AI-Exposed Jobs Fell 0.6% in August. The Gap Narrowed From the Wrong Side (35 Months)

Employment in America's most AI-exposed occupations fell 0.6% in August 2026 from a year earlier. In July the same figure was +0.1%, and we wrote then that if it went back below zero, the July rise would have been within the noise. It went back below zero. [Fact]

That is the smaller of the two things the August note says. The larger one is about the youngest workers. For 22-to-25-year-olds, employment in high-exposure jobs fell 4.4%, the 35th consecutive year-over-year decline. But for the first time in the three months we have tracked this series, their employment in the least-exposed jobs fell too, by 2.0%. The whole cohort is now shrinking, not just its AI-exposed half. [Fact]

This is the third reading in a monthly series we follow from the June and July notes. Three months is enough to test the questions we left open in July. Most of them now have answers, and one of the answers is the opposite of what it looks like.

The August figures

The Canaries Dashboard is a joint project of the Stanford Digital Economy Lab and ADP Research. It observes payroll records of millions of workers in more than 730 occupations at tens of thousands of private U.S. employers each month, links each occupation to an established measure of AI exposure, and compares employment in the most-exposed occupations against the least-exposed ones, with the clock starting at the wide release of generative AI in late 2022. [Fact]

The August 2026 reading, published September 23, 2026 under ADP chief economist Nela Richardson, reports the following. Across all ages, employment in high AI-exposure occupations was down 0.6% year over year while employment in the least-exposed occupations was up 0.2%. Workers aged 22 to 25 saw overall employment contract 3.1%. Within that band, high-exposure employment fell 4.4%, a streak ADP dates to October 2023, twelve months after the public release of ChatGPT, and least-exposed employment fell 2.0%. For workers aged 26 to 30, high-exposure employment fell 3.1% and least-exposed employment rose 0.1%. [Fact]

That is the whole release: six numbers and a methodology paragraph. It names no occupations, no sectors, and no confidence intervals. Everything below that goes beyond those six numbers is our arithmetic or our reading, and is labeled as such.

Three months side by side

Line up June, July and August and the series stops looking like a wobble around zero.

All ages, high exposure: -0.2%, then +0.1%, then -0.6%. All ages, least exposed: +0.6%, then +1.1%, then +0.2%. The gap between them was 0.8 points in June, 1.0 in July and is 0.8 points again in August. [Estimate — our subtraction of ADP's published rates.] The August high-exposure figure is the weakest all-ages reading of the three. The August least-exposed figure is also the weakest of the three. Both groups slowed; the penalty for being in an exposed occupation stayed roughly where it was.

Ages 22 to 25, high exposure: -4.3%, -3.0%, -4.4%. The July easing that looked like it might be a turn has been fully given back. August is the worst reading of the three for this group, by a tenth. [Estimate]

Ages 22 to 25, least exposed: flat, flat, -2.0%. This is the new number, and it changes the shape of the story more than any other figure in the release.

Ages 26 to 30: high exposure went -2.6%, -2.1%, -3.1%. Least exposed went +0.9%, +1.8%, +0.1%. The gap for this band was 3.5 points, then 3.9, and is now 3.2 points. [Estimate]

The gap narrowed. That is the bad news.

Here is the answer that reads backwards.

In July we said that if the gap for 22-to-25-year-olds kept narrowing while the streak continued, the base-effect explanation would gain weight: the decline aging out of the comparison window without the level recovering. The gap did keep narrowing. It was 4.3 points in June, 3.0 in July and is 2.4 points in August. [Estimate] On its own that looks like continued improvement.

It is not. Decompose the move. From July to August, high-exposure employment growth for this age band fell by 1.4 points (from -3.0% to -4.4%). Least-exposed employment growth fell by 2.0 points (from flat to -2.0%). The gap closed by 0.6 points because the comparison group deteriorated faster than the exposed group did, not because the exposed group recovered. [Estimate] A narrowing gap driven by the safe side collapsing is not convergence. It is the floor dropping.

The same thing happened one band up. The 26-to-30 gap narrowed from 3.9 to 3.2 points, and in July we had said a widening gap would move the retention story from hint to pattern. It did not widen. But high-exposure employment for this band got a full point worse (-2.1% to -3.1%) while least-exposed growth fell from 1.8% to barely positive. [Estimate] Again the gap closed from the wrong side.

So the base-effect hypothesis from July loses ground rather than gaining it. If the youngest band's decline were simply aging out of the year-over-year window, the high-exposure figure should have kept easing. It reversed. What the August note shows instead is a cohort-wide weakening for the youngest workers, with the AI-exposed occupations still worst, and a general softening across all groups that pulled the least-exposed side down toward it.

