AI Employment Squeeze Hits Young Workers The Hardest Now todayheadline

AI employment is already reshaping who gets hired first. A new working paper from Stanford researchers, drawing on millions of ADP payroll records through July 2025, finds that early-career workers in the most AI-exposed occupations have lost ground while older peers hold steady or advance.

The authors report a 13 percent relative decline in employment for ages 22 to 25 in highly exposed roles, with losses concentrated where generative AI automates tasks rather than augments them. Wages barely budged. The timing matters, too, with the shift beginning in late 2022 as LLM use surged.

AI Employment Effects, With A Twist

The team links job titles to established exposure metrics and to real usage patterns of large models. One measure comes from task-level exposure scores adapted to occupations. Another comes from observed queries to a leading model, which the researchers sort into automative or augmentative use. Put plainly, if AI mostly does the task for you, young workers in that job are hurting. If AI mostly helps you do the task, the entry-level squeeze is less pronounced.

These six facts provide early, large-scale evidence consistent with the hypothesis that the AI revolution is beginning to have a significant and disproportionate impact on entry-level workers in the American labor market.

What The Data Actually Show

The analysis uses monthly, individual-level payroll records from the largest U.S. provider, keeping a consistent firm sample from January 2021 to July 2025. That scale lets the authors see patterns that national surveys often blur. Overall employment keeps rising. But for ages 22 to 25 in the top exposure quintiles, headcount slides, notably in software development and customer service. In less exposed jobs, or for workers over 30 in the same exposed jobs, employment is flat or up. When the authors absorb firm-time shocks in an event-study framework, the early-career hit in highly exposed occupations remains large and statistically significant. Compensation, deflated to 2017 dollars, shows little divergence by exposure or age, hinting at sticky wages and adjustment via hiring instead.

Automation Hurts More Than Help

Using the Anthropic Economic Index to characterize real model use, the declines concentrate in occupations where queries look automative, not augmentative. In jobs where model use mostly complements human effort, the entry-level trend is muted or even positive. That split supports a simple story: codified, trainable tasks are easier for AI to take over, and those tasks are often what junior hires do first. Tacit knowledge, learned on the job, still buys resilience for experienced workers.

Context Without Hype

The results do not rest on tech sector gyrations or remote-work shuffles. Patterns are similar after excluding computer occupations and information-sector firms, and they appear in both teleworkable and non-teleworkable jobs. The divergence intensifies after late 2022, lining up with the spread of generative AI at work. Overall, the six facts describe an economy where AI’s earliest labor shock lands at the entry point, not across the board.

Key Findings

• Population and period: 3.5–5 million workers per month, consistent firm sample, Jan 2021 to Jul 2025, ADP payroll data.
• Main effect: Ages 22–25 in the most AI-exposed occupations saw a 13 percent relative employment decline versus least exposed peers.
• Who is spared: Older workers in the same occupations generally stable or growing, less exposed jobs show no comparable decline.
• Mechanism split: Declines concentrate where AI use is automative; muted or positive entry-level trends where use is augmentative.
• Adjustment margin: Employment changes outpace wage changes, suggesting short-run wage stickiness.
• Robustness: Effects persist with firm-time fixed effects, after excluding tech roles, and across teleworkable and non-teleworkable occupations.
• Timing: Divergence starts in late 2022, coincident with rapid LLM adoption at work.

Why It Matters

Early-career roles are designed to teach the craft. If AI absorbs a chunk of the beginner tasks, the first rung gets thinner. That risks a pipeline problem for skills development. It also puts pressure on firms and training programs to redesign apprenticeships around the tasks that teach tacit judgment. The authors caution that other forces may be at play, but the near-real-time payroll view strengthens the case that generative AI has begun to shift entry-level hiring in exposed jobs.

What To Watch Next

1. Augmentation investments: Teams that architect workflows so AI assists rather than replaces could stabilize junior hiring.
2. On-ramp design: Rotations that prioritize tacit skills, code review, troubleshooting, and stakeholder work may preserve learning ladders.
3. Metrics beyond headcount: Track time-to-proficiency, mentor bandwidth, and quality outcomes as junior tasks evolve.

Takeaway

The earliest measurable labor impact of generative AI is concentrated at the entry level in highly exposed occupations, with hiring cooling more than wages. Where AI augments human work, the squeeze is smaller. The pipeline for tacit skills now depends on redesigning first-year tasks.

Sources and further reading: Stanford Digital Economy Lab working paper; Stanford AI Index; Anthropic Economic Index; BEA PCE Price Index; background on exposure measures via Eloundou et al., Occupational exposure to Generative AI.

Journal: Stanford Digital Economy Lab Working Paper
DOI: None assigned (latest version at Stanford Digital Economy Lab)