A theory-based AI automation exposure index: Applying Moravec's Paradox to the US labor market
Jacob Schaal
TL;DR
The paper develops a theory-based AI automation exposure index anchored in Moravec's Paradox, evaluating 19,000 O*NET tasks across four dimensions (Performance Variance, Tacit Knowledge, Data Abundance, Algorithmic Gap) and weighting core tasks more heavily. Using OEWS wage and employment data and AI-annotation validation against Eloundou et al. (2024), it finds that management, STEM, and sciences occupations face the highest exposure while maintenance, agriculture, and construction face the lowest, with a positive wage-exposure relationship and a distinct tacit-knowledge wage pattern. The index demonstrates robustness across multiple AI models (GPT-4, Claude, Gemini) and shows a strong correlation (≈0.72) with the Eloundou benchmark, suggesting a paradigm shift away from capability-based indices. The results imply that AI exposure could compress wages for high-skill cognitive roles while sparing many manual jobs, and they call for further empirical testing as AI capabilities evolve. Overall, the study offers a principled, theory-grounded framework for measuring automatability that remains relevant as AI technology advances and provides actionable insights for policy and labor-market research.
Abstract
This paper develops a theory-driven automation exposure index based on Moravec's Paradox. Scoring 19,000 O*NET tasks on performance variance, tacit knowledge, data abundance, and algorithmic gaps reveals that management, STEM, and sciences occupations show the highest exposure. In contrast, maintenance, agriculture, and construction show the lowest. The positive relationship between wages and exposure challenges the notion of skill-biased technological change if AI substitutes for workers. At the same time, tacit knowledge exhibits a positive relationship with wages consistent with seniority-biased technological change. This index identifies fundamental automatability rather than current capabilities, while also validating the AI annotation method pioneered by Eloundou et al. (2024) with a correlation of 0.72. The non-positive relationship with pre-LLM indices suggests a paradigm shift in automation patterns.