Somewhere between “AI is coming for everyone’s job” and “AI anxiety is overblown” sits the actual, measured, slightly less dramatic truth: AI does appear to be reshaping who gets hired for a very specific slice of the labor market — and that slice is entry-level. The best evidence for this isn’t a vibes-based LinkedIn post or a consulting-firm survey. It’s a running research project out of Stanford’s Digital Economy Lab, led by economist Erik Brynjolfsson along with Bharat Chandar and Ruyu Chen, that has been tracking millions of real payroll records since the launch of ChatGPT.

The project has a title that doubles as a warning label — “Canaries in the Coal Mine?” — and it’s now old enough to have a sequel. The original paper landed in August 2025. A year of monthly updates and robustness checks later, the lab published a follow-up in August 2026 confirming the trend hadn’t just continued, it had gotten worse. Here are six numbers from that body of work, what they actually mean, and where the story gets more complicated than a single headline stat.

1. 19% — how far behind AI-exposed young workers have fallen, as of mid-2026

The lab’s August 2026 update puts it plainly: workers aged 22 to 25 in the most AI-exposed occupations are now about 19% below where their employment would be if they’d simply kept pace with similarly aged workers in less-exposed jobs. That’s not a one-time dip. It’s a gap measured against a credible counterfactual — what young workers in exposed fields would look like if AI weren’t a factor — and it has been widening steadily since researchers first started tracking it.

Crucially, the researchers found no comparable gap for more experienced workers doing the same AI-exposed jobs. This is the detail that keeps this study in the news: it isn’t “AI is killing jobs.” It’s “AI is killing the easiest way in.”

2. 13% — the original number that made this study go viral in 2025

Before the gap was 19%, it was 13%. That was the headline figure from the original paper, as the lab’s own October 2025 research recap put it: within firms, entry-level hiring in AI-exposed jobs had fallen 13% relative to hiring in less-exposed jobs, using payroll data running through mid-2025.

That 13% is the number most people remember, because it was the first hard evidence that something structural — not just “a tough grad job market” — was happening to the bottom of the career ladder in AI-adjacent fields. Thirteen months later, the lab’s own numbers say it undersold the problem.

3. -11% and +10% — the plain-math version of the divergence

Model specifications and “relative to counterfactual” framing can obscure how stark the raw numbers are. So here’s the un-adjusted version: from November 2022 (right around ChatGPT’s launch) through June 2026, employment for 22- to 25-year-olds in the two most AI-exposed occupation quintiles fell by roughly 11%. Over that same stretch, employment for the same age group in the three least-exposed quintiles grew by roughly 10%.

Put those next to each other and you get a roughly 21-point swing between “exposed to AI” and “not exposed to AI” for workers at the exact same career stage. No regression needed to see that one.

4. 3 — the job categories doing most of the damage

The researchers didn’t just wave at “AI-exposed jobs” as an abstraction. Their case studies repeatedly point to three categories: software development, customer service, and clerical work. These are jobs where a large share of the day-to-day task list is “codified knowledge” — the kind of formal, documented, teachable work that’s relatively easy to hand to a large language model, as opposed to the tacit, learned-on-the-job know-how that still favors people with years in the seat.

That distinction — codified versus tacit knowledge — is one of the more useful additions in the 2026 update. It’s less “which industry” and more “which kind of knowledge your job actually runs on,” and it’s a far better predictor of who’s getting squeezed than job title alone.

5. Zero — the number of mass layoffs this study actually documents

Here’s the part that gets lost whenever this study goes viral: the researchers are explicit that they found no evidence of broad, economy-wide job destruction. The effect shows up almost entirely through hiring, not firing. Companies aren’t mass-cutting young AI-exposed workers; they’re just not hiring as many of them in the first place, while headcount for experienced workers in the same roles holds steady or grows.

“We find no evidence of widespread, economy-wide job displacement.”

That’s a direct line from the researchers themselves, and it matters for how you read every other number on this list. This isn’t a story about robots clearing out the office. It’s a story about a door quietly getting narrower for the people trying to walk through it for the first time.

6. 2024 — the year the effect actually starts to show up, under the toughest test

One of the sharpest, least headline-friendly findings came from a February 2026 follow-up that stress-tested the original paper against an obvious objection: maybe this is really about interest rates, not AI. The researchers checked, and found that AI-exposed occupations are, if anything, less sensitive to interest-rate swings than average — construction, for instance, is the opposite profile: high rate exposure, low AI exposure.

More interesting is what happened when they controlled for every firm-specific factor they could using the strictest statistical specification available: the AI-exposed employment decline for young workers doesn’t become statistically significant until 2024, not late 2022 right after ChatGPT launched. Under that specification, the gap reached about 16% by October 2025, versus the simpler 13% figure from the original paper covering a similar period. Translation: the effect is real, it’s not just a 2022 blip, and depending on how conservatively you measure it, the honest range has probably always sat somewhere between “bad” and “worse than the headline number.”

So what does an actual early-career person do with this?

Not panic, mostly. The lab’s own researchers keep repeating, almost like a disclaimer, that these are descriptive patterns rather than proof of causation, and that the overall effect on total employment is still small. But “small in aggregate” is cold comfort if you’re a 23-year-old trying to break into software, support, or admin work specifically. The practical read: codified, documented, easily-explained tasks are the ones getting automated around young hires first, so the hedge is the same one career advisors have been pushing anyway — build a portfolio of things you’ve actually shipped, lean into the tacit, judgment-heavy parts of a role that are harder to hand to a model, and treat “I used the AI tools to do more, faster” as a selling point rather than something to downplay.

Frequently asked questions

Is this Stanford study saying AI is destroying jobs overall?

No. The researchers are explicit that they found no evidence of widespread, economy-wide job displacement. The effect is concentrated and specific: fewer new hires for young workers in a handful of AI-exposed occupations, not a broad wave of layoffs.

What is the actual size of the gap right now?

As of the lab’s August 2026 update, workers aged 22 to 25 in highly AI-exposed occupations were about 19% below where their employment would be had they kept pace with similarly aged workers in less-exposed fields — up from roughly 13 to 15% when the original research came out a year earlier.

Which jobs are most affected?

The research repeatedly points to software development, customer service, and clerical work as case studies, with the common thread being “codified” task knowledge — formal, documented, teachable work that large language models can pick up relatively quickly, as opposed to tacit expertise built through years on the job.

None of this is the last word — the researchers themselves keep adding robustness checks, caveats, and new data vintages, and they’ve said this is a trend worth monitoring rather than a settled verdict. For more on how AI is reshaping entry points into creative and knowledge work, head over to our Perspectives section.