Performance or Principle: Resistance to Artificial Intelligence in the U.S. Labor Market
Citation. Friis, S., & Riley, J. W. (2025). Performance or Principle: Resistance to Artificial Intelligence in the U.S. Labor Market. Harvard Business School Working Paper 26-017. https://doi.org/10.2139/ssrn.5560401
Summary. Friis and Riley distinguish performance-based resistance to AI automation from principle-based resistance that persists even when AI is described as better and cheaper than human labor. In a quota-matched U.S. survey covering 940 occupations and 23,570 occupation ratings, public support for full automation rises sharply when AI capability is improved, but a smaller set of care, therapy, and spiritual or moral-authority occupations remains morally protected. The paper turns AI labor-market resistance into a social-feasibility problem rather than a pure capability or cost problem.
Key claims
- In the current-AI condition, 30.0% of occupations receive majority support for full automation; when AI is described as outperforming humans at lower cost, 58.3% receive majority support.
- Augmentation is far more acceptable than replacement: respondents endorse AI assisting humans in 94.4% of occupations under current AI and 95.6% under advanced AI.
- Moral repugnance toward AI use is concentrated rather than generalized: 36.8% of occupations are classified as morally permissible, 51.3% as morally ambivalent, and 11.9% as morally repugnant.
- Moral repugnance and technical suitability are only weakly aligned (reported R² = 0.134), implying that capability/exposure estimates miss a distinct social-acceptability boundary for adoption and substitution.
- Occupations classified as morally repugnant for AI use represent an estimated 17.3–21.9 million U.S. workers, with a weighted estimate of 19.8 million workers or 13.2% of the workforce.
- Occupations with higher wages, fewer employees, higher proportions of women, and higher proportions of White workers are associated with higher moral repugnance scores, so social protection against AI automation may map unevenly onto existing labor-market inequalities.
Evidence & limitations
- The main study surveyed 2,357 U.S. adults recruited through Prolific with quotas for age, gender, ethnicity, and political affiliation; each participant rated 10 occupations, yielding 23,570 occupation ratings across 940 O*NET occupations.
- The design compares current AI to an advanced-AI scenario and measures moral repugnance toward AI use, so it separates performance-contingent resistance from principle-based resistance more directly than generic algorithmic-aversion studies.
- The unit of analysis is the occupation rather than the task; the authors justify this because moral boundaries often attach to socially recognized roles, but task-level variation inside morally contested occupations remains unresolved.
- The paper is a Harvard Business School working paper / SSRN preprint, not a peer-reviewed article; it measures public attitudes and social feasibility, not actual firm adoption, policy decisions, wages, hiring, or realized substitution.
- Extraction was high confidence from a 79-page PDF; no source-manipulation or model-directed prompt-injection language was found.