On October 28, 2025, Amazon announced it was cutting roughly 14,000 corporate roles. The headlines were nearly unanimous: “Amazon cuts 14,000 jobs as it embraces AI”. But read the actual memo and the language is more careful than the coverage. The memo attributes the cuts themselves to organizational structure — too many layers, a need to “be organized more leanly, with fewer layers and more ownership.” AI appears as the reason for getting leaner, not as the direct cause of the layoffs: “This generation of AI is the most transformative technology we’ve seen since the Internet,” so the organization must move faster to keep up. And the actual forecast — that AI efficiency gains “will reduce our total corporate workforce” over the next few years — isn’t in the layoff memo at all. It comes from a separate memo CEO Andy Jassy wrote in June 2025.

So: the layoffs are present tense, attributed to org structure. “AI will shrink headcount” is future tense, written in a different document. Most headlines fused the two into one sentence.

And the cuts keep coming. The week this post went up (July 22, 2026), CNBC reported that Amazon laid off employees in its AGI unit — the team building Amazon’s own frontier models. This round hit the people building the AI. Two years of headlines like these raise an obvious question: with all these layoffs, did everyone land somewhere? Why doesn’t a layoff wave show up in the employment data? I’ll come back to that below.

This mismatch — layoffs now, attribution deferred to the future — is exactly what a July 2026 policy brief from the Stanford Institute for Economic Policy Research sets out to untangle (the brief). The author list is worth noting: SIEPR director Neale Mahoney, former Bureau of Labor Statistics Commissioner Erika McEntarfer, and researcher Karsen Wahal. McEntarfer was running the agency that produces the official US employment statistics until August 2025.

No layoff wave in the aggregate data

The brief’s core method is a comparison across exposure groups: score every occupation on how much its tasks overlap with what current AI can do. The scores aren’t the authors’ own — they use the AI Occupational Exposure (AIOE) index published by economists Felten, Raj, and Seamans in 2021, which maps the abilities each occupation relies on (from the O*NET database) against progress in different AI applications. Sort all occupations into quintiles by that score, then compare unemployment trajectories using IPUMS-CPS microdata, quarterly from 2015 through 2026. The two ends of the scale look roughly like this: the most exposed quintile includes genetic counselors and financial examiners — jobs whose core work is standardized processing of text and information; the least exposed includes dancers and construction helpers — jobs built on physical work done in person.

The result is the hardest number in the brief: since 2022, unemployment in the most exposed quintile has risen 0.77 percentage points; in the least exposed quintile, 0.85 percentage points. The workers who should be getting hit hardest by AI have seen their unemployment rise slightly slower than the workers who should be safest. The authors’ conclusion: this is a labor market that is softening across the board, not one being reshaped by AI-driven job destruction. What’s actually slowing hiring looks more like macro forces that hit the whole market rather than just exposed occupations — the Fed’s rate hikes starting in March 2022 and the unwind of pandemic-era overhiring.

The firm-side data is just as flat. In the Census Bureau’s business survey, only 5% of firms that have adopted AI report any effect on headcount — and within that 5%, increases and decreases split roughly evenly. In the Atlanta Fed’s executive survey, about 80% of executives say their AI investments haven’t yet changed headcount or productivity.

Back to the opening question: two years of layoff headlines, so why is the aggregate data quiet? Mostly scale and flow. The unemployment rate is a net figure. The US has about 160 million people employed, and per the BLS JOLTS data (latest release covers May 2026), even now roughly 1.7 million people are laid off or discharged every month — a monthly layoff rate of about 1.1%, near historic lows — while about 5.2 million are hired. A single company cutting ten or twenty thousand people is a rounding error in that flow; layoff announcements make the front page because of the company’s name, not the headcount. The other thing to see is that this is a low-hiring, low-firing market: companies aren’t cutting much, but they aren’t hiring much either. That squares with unemployment drifting up slowly across every exposure quintile, and it’s what the brief’s “broad softening” concretely means. One boundary condition worth stating: the brief’s data runs through mid-2026, so the late-July AGI-unit layoffs postdate it.

Then why does every layoff announcement invoke AI? The brief’s authors are explicitly skeptical of self-reported attribution and offer two hypotheses: firms may be using an AI narrative to explain contraction after pandemic-era overhiring, or using layoffs to free up cash for AI capital spending. My own read is that the incentives are simply asymmetric: “restructuring for the AI transition” sounds far better than “we overhired during the pandemic,” and it’s the story investors want to hear. A layoff announcement is first a document written for capital markets, and only second a description of the labor market.

The real signal: narrow, but real

The brief doesn’t stop at “false alarm.” It concedes that one signal is genuine: entry-level jobs.

In Q1 2026, unemployment among new US graduates hit 5.6%, 1.6 percentage points higher than three years earlier. The finer-grained evidence comes from the Brynjolfsson team’s “Canaries in the Coal Mine” work (Stanford Digital Economy Lab, using ADP payroll data): in the most AI-exposed occupations, employment of 22–25-year-olds has fallen about 16% relative to trend, while senior workers in the same occupations held roughly steady. The starkest case is software development, where employment of 22–25-year-old developers fell nearly 20% from its late-2022 peak through July 2025. And the declines concentrate in occupations where AI is more likely to automate human tasks — take the work over outright — rather than augment them, helping people do the same work faster. Same occupation, only the juniors hit, and concentrated in automation-type roles: that pattern does look like AI’s fingerprint.

