Does AI Cost Jobs, and Lead to Layoffs?

11.1  What the payroll data shows

The central study is Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, by Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen of the Stanford Digital Economy Lab. It uses high-frequency administrative payroll records from ADP — the largest US payroll processor, covering millions of American workers — running through June 2026, and takes the November 2022 ChatGPT release as its index date.1

Administrative payroll data matters here. This is not a survey of what executives say they intend, and not job postings. It is a record of who was actually on a payroll, month by month.

The six facts, stated precisely:

1. No economy-wide displacement. Across job categories and age groups, the researchers find no evidence of widespread AI-associated job displacement.
2. A sharp age effect. Employment of workers aged 22–25 in AI-exposed occupations stands roughly 19% below where it would be had it tracked their less-exposed peers. Experienced workers show no comparable gap.
3. The divergence is widening. It has grown steadily since August 2025.
4. Hiring, not firing. The gap operates primarily through reduced hiring rather than increased separations.
5. Substitution versus complementarity. Declines concentrate in occupations where AI substitutes for tasks; employment is flat or rising where AI complements the worker.
6. Employment, not pay. Adjustment is occurring through headcount rather than through base compensation.1

Fact 4 is the one most often lost in the retelling, and it changes the whole picture. This is not a wave of layoffs. It is a door closing quietly. Nobody is fired, no announcement is made, and the entry-level opening is simply not posted — which is precisely why it took administrative data to see it at all.

Aggregate employment in the ADP sample rose about 6% between November 2022 and June 2026, while employment in the most-exposed quintile grew about 4%.2 Both grew. Put that on screen next to the 19% figure, because both are true and quoting either alone misleads.

FIG-11-01

Two lines that separate

A time series from November 2022 to June 2026, indexed to 100 at the start. Two lines: workers aged 22–25 in AI-exposed occupations and the same age group in less-exposed occupations. The lines track together, then begin separating around August 2025, with the gap widening to the right edge. Mark the index date (Nov 2022) and the divergence point (Aug 2025). Add a third, faint line for experienced workers in exposed occupations that does not separate — that flat line is half the argument. Label the terminal gap ≈19%. Axis must start at 100, not at zero, and the caption must say so.

Aspect 16:9 · 3200×1800 px · Brand palette: navy #00040c, silver-blue #a8c0d8, column gold #a87848, caution #8f3a2a · Style: rigorous statistical chart with visible axis labels and source line

11.2  A number that keeps moving — and why that is the lesson

The 19% figure has a history, and teaching the history is more valuable than teaching the number.

Three readings of that table, all legitimate, and you should be able to hold all three:

• The effect is real and growing. Consistent with fact 3 — the divergence has widened steadily since August 2025.
• Any citation without a version is unreliable. All three numbers are “the Stanford study.” Someone quoting 13% is not wrong; they are quoting an earlier vintage.13
• Live research is not settled research. A figure that has moved three times in eighteen months may move again, in either direction.

11.3  Corroboration, and genuine dissent

Independent work pointing the same way

Hosseini and Lichtinger find that firms adopting generative AI reduced junior employment by about 9% relative to non-adopters, with senior employment unaffected.4 Different data, different method, same shape: the effect lands on the junior end.

The IMF has folded this evidence into its own analysis. Its January 2026 Staff Discussion Note Bridging Skill Gaps for the Future: New Jobs Creation in the AI Age reports that vacancies demanding AI skills post higher wages, but that the diffusion of those skills is linked to lower employment in occupations with high exposure and low complementarity to AI — and states explicitly that this poses challenges for young workers.5 Brynjolfsson has stress-tested the finding against the obvious confounders — interest rates, the technology sector specifically, remote work — and reports the pattern holding.6

The dissent, presented rather than dismissed

Economists reading the same data disagree about cause, durability, and appropriate response, with Daron Acemoglu the most prominent sceptic.6 The disagreement is not about whether the correlation exists. It is about attribution: 2022–2026 also contained a sharp interest-rate cycle, a post-pandemic hiring correction, and a technology- sector contraction with its own causes. Disentangling AI from those remains genuinely unresolved, and recent-graduate unemployment reaching 5.6% in early 2026 — up 1.6 percentage points in three years — is consistent with several stories.7

A methodological caution from a different direction is worth equal weight. Anthropic's own labour research introduces a measure it calls observed exposure, built from O*NET occupational task data, actual model-usage data, and task-level capability estimates — tracking deployment rather than theoretical possibility. It reports no detectable increase in unemployment for exposed workers since the ChatGPT release, alongside suggestive evidence of slowed hiring of younger workers in exposed occupations. And it cautions explicitly that employment and job-posting counts do not by themselves signal harm, since a decline in one occupation may be offset by openings in a related one — which is why the researchers prioritise unemployment as the outcome that most directly captures the potential for economic harm.8

Note who is making that argument: a company with an obvious commercial interest in a benign reading, publishing a methodological caution that happens to favour the benign reading. Both facts are true and neither cancels the other. The measure is a genuine contribution; read it with the interest in view. That is how one reads any interested party — including the vendors cited in Lessons 7 through 10.

11.4  What to do about it

Do not close on reassurance. Close on the actionable version.

The divide in fact 2 tracks the distinction Lesson 1 introduced as Polanyi's paradox. Codified knowledge — procedures that can be written down, and which juniors have historically been employed to execute — is what these systems replicate well. Tacit knowledge, accumulated through experience and difficult to articulate, is what they replicate badly. That is why the gap appears at 22–25 and not at 45.

So the defensible position is the one where judgement, relationships, physical presence, or accountability are the product. Note that these are not exotic skills. They are the parts of a job that are usually described as the parts you learn on the job.

Which surfaces the honest structural worry, and it should be named on screen rather than implied: if the bottom rung of the ladder is cut this year, the leadership pipeline is damaged five years out. Tacit knowledge is acquired by doing codified work under supervision. An organisation that stops hiring juniors because AI does the junior tasks has solved a cost problem this quarter and created a capability problem this decade.3

Demonstration protocol

1. List your own weekly tasks on screen. Actual ones, roughly fifteen, no editorialising.
2. Mark each C (codified) or T (tacit). Do it live. The hesitations are informative and should stay in.
3. Take one C task and run it through a model. Show the real result, including where it falls short.
4. Take one T task and try the same. Show why it fails — usually missing context, relationships, or accountability that cannot be handed over.
5. The audit is the deliverable. Not the model output. Say so explicitly and point the viewer at the handout.