The stat inside the stat

Stanford's economists put out an update to their jobs paper this week, and the number I keep turning over isn't the headline 19%. It's that the same measurement said 13% a year ago. A six-point jump in twelve months, on a metric tracking whether 22-to-25-year-olds can get hired at all, doesn't show up as a mass layoff on the news. It shows up as a stack of rejected applications nobody writes a story about, one at a time, until an economist adds them up.

The paper is called "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence," from a team led by Erik Brynjolfsson, and this August edition revises last year's version with fresh ADP payroll data. ADP processes payroll for a huge slice of the US workforce, so the sample is real hiring and firing, not a survey. The researchers scored occupations for "AI exposure" two ways: a labor-market impact index built by earlier researchers (Ars Technica noted its own critical read of that methodology back in February), and Anthropic's own Economic Index, which tags occupations by how people actually use Claude day to day. Google published a comparable Gemini-usage occupational report last month, and the two datasets apparently point the same direction, which is the kind of independent confirmation this field doesn't get often.

Here's what the numbers actually say. Since 2022, employment for 22-to-25-year-olds in the top 40% of AI-exposed occupations has fallen about 11%. In the bottom 60%, least exposed, it's grown about 10% over the same stretch. Zoom out to the whole workforce, all ages, and the gap mostly disappears. This is not "AI is destroying jobs" in aggregate. It's specifically young workers, specifically in exposed fields, and it's specifically about not getting hired in the first place. The researchers checked firings and quits separately. Neither moved much. Companies aren't pushing young workers out. They're just not opening the door.

That's the mechanism that makes this different from a normal downturn story, and it's why I don't think the framing "AI is coming for jobs" is quite right here. What's actually happening is narrower and, honestly, more interesting: the first rung of a lot of career ladders is being quietly removed while the ladder above it stays intact.

Automative versus augmentative is the real predictor

The paper's most useful move is a distinction Anthropic's index already draws: whether people use AI on a task to replace the work (automative) or to help with the work (augmentative). Accountants, auditors, receptionists, and information clerks show heavy automative usage, and their entry-level numbers are the worst in the dataset. CEOs and registered nurses sit at the augmentative end, and their young-worker employment lines look flat to rising. The researchers' own line: "automation-oriented uses of AI substituting for labor while complementary uses are associated with flat or rising employment." That's not a surprising sentence in isolation. What's useful is that it's now a testable prediction, not a vibe.

The honest pushback here is correlation. Tech hiring has been soft generally in 2026 for reasons that have nothing to do with AI: rate environment, a post-2024 hiring hangover, sector-specific overcapacity from the 2021-2023 boom. Maybe "AI-exposed" occupations just happen to overlap with sectors already cutting junior headcount for unrelated reasons, and the AI-exposure score is picking up that correlation, not causation. I take that seriously. But the paper's own comparison group is other young workers in the same broad economy, same interest-rate environment, split only by occupational AI exposure. The gap between those two groups is what's widening, not shrinking, as the wider hiring market has actually started to loosen elsewhere. If it were purely macro, I'd expect the gap to compress this year. It didn't. It grew.

What to do with this if you're on either side of it

If you're early-career and job hunting in accounting, basic bookkeeping, entry-level customer support, or general administrative work, this data says the field itself is worth questioning, not just the specific employer that rejected you. If you're hiring, the paper is a decent proxy for a question worth asking directly: for the junior role you're not filling this quarter, is that a real headcount decision or a quiet default you haven't examined?

For builders

• Check your own occupation against Anthropic's Economic Index — is your daily AI usage mostly automative (replacing a task) or augmentative (helping you do it)? That split predicts your field's hiring trend better than the job title does.
• If you manage a team, look at your last four open reqs: how many were junior roles that got quietly closed or downleveled instead of backfilled? That's the leading indicator, not the layoff number.
• If you're early-career in an exposed field, weight roles by whether the job description reads as "does the task" or "reviews and directs the task" — the second category is where the augmentative jobs are.

Further reading

• Stanford Digital Economy Lab: Canaries in the Coal Mine? (August 2026 update) — the primary paper and underlying charts
• Ars Technica — AI is hitting entry-level jobs hardest, Stanford study finds — source article for this piece
• Anthropic Economic Index — the automative/augmentative occupational classification used in the study