But the brief flags a timing problem before you convict. The youth employment decline starts in late 2022 — and ChatGPT launched at the end of November 2022. It’s hard to believe firms were replacing workers with it in its first weeks. What was actually happening then was two other things: the Fed began hiking in March 2022 and tech hiring contracted on cue, and post-pandemic remote work stripped new hires of the learn-by-osmosis apprenticeship an office provides, so demand for juniors was already falling. Responding to exactly this critique, the Brynjolfsson team ran a revised analysis controlling for these confounders. After the controls, the visible decline in entry-level employment doesn’t appear until 2024. In other words, for the 2022–2023 stretch, non-AI explanations — rate hikes, remote work — fit better; if AI is genuinely compressing entry-level hiring, its identifiable effect only starts showing in 2024.

The accurate reading, then: the entry-level squeeze is real, AI is a suspect, but the current evidence can’t convict it alone.

The measurement problem: one question, four answers

The brief’s most useful methodological contribution is lining up the different measures of “AI adoption” side by side. Ask firms (Census Bureau survey): about 20% are using AI. Ask workers (household surveys): over 40% say they use AI at work. Ask executives: they claim over 80% of their employees use it. Look at what firms actually pay for (Ramp’s payments data — a non-representative sample): over half of customers are spending on AI tools. One question, answers from 20% to 80%, depending on whom you ask and how. Any article drawing conclusions from a single measure deserves a raised eyebrow.

The productivity evidence is just as split. On one side, the lab results: a call-center study measured a 15% overall productivity gain, with novices gaining about 30% (Brynjolfsson, Li, and Raymond’s original study, QJE 2025); in the GitHub Copilot experiment, developers finished a task about 56% faster — with the gains concentrated among less experienced developers. But METR’s randomized controlled trial supplies the counterexample, and its design is worth spelling out. Sixteen experienced open-source developers worked in large codebases they had maintained for years (averaging 22k+ GitHub stars and over a million lines of code), on 246 real tasks — bug fixes, features, refactors they were going to do anyway, averaging about two hours each. Each task was randomly assigned to “AI allowed” or “AI disallowed,” and each task was done exactly once — not the same task twice with a stopwatch, so there’s no learning effect from familiarity; the comparison is between the two groups of tasks, with randomization keeping their difficulty distributions comparable. The result: on AI-allowed tasks, completion took 19% longer on average — while the developers themselves, even afterward, still believed AI had sped them up by 20%. A roughly 40-percentage-point gap between perception and measurement, which also offers a clue to why executive surveys and measured data so often disagree: everyone feels the productivity gain.

None of this is a new story. In 1987, economist Robert Solow famously remarked: “You can see the computer age everywhere but in the productivity statistics.” US firms were buying computers in bulk, yet aggregate productivity growth stayed flat for years and didn’t visibly accelerate until the mid-to-late 1990s. The reason wasn’t that computers were useless — it was that converting a new technology into output requires reorganizing workflows, retraining workers, and building complementary investments, which takes years, sometimes a decade. The brief uses that history as the template for AI: model capabilities can jump a level a year, but the pace at which they seep into daily business operations — and from there into economic statistics — is much slower. Not “AI has no economic effect,” but “the effect can take a long time to become visible.”

Three judgments for industry watchers

First, before accepting any “AI caused X” claim about jobs, check three specifications. Whose data (firm self-reports, payroll records, or household surveys)? What time window (does it span the rate-hike and pandemic-unwind confounders)? How is exposure defined? A conclusion that doesn’t show its specifications is an opinion, not a fact.

Second, read layoff attributions as narrative, not evidence. Amazon is the type specimen: the layoff memo attributes the cuts to org structure; the “AI will shrink headcount” forecast lives in a different memo, in the future tense; the headlines merged the two into the present tense. Firms have every incentive to package restructuring as technological transformation, and the data — 5% of adopters reporting any headcount effect, split evenly between up and down — doesn’t currently support the announcements’ face value.

Third, watch the narrow signal, not the broad narrative. Aggregate unemployment won’t yield an AI signal any time soon. What’s worth tracking is entry-level hiring and unemployment in high-exposure occupations — young software developers, customer service reps, junior clerical and analysis roles (the Stanford Digital Economy Lab maintains a continuously updated dashboard). And keep the brief’s boundaries in mind: its core employment data covers only the US and stops in mid-2026, while agent-style tools have only just started entering real corporate workflows — by the “diffusion lags breakthrough” logic, today’s data simply can’t see them yet. That’s not a flaw in the brief; it’s an uncertainty the authors themselves repeatedly flag.

My overall judgment: AI’s effect on employment right now is real but narrow. It’s nearly invisible in the aggregates; the pressure on entry-level roles in high-exposure occupations is real, and AI is plausibly one of the drivers. Both loud narratives are distorted: the jobs-apocalypse story inflates a narrow signal into a tsunami, and corporate announcements dress up a macro correction as technological inevitability. Technology moves fast and diffuses slowly — and that lag is precisely the window policy and individuals have to adjust. Provided we watch the data, not the headlines.

References