. Experience might be the one that matters.

Key findings
  • Forecast methodology drives most of the disagreement: ZEW Mannheim estimates 9% of jobs at high risk; the OECD's task-based approach — widely regarded as the most accurate method to date — finds 14% at high risk and 32% facing significant change; long-horizon 2050 forecasts run as high as 60–80%.
  • McKinsey's own position has shifted. Its 2017 report treated "technically automatable" as equivalent to "will be automated." Its 2025 report explicitly separates technical potential (57% of work hours) from actual deployment, noting 42% of organizations have already abandoned AI pilot projects.
  • The Dallas Fed (June 2025) found very little evidence of AI displacing jobs at scale to date, and low correlation between an occupation's AI exposure and its projected 10-year job growth or decline.
  • The International Labour Organization, citing research across seven countries, reported in May 2025 that there is "little evidence so far of large-scale job losses directly attributable to GenAI."
  • Beneath that aggregate calm: US workers aged 22–25 in the most AI-exposed occupations saw employment fall by roughly 6–12% since late 2022, while workers in the same occupations aged 30 and older saw employment rise 6–9% over the same period, per Stanford and Dallas Fed analyses.
  • The pattern is not US-specific. Ireland's Department of Finance found youth employment in AI-exposed sectors fell 20% between 2023 and 2025, while employment for prime-age workers (30–59) in the same sectors grew 12%.
Why forecasts disagree by a factor of nine

Most of the spread between 9% and 80% comes down to what's being counted. Whole-job forecasts ask whether an entire occupation could theoretically be automated, which tends to produce large, dramatic numbers. Task-based forecasts, like the OECD's, instead ask what share of a job's individual tasks are automatable — and since almost no real job is 100% one type of task, this method consistently produces smaller, more conservative estimates.

The OECD's approach has held up best against real-world evidence so far, which is itself informative: the more granular the method, the smaller and more defensible the number tends to be. McKinsey's own retreat from its 2017 framing — separating "could be automated" from "is being automated" — is a rare case of a major forecaster publicly correcting its own methodology in light of slower real-world deployment.

The one signal already visible

If occupation-wide displacement isn't showing up yet in aggregate data, something else is. J. Scott Davis's Dallas Fed analysis found that since ChatGPT's late-2022 debut, overall US employment rose about 2.5%, while employment in the most AI-exposed industries slipped roughly 1% — a modest gap, not a collapse. Wages in those same AI-exposed industries actually grew faster than the economy overall, 8.5% versus 7.5%, which is the opposite of what a straightforward displacement story would predict.

Davis's explanation is a distinction between codified knowledge — the kind learned from textbooks, which AI reproduces easily — and tacit knowledge gained through hands-on experience, which it doesn't. "Returns on job experience are increasing in AI-exposed occupations," as Davis put it; young workers with mostly codified knowledge and little experience are facing a genuinely harder hiring market, while experienced workers in the same fields are becoming more valuable, not less.

A separate Stanford analysis of ADP payroll data across millions of records reinforces the pattern with a sharper mechanism: when AI automates a task outright — writing code, handling routine customer chats — entry-level hiring falls. When AI merely augments a task, supporting rather than replacing judgment, employment holds steady or grows. Automation and augmentation are producing opposite effects, and young workers are disproportionately exposed to the automation side.

The honest gap in the evidence

The clean version of this finding — "AI is quietly hollowing out entry-level jobs" — is more solid than the sweeping occupation-wide forecasts, but it isn't fully settled either. Stanford's own research flags that hiring in AI-exposed occupations began declining after the Federal Reserve's 2022 interest-rate hikes but before ChatGPT's public release, meaning some of this effect may be monetary policy and pandemic-era over-hiring correcting itself, tangled together with any genuine AI effect.

Nobody yet knows whether this is temporary — young workers catching up once they accumulate experience — or the leading edge of a lasting structural shift in how career ladders work. And the data covers barely three years since ChatGPT's release, against forecasts that stretch to 2050; that mismatch in time horizon is itself a reason for caution in either direction.

Why this matters for the automation question

The Labour Machine's automation control treats exposure as a single economy-wide dial. The evidence in this paper suggests that's a simplification worth naming plainly: the real, currently measurable effect isn't evenly distributed across an occupation or an economy — it's concentrated at the start of a career. A model built to explore "what happens as automation rises" should be read with that in mind, particularly for any future work on generational or age-based effects the current version doesn't yet capture.

Methodology

This paper synthesizes published research and data analysis from 2025–2026 rather than presenting original data collection. Primary sources include the Federal Reserve Bank of Dallas's research division, Stanford's Institute for Economic Policy Research, the OECD, the International Labour Organization, and Ireland's Department of Finance, cross-referenced against longer-horizon industry forecasts from McKinsey Global Institute and others for methodological contrast.

Assumptions
  • US-focused data (Dallas Fed, Stanford) is treated as broadly informative for other advanced economies, supported by corroborating findings from Ireland; this is not assumed to generalize to economies with substantially different labour-market structures.
  • The age-based employment gap identified in 2025–2026 data is treated as a real, currently observable pattern — not as confirmation of a permanent structural shift, given the short time horizon available.
Limitations
  • Under three years of post-ChatGPT data exist; this paper cannot distinguish a lasting structural change from a temporary hiring-market adjustment.
  • Some of the youth employment decline may be attributable to 2022's interest-rate increases and pandemic-era hiring corrections rather than AI specifically, per Stanford's own research.
  • Long-horizon forecasts (2050) are inherently speculative and not directly comparable to the near-term empirical findings this paper otherwise relies on.
Data sources
Citation

After Labour (2026). Who Goes First? Mapping which jobs are most at risk from AI and automation.