Canaries in the Coal Mine August revision
S · Shapes the field Erik Brynjolfsson and Stanford colleagues use ADP payroll records for 3.5-5 million workers a month; the 2025 version became the most cited evidence on AI and entry-level jobs and this revision extends it to June 2026.
Summary based on the full text·Original: about 90 minutes to read
In short
The August 2026 revision of "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence" by Erik Brynjolfsson, Bharat Chandar and Ruyu Chen (Stanford Digital Economy Lab). With ADP payroll data through June 2026 it finds no broad AI-driven job loss, but young workers in AI-exposed occupations have fallen 19% behind peers in less-exposed ones. The gap comes from reduced hiring, sits where AI automates rather than complements work, and has widened since the first version in August 2025. The authors present it as early descriptive evidence, not causal proof.
Key numbers
- 19%employment shortfall for 22-25-year-olds in AI-exposed occupations vs less-exposed peers, June 2026
- 21 ppdivergence between -11% (two most exposed quintiles) and +10% (three least exposed) for 22-25-year-olds, Nov 2022-Jun 2026
- -0.179long-difference estimate: most vs least exposed quintile, workers aged 22-25, no controls
- 6%overall employment growth in the ADP sample, Nov 2022-Jun 2026 (4% in the most exposed quintile)
As reported by the source.
English·PDF·opens digitaleconomy.stanford.edu
What it reports
- Fact 1: average employment in the ADP sample rose about 6% from November 2022 to June 2026; the most exposed quintile grew about 4%.
- Fact 2: employment of 22-25-year-olds in the two most exposed quintiles fell about 11%, while the three least exposed grew about 10%. Total employment for this age group was roughly flat (-1.9%); workers aged 35-40 grew as much as 11%.
- About 60% of occupations in the least exposed quintile saw rising early-career employment, against about 30% in the most exposed quintile.
- Fact 3: the kept-pace shortfall was 15% at the July 2025 data vintage and 19% by June 2026. Earlier versions headlined firm-adjusted estimates of 13% (July 2025 data) and 16% (September 2025 data).
- Fact 4: adjustment runs through reduced hiring, not more separations. Fact 6: it shows in employment, not base pay.
- Fact 5: declines concentrate where Anthropic Economic Index usage is automating; complementary usage goes with flat or rising employment.
- Mechanisms: losses for young workers in codified-knowledge occupations, gains for experienced workers in tacit-knowledge occupations; women face higher average exposure.
How they got it
A balanced panel of firms from ADP payroll records, January 2021 to June 2026 (plus a 2018-2026 panel to test pre-trends), covering 3.5 to 5 million full-time workers a month. Job titles are mapped to SOC codes and linked to GPT-4-based exposure scores (Eloundou et al., 2024) and Anthropic Economic Index usage data. Occupation-level long-difference regressions add controls for interest-rate exposure, college share, remote work and teleworkability, and results are benchmarked against CPS and ACS.
Limits
- Descriptive, not causal; the authors say no single study is definitive.
- Estimates attenuate when controlling for occupational education, some divergence predates ChatGPT (notably around COVID-19), and the gap is larger in ADP data than in national surveys.
- The ADP sample over-represents manufacturing, wholesale, larger firms and high-exposure occupations; job titles are missing for roughly 30% of records.
Why it matters for companies
This is the strongest current evidence that AI is narrowing the entry route into codified-knowledge jobs while leaving experienced staff untouched. For an employer, it means the first visible cost of AI shows up in graduate and junior hiring, years before any headcount plan says so. Firms that stop hiring juniors in exposed roles should plan how future seniors will be trained. The monthly Canaries Dashboard lets HR compare its own hiring mix with the national pattern.
Summary based on the full text. Read the original for the full argument and data.Open PDF on digitaleconomy.stanford.edu
- Tags
- Hiring