1. Motivation and Context

The rapid proliferation of Generative Artificial Intelligence (GenAI) has instigated a critical global discourse concerning its potential transformative impact on the labor market.

Much like the debates surrounding the social cost of carbon or the appropriate discount rate in climate economics, the discussion on AI often generates more heat than light. It spans utopian predictions of enhanced productivity ^[14], dystopian fears of widespread job displacement ^[3], and skeptical views suggesting minimal short-term effects ^[15].

Historically, technological change has affected tasks, occupations, and industries heterogeneously. This suggests the existence of "canaries in the coal mine"—early indicators that serve as harbingers of broader economic shifts. To date, however, the debate has often lacked a sound empirical base, relying heavily on anecdotal evidence or theoretical models without sufficient data calibration.

2. Data and Methodology

2.1 The ADP Administrative Dataset

To confront the empirical deficits characterizing the current AI debate, the research employs an exceptionally rich dataset: high-frequency administrative payroll records sourced from ADP ^[2]. As the largest payroll software provider in the United States, this sample tracks monthly, individual-level employment dynamics through July 2025, encompassing millions of workers across tens of thousands of firms.

This data architecture offers a distinct advantage over traditional sources like the Current Population Survey (CPS). It provides a near real-time view of labor market adjustments ^[8], allowing for a degree of granularity and precision necessary to disentangle the specific impacts of technological shocks from general economic volatility.

2.2 Measuring AI Exposure

The core analysis links these payroll records to established measures of occupational AI exposure, creating a framework similar to how one might index technological change or energy intensity. We rely on two primary indices:

1. GPT-4 Exposure Measures: Based on the seminal contribution of Eloundou et al. ^[11], these metrics estimate AI exposure via O*NET tasks, validated using ChatGPT and human labeling.
2. The Anthropic Economic Index: Derived from Handa et al. ^[13], this index estimates generative AI usage across O*NET tasks based on millions of user conversations with the Large Language Model (LLM) Claude.

Crucially, the measure from Handa et al. allows us to distinguish tasks based on a fundamental dichotomy: whether the technology functions as a substitute (automative) or a complement (augmentative).

3. Empirical Findings

The analysis documents six key facts characterizing the labor market since the widespread adoption of generative AI began around late 2022. These findings highlight the distributional consequences of technological change across different age cohorts and skill levels.

3.1 Heterogeneity in Employment Effects across Age Cohorts

The most pronounced finding is a 13 percent relative decline in employment experienced by early-career workers (ages 22–25) in the most AI-exposed occupations. This holds true even after controlling for firm-level shocks.

For professions deemed highly exposed—such as software developers and customer service representatives—employment for the youngest workers declined considerably after October 2022. For instance, employment for software developers aged 22–25 decreased by nearly 20% compared to its late 2022 peak by July 2025.

In stark contrast, employment for more experienced workers in the exact same exposed occupations has remained stable or continued to grow. This divergence suggests a substitution effect concentrated at the entry level ^[5].

3.2 Stagnation in Employment Growth

While overall employment continues to grow robustly, employment growth for young workers (22–25 year-olds) has been tepid or stagnant since late 2022 ^[7]. The disparity becomes evident when decomposing growth rates:

This divergence strongly suggests that declining employment in AI-exposed jobs is the key driver of the overall stagnation for the 22-25 age cohort ^[9].

3.3 The Automation vs. Augmentation Dichotomy

Labor market effects are heterogeneous, depending on whether AI substitutes for or complements human tasks ^[1]. Using the query data from Handa et al. ^[13] to distinguish automative from augmentative uses, a clear pattern emerges:

Entry-level employment declined significantly in applications of AI that primarily automate work but demonstrated muted changes or even growth in applications that are highly augmentative.

3.4 Robustness to Firm-Level Shocks

To eliminate the confounding influence of industry- or firm-level shocks (e.g., interest rate changes) that might correlate with sorting patterns, we employ a high-dimensional Poisson event study regression. This approach controls for firm-time effects (β_f,t) and firm-quintile effects (α_f,q).

This suggests the observed employment trends are intrinsic to the AI exposure dynamic, rather than being driven by differential shocks to firms.

3.5 Adjustment Mechanisms: Quantity vs. Price

A crucial distinction in the observed adjustment is its manifestation: it is visible in employment quantities rather than compensation prices. Analysis of annual base compensation reveals little significant difference in salary trends by age or AI exposure quintile.

This indicates that the labor market adjustment is, at least initially, occurring through quantity adjustments (job counts) rather than price adjustments (wages). This outcome is consistent with the hypothesis of wage stickiness in the short run ^[10].

4. Theoretical Implications: Codified Knowledge and Tacit Expertise

The question of why these effects disproportionately target entry-level workers (22–25) can be addressed by examining the differential nature of knowledge production.

We posit that AI is highly effective at replacing codified knowledge—the formal, explicit "book-learning" acquired in formal education ^[6]. This codified knowledge forms a larger proportion of the skills supplied by young workers.

Conversely, AI appears less capable of replacing tacit knowledge—the idiosyncratic tips, tricks, and specialized skills accumulated through practical, on-the-job experience or "learning-by-doing" ^[4]. Since older, more experienced workers supply a greater share of tacit knowledge, they may be less vulnerable to task replacement ^[5], leading to greater employment reallocation risk for the young, AI-exposed cohort.