New Payroll Data Ties AI to Early-Career Job Losses in Exposed Fields

A new Stanford study using ADP payroll data finds a 13% employment decline for 22–25-year-olds in highly AI-exposed jobs since ChatGPT, while older workers and less-exposed occupations saw steady or rising employment. The effect is stronger where AI automates tasks and persists even within firms, suggesting targeted displacement rather than economy-wide weakness. Thompson concludes this is the strongest evidence so far that AI is already reshaping early-career white-collar work, warranting curricular changes and continued scrutiny.

Key Points

New Stanford analysis of ADP payroll data shows a roughly 13% employment decline since ChatGPT for 22–25-year-olds in highly AI-exposed roles, especially software development and customer service.
Employment has risen or held steady for older workers and for less-exposed occupations such as home health aides, indicating the effect is concentrated, not economy-wide.
Within the same firms, highly AI-exposed jobs decline relative to less-exposed jobs, suggesting the pattern is not driven solely by firm-level shocks like rates or macro conditions.
Jobs where AI usage is automative show youth employment declines; where AI is augmentative, similar declines are not observed.
Younger workers’ tasks overlap more with what LLMs can replicate (codified, short-horizon, easily evaluated work), while older workers rely more on tacit, strategic capabilities; colleges should teach AI tool use and emphasize capabilities AI lacks.

Sentiment

Mixed but leans skeptical that AI is the primary cause of youth job losses to date; consensus is multi-causal with macro, policy, and offshoring leading early, and AI increasingly contributing as an accelerant and justification in 2024–2025.

In Agreement

AI now reduces demand for junior workers in automatable, well-specified tasks (entry-level coding, customer service, basic marketing), matching the paper’s automate-vs-augment distinction.
Management has updated staffing models to expect AI efficiencies, pausing or cutting junior hiring and using AI to ‘do more with less.’
Significant AI infrastructure spending is crowding out budgets for hiring and training, especially for entry-level roles.
Within-firm declines in AI-exposed roles and the timing (mid-2024 onward) suggest AI is increasingly a real factor, even if not the sole cause.
Fields like translation, copywriting, and illustration show visible displacement from ‘good enough’ AI, aligning with the study’s findings.

Opposed

The timeline doesn’t fit: declines began before widespread LLM adoption; macro shocks (end of ZIRP, rapid rate hikes, COVID overhiring correction) explain more of the early drop.
Section 174’s 2022 change (amortizing R&D) made engineering headcount more expensive and likely drove layoffs; studies must control more explicitly for this, especially in tech.
Offshoring/outsourcing—turbocharged by remote work normalization—is a primary driver for entry-level losses (e.g., customer service moving abroad), not AI.
Musk’s Twitter layoffs signaled headcount cuts were possible, independent of AI performance; ‘lean headcount’ is an investor fashion rather than an AI necessity.
Methodological concerns: narrow windows confounded by pandemic-era volatility; need longer pre-2020 baselines, exclusion tests (e.g., remove SWE), and checks for cohort-aging artifacts in age bucket analyses.
AI is often a scapegoat for cost-cutting and hiring freezes; observable service quality declines suggest replacement hasn’t matched human output.