Two stories still fit one payroll series

The obvious reading is that AI adoption is now spilling from the most-exposed occupations into hiring of young workers generally. The August numbers are consistent with that. They are also consistent with something that has nothing to do with AI: a broadly cooler U.S. labor market in which the youngest workers, who are always the last hired, are the first to feel it, and in which AI-exposed occupations happen to be the ones that also over-hired in 2021 and 2022 and are still correcting. [Claim]

We cannot separate those from a six-number release, and we want to be direct about that. The Canaries Dashboard measures employment in occupations rated as exposed; it does not observe which employers deployed a tool or in which task. When the least-exposed group softens along with the exposed group, the share of the move attributable to AI specifically becomes harder to identify, not easier. The one thing that survives both readings is the ordering: within every age band ADP reports, and in every one of the three months, the AI-exposed occupations did worse. Thirty-five consecutive months of that ordering for the youngest workers is not noise.

The sample caveat also stands. ADP's client base is large but not a random draw of U.S. employers, and the monthly note publishes no revision history, so we do not know whether the July +0.1% was itself later revised. We are treating each month's note as published. [Claim]

Our own instrument shares the dashboard's weakness. The exposure profiles on our occupation pages come from the same family of task-based indices the dashboard links its payroll data to. If those indices mislabel which occupations are actually being changed, we inherit the error together, and agreement between the two is not independent confirmation.

Which occupations this reaches

The August note, like June and July, names no occupations. Richardson's June commentary pointed to software development and customer service at the high-exposure end and home health care at the low-exposure end, and nothing published since changes that grouping. [Fact] The exposure and task profiles for those roles are on our pages for software developers, customer service representatives and home health aides. The mapping from ADP's exposure bands to those pages is ours, not ADP's. [Claim]

What to do with an August like this

If you are 22 to 25 and looking at an AI-exposed field, the August note removes one comfort we offered in June: that the least-exposed path for your age group was at least holding flat. It is not, this month. The advice does not change, but the margin for error did. Target entry roles whose task bundle contains something a model cannot do unsupervised — a client relationship, accountability for a decision, a licensed or physical setting — and treat the first role as the hard part, because in this data the first rung is where the contraction is deepest.

If you are 26 to 30 in an exposed occupation, August is the worst reading yet for your band, and the narrowing gap to your least-exposed peers is not good news for the reason set out above. The durable protection is unchanged: make the part of your job your employer can see the part a model does not do.

For everyone reading monthly headlines: a note that says "the gap narrowed" and a note that says "employment in AI-exposed jobs fell at the fastest pace of the summer" are describing the same six numbers. Read for which side of the gap moved.

What September would need to show

Three months now give a direction for each band. The test for September is simple. If least-exposed employment for 22-to-25-year-olds returns to flat while high-exposure employment stays down near 4%, August was a broad-market wobble layered on a persistent AI-exposure penalty. If both stay negative, the youngest workers are facing a general hiring freeze that AI exposure makes worse but did not start. Either way the 35-month streak is the number to watch, and we will read the September note when ADP publishes it.

Sources

• ADP Research, Nela Richardson, "Canaries Dashboard: Employment in AI-exposed occupations slowed in August," published September 23, 2026 — https://www.adpresearch.com/research/canaries-dashboard-employment-in-ai-exposed-occupations-slowed-in-august
• ADP Research, "Canaries Dashboard: Employment in AI-exposed occupations rose slightly in July," published August 19, 2026 — July comparison figures — https://www.adpresearch.com/research/canaries-dashboard-employment-in-ai-exposed-occupations-rose-slightly-in-july
• ADP Research, "Canaries Dashboard: Employment in AI-exposed occupations contracted in June," published July 22, 2026 — June comparison figures — https://www.adpresearch.com/research/canaries-dashboard-2026-june
• Our earlier analyses: June and July

About this analysis

This article was written with AI assistance. Every percentage attributed to ADP is quoted from its published monthly commentary; the underlying dashboard is hosted by the Stanford Digital Economy Lab and we did not access its raw data for this article. The three-month comparisons, percentage-point gaps and the decomposition of the gap changes are our own subtraction of ADP's published figures and are labeled [Estimate]. The two-stories reading, the sample and revision caveats, and the mapping to our occupation pages are our interpretation, not ADP's. ADP Research material is quoted briefly under its copyright; no figures or charts are reproduced.