Tech and AI-exposed workers hit by downturn before generative AI

Across nearly all measures, deterioration in AI-exposed jobs began in early 2022. Unemployment risk rose faster in occupations with high exposure to language-model tasks than in less exposed jobs. This increase occurred months before ChatGPT’s release in November 2022 and before generative AI tools entered routine workplace use.

A new working paper on arXiv reveals that employment conditions in jobs most exposed to AI began deteriorating months before ChatGPT entered public use, suggesting that economic pressures rather than generative AI itself were already reshaping career prospects in key sectors.

The study, titled AI-Exposed Jobs Deteriorated Before ChatGPT, reconstructs how AI-exposed occupations evolved before and after the launch of large language models.

The findings challenge a growing narrative that places generative AI at the center of recent labor market disruptions. Instead, the evidence shows that structural shifts were already underway, particularly in technology-heavy roles, well before conversational AI tools became widely accessible.

Job losses preceded generative AI’s public debut

To test whether ChatGPT marked a turning point for AI-exposed jobs, the researchers analyzed U.S. unemployment insurance records, millions of LinkedIn career trajectories, and thousands of university course syllabi. Together, these data sources allowed the team to track job separations, reemployment rates, wage outcomes, and skill alignment over time.

Across nearly all measures, deterioration in AI-exposed jobs began in early 2022. Unemployment risk rose faster in occupations with high exposure to language-model tasks than in less exposed jobs. This increase occurred months before ChatGPT’s release in November 2022 and before generative AI tools entered routine workplace use.

After ChatGPT became widely available, the pattern did not accelerate. Instead, unemployment risk in AI-exposed roles stabilized. This timing runs counter to the idea that generative AI triggered a sudden collapse in demand for affected jobs. The data instead point to a labor market already adjusting to slower growth, tightening financial conditions, and post-pandemic corrections in the technology sector.

The effect was most pronounced in computer and mathematical occupations, fields that had expanded rapidly during the pandemic years. As hiring slowed and layoffs increased in tech-heavy industries, AI-exposed workers experienced sharper employment volatility than those in less exposed roles. The study finds that this divergence emerged well before generative AI tools became a realistic substitute or complement in everyday work.

Geographically, the pattern held across most U.S. states, indicating that the shift was not driven by isolated regional shocks. The consistency strengthens the argument that macroeconomic forces, rather than a single technological breakthrough, were the dominant drivers of early job market deterioration.

Early-career workers bear the brunt of the slowdown

While AI exposure mattered, the study identifies career stage as a critical factor shaping outcomes. Early-career workers faced significantly worse prospects than mid- and late-career professionals, particularly in AI-exposed fields.

Graduates entering the labor market from 2021 onward were less likely to secure jobs in highly AI-exposed occupations than earlier cohorts. Those who did find work took longer to do so and often entered at lower starting wages. Importantly, these gaps opened before ChatGPT’s release, undermining claims that generative AI directly displaced new entrants.

The findings suggest that hiring freezes, reduced entry-level openings, and cautious corporate expansion played a larger role than automation. Employers appeared to prioritize experienced workers who could navigate uncertainty, while cutting back on junior roles that typically serve as pipelines into technical careers.

This pattern has long-term implications. Early-career setbacks can compound over time, affecting lifetime earnings and career mobility. The study warns that focusing policy debates solely on AI displacement risks overlooking a cohort of workers already affected by broader economic realignment.

At the same time, the research shows that not all early-career workers were equally vulnerable. Those whose education aligned closely with AI-related skills fared better after ChatGPT’s launch. Graduates with coursework emphasizing tasks relevant to language models and data-driven work secured higher first-job wages and found employment more quickly than peers without such alignment.

This finding complicates simplistic narratives of AI replacing workers. Instead, it highlights a growing divide between workers whose skills complement AI systems and those whose training does not map as clearly onto evolving job demands.

Education and skills still matter in an AI-shaped labor market

By analyzing university course syllabi, the researchers identified which academic programs emphasized skills closely related to AI-exposed tasks, such as text analysis, coding, data interpretation, and complex problem-solving.

Graduates from programs with stronger alignment to these skills saw tangible benefits after ChatGPT’s release. They experienced faster job placement and higher starting salaries, even within occupations classified as highly AI-exposed. This suggests that AI has not erased the value of human capital but has reshaped which forms of it are rewarded.

The results challenge fears that investing in AI-related education is futile in a world of increasingly capable machines. Instead, the data indicate that AI-complementary skills continue to command a premium, even amid labor market uncertainty.

At the same time, the study cautions against interpreting AI exposure as a static risk. Jobs are not replaced overnight, nor are they uniformly affected. Tasks within occupations evolve, and workers who adapt their skill sets can remain competitive even as tools change.

The authors argue that policymakers and educators should resist framing generative AI as a singular shock. Treating ChatGPT as a clean break obscures the structural forces already reshaping work and risks misdirecting responses. Training programs focused solely on avoiding AI exposure may leave workers less prepared for a labor market where AI is embedded across industries.

Instead, the study points toward a more nuanced approach. Strengthening pathways into AI-complementary roles, supporting early-career workers through hiring downturns, and improving alignment between education and labor market needs may do more to stabilize employment than attempting to slow or restrict AI adoption.