Barriers to AI Adoption: Image Concerns at Work
Citation. Almog, D. (2026). Barriers to AI Adoption: Image Concerns at Work (Job market paper, March 26 version).
Summary. Almog studies whether workers underuse AI because visible reliance on AI changes how they expect to be evaluated. In a field experiment with 450 U.S.-based Upwork workers completing an AI-assisted image-categorization task, making AI reliance visible to an HR evaluator reduced AI reliance and lowered performance. The paper argues that adoption barriers are not only technical or skill-based: the social meaning attached to using AI can make workers avoid useful assistance even when the evaluator is instructed to value accuracy on the same AI-assisted task.
Key claims
- Making AI reliance visible to an HR evaluator reduced workers’ AI recommendation following by 4.3 percentage points overall and 8.0 percentage points when the worker’s initial answer disagreed with the AI recommendation, relative to a 30.5% control-group reliance rate.
- Visible AI reliance lowered final task accuracy from 79.1% to 76.4%, a 2.7 percentage-point decline, even though workers spent about two seconds longer on initial choices and did not improve initial accuracy or selectivity about when to follow AI.
- The reduction persisted in a treatment that reassured evaluators about workers’ strong platform performance history, suggesting that image concerns were not easily mitigated by closing a simple information gap.
- Workers shifted sharply toward emphasizing confidence in their judgment when feedback would also disclose above- or below-average AI use; the share choosing confidence more than doubled relative to control, consistent with a concern that heavy AI reliance signals weak independent judgment.
- In a second-stage evaluator task, workers penalized AI reliance in others’ profiles: each additional percentage point of AI reliance reduced profile scores by 0.36 points, while each additional percentage point of accuracy increased scