Did ChatGPT Reprice AI-Exposed Labor? Substitution, Complementarity, and the Sign of AI Exposure

Abstract

Generative artificial intelligence can now perform work such as writing, coding, and tutoring, raising a question in labor economics: does this new technology displace workers or make them more productive? Because asset prices incorporate expectations within days, returns around a technology shock reveal what investors believe. Existing exposure measures rank industries by how much their tasks overlap with AI capabilities, but say nothing about direction: they establish that AI is relevant to an industry's work without indicating whether it is a threat or a benefit. We map the Felten–Raj–Seamans AI Industry Exposure (AIIE) index to firms by 4-digit NAICS, compute cumulative abnormal returns (CARs) for 3,056 U.S. firms around ChatGPT (November 30, 2022) and GPT-4 (March 14, 2023), and classify industries by whether AI substitutes for or complements their output. A portfolio long the most and short the least exposed industries earned +1.3% over [0,+1] (t=3.96) and +2.7% over [0,+10] (t=3.46). Cross-sectionally, a one-standard-deviation increase in AIIE predicts a ten-day CAR 0.71 percentage points higher (t=2.49, industry-clustered), rising to 0.97pp among non-financials, with a null pre-event placebo. This average, however, masks a sign reversal: AI-supplier industries gained 0.6%, while substitution industries whose product is cognitive labor (staffing, education, publishing, legal, consulting) lost 2.8% (t=-2.32). Firm-level 10-K AI-mention intensity from 650 EDGAR filings corroborates the industry proxy (ρ=0.25). Markets priced generative AI within days as a cost saver for most firms, but as a threat wherever AI can produce what the firm sells.