The AI Debate

Ask whether AI is killing entry-level jobs and you’ll get a confident yes from one camp and a confident no from another, each backed by real data. The honest answer, on the evidence available in mid-2026, is that this is a live empirical disagreement — and treating it as settled in either direction is a worse mistake than sitting with the uncertainty.

What It Is

Two credible research efforts reach different emphases using different methods on overlapping questions, plus a third data source — employers’ own stated reasons for layoffs — that measures something different from either: not what’s actually happening, but what companies find advantageous to say is happening.

How It Works

The strongest single piece of evidence for a real AI effect comes from the Stanford Digital Economy Lab, using ADP payroll microdata — real administrative payroll records, not a survey. Their November 2025 working paper found workers aged 22–25 in the most AI-exposed occupations experienced roughly a 16% relative decline in employment after generative AI’s spread, with replication studies finding similar effects (~13%). Young software developer employment specifically fell about 20% by mid-2025. Two details matter for how to read this: the decline runs through hiring, not layoffs — specifically, fewer people transitioning directly from outside the workforce into a first job, not more people losing jobs they already have — meaning AI isn’t (on this evidence) pushing young workers out, it’s narrowing the door they’d use to get in. And the effect concentrates specifically where AI automates a task a junior worker would otherwise have done; where AI merely augments a worker who still performs the task, the effect largely disappears. Consistent with that split, employment for young workers in health-aide and nursing-adjacent roles — where AI isn’t substituting for the entry-level task — is growing faster than for older workers in the same roles.

The Economic Policy Institute’s “Class of 2026” analysis complicates the clean AI story without dismissing it. Their case: the depressed overall hires rate (a fifteen-year low — see Macro Labor Data) explains most of the young-graduate weakness on its own, independent of any AI-specific mechanism. Both young college graduates and young non-college workers have seen rising unemployment over the past three years — an AI-exposure story predicts a college-specific or exposed-occupation-specific pattern, and the data doesn’t track that cleanly. The weakness shows up broadly across industries, not concentrated in AI-exposed ones. The information sector, the industry most commonly cited as AI-exposed, employs only 2.3% of young college graduates — even a severe, concentrated effect there can’t explain broad-based graduate weakness. And EPI points to evidence that new-graduate employment weakness has been building for more than two decades — well before generative AI existed as a labor-market factor at all.

A third data source measures something different from either of the above: what employers say. Challenger, Gray & Christmas tracks the stated reasons companies give publicly for layoffs, and AI has been the leading cited reason for four consecutive months as of mid-2026 — a streak with no precedent in their data. The monthly share cited AI rose from 7% in January to roughly 40% by May; cumulative 2026 AI-cited cuts already exceed the full-year 2025 total. But this figure has a critical limit: Challenger counts announcements, and records whichever reason the employer chooses to state publicly. “AI efficiency gains” is a far more market-flattering explanation than “we over-hired in 2021–22,” “demand is soft,” or “we’re absorbing tariff-driven cost pressure.” Read this figure as an upper bound on narrative attribution, not a measurement of actual causal AI displacement.

Synthesis

The most defensible reading holds two things true at once: the AI effect on entry-level hiring is real and identifiable in high-quality administrative data, but it’s riding on top of a much larger cyclical freeze — the fifteen-year-low hires rate — and cleanly separating the two isn’t currently possible with available data. The EPI critique is a legitimate scientific check on over-attribution, not a dismissal of the Stanford finding. AI is plausibly suppressing entry-level, automatable hiring at the margin, while the dominant driver of aggregate graduate weakness is the broader freeze that predates and exceeds any AI-specific effect. That’s a less satisfying headline than either “AI is destroying entry-level work” or “AI’s role is overblown” — and it’s the honest one.

Why It Matters