TL;DR
- Using millions of U.S. payroll records from ADP through July 2025, a Stanford University team finds that AI’s earliest labor market impacts are concentrated among entry-level workers in highly AI-exposed occupations.
- The biggest hit is to “first jobs” (ages 22–25) in roles like software development and customer service.
- Declines are strongest where AI automates tasks; where AI augments workers, entry-level employment holds up or grows.
- Adjustments are happening through headcount, not pay—wages are stickier.
- These patterns persist after controlling for firm-level shocks and across multiple robustness checks.
What the Stanford team analyzed
Stanford University researchers matched:
- High-frequency payroll data from ADP (millions of workers, tens of thousands of firms) through July 2025.
- Two occupation-level AI exposure measures:
- GPT-4-based exposure (Eloundou et al., 2024).
- Task-level usage of Anthropic Claude (Handa et al., 2025), including whether AI use is automative or augmentative.
This let them track who is most affected, when, and how, in near real-time.
Six facts shaping the future of work
- Early-career employment is falling in AI-exposed jobs
- For software developers and customer service reps, employment for ages 22–25 declined sharply after late 2022. Software developers saw nearly a 20% drop from peak for 22–25-year-olds (Figure 1, page 10).
- Controlling for company-by-month shocks, the relative employment decline for the most exposed quintile among 22–25-year-olds is about 12–13 log points (Figure 9, page 20).
- Overall employment is still growing—but young workers are stalling
- Aggregate employment remains robust, but 22–25-year-olds have flattened out (Figure 4, page 13).
- From late 2022 to July 2025, 22–25-year-olds in the most AI-exposed occupations fell ~6%, while 35–49-year-olds in the same occupations rose 6–9% (Figure 5, page 14).
- Automation hurts entry roles; augmentation doesn’t
- Occupations where AI usage is primarily automative show pronounced entry-level declines (Figure 7, pages 17–18).
- Occupations where AI usage is primarily augmentative don’t show the same entry-level drop; some even grow faster (Figure 8, page 18).
- It’s not just macro or firm-specific shocks
- Even after absorbing firm-time shocks (e.g., interest rates, sector slowdowns), the exposure-linked decline for 22–25-year-olds remains large and significant (Figure 9, page 20).
- The adjustment margin is headcount, not pay
- Wages show little divergence by age or exposure; the visible action is in employment levels (Figures 10–11, pages 22–23).
- Robust across tech/non-tech and remote/non-remote
- Excluding computer occupations and information-sector firms yields similar patterns (Figures A4–A5, pages 38–39).
- Patterns persist in both teleworkable and non-teleworkable occupations (Figures A6–A7, pages 40–41).
- Longer time windows show divergence beginning with generative AI’s rise in late 2022 (Figures A8–A11, pages 42–45).
Why are young workers hit hardest? Stanford’s hypothesis
AI substitutes best for codified “book” knowledge, not tacit knowledge (experience, judgment, organizational know-how). Entry-level workers supply more of the former, while experienced workers lean more on the latter—so early-career roles in exposed occupations are first to feel the squeeze (Discussion, pages 4–5).
What this means for your career
Move from “tasks AI can fully automate” to “work AI helps you do better.”
Examples: product/ops with customer and compliance touchpoints, sales engineering and solution design, supply chain and field ops, clinical and care roles with hands-on complexity.Build tacit advantage:
Business context, client empathy, risk tradeoffs, cross-functional orchestration, and change management.Operationalize AI augmentation:
- Prompt design and critique (ask better, verify better).
- Chain tools into workflows; measure outcomes with A/B tests.
- Data governance and compliance-by-design.
Show your receipts:
Portfolio projects that document end-to-end “AI + you” value: problem framing → AI workflow → validation → deployment → ROI.
What this means for employers
Don’t chop the ladder:
Replacing entry roles with AI may save now but depletes future talent. Consider “AI + apprenticeship” models where juniors ship faster with AI and seniors gatekeep quality and pass on tacit knowledge.Measure the right things:
Track both automation savings and human capability growth. Balance near-term productivity with long-term bench strength.
Limitations and scope
- The study leverages U.S. ADP payroll data and occupation-level AI exposure proxies; it infers realized changes but doesn’t claim universal causality.
- Still, the age-by-exposure patterns are consistent, sizeable, and start right when generative AI adoption surged (late 2022).
- Results align with broader evidence that general-purpose tech shifts begin with task reallocation before wages fully adjust.
Source and citation
- Stanford University Digital Economy Lab: “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence” (Aug 26, 2025).
Authors and affiliations:- Erik Brynjolfsson (Stanford University & NBER)
- Bharat Chandar (Stanford University)
- Ruyu Chen (Stanford University)
Data: ADP payroll records (through July 2025).
AI exposure: Eloundou et al. (2024, GPT-4 β); Anthropic Economic Index (Handa et al., 2025).
Link: https://digitaleconomy.stanford.edu/publications/canaries-in-the-coal-mine/
Key figures referenced:
- Early-career declines in exposed jobs: Figure 1 (page 10)
- Aggregate trends and young worker stall: Figure 4 (page 13); Figure 5 (page 14)
- Automation vs. augmentation split: Figure 7 (pages 17–18); Figure 8 (page 18)
- Firm-time controlled effects: Figure 9 (page 20)
- Wages vs. employment: Figures 10–11 (pages 22–23)
- Robustness (non-tech, non-remote): Figures A4–A7 (pages 38–41)
- Longer window divergence: Figures A8–A11 (pages 42–45)
If you run a team or you’re starting a career, the Stanford University takeaway is clear: let AI do the repeatable work, and double down on the human work—context, judgment, collaboration—that machines can’t yet replace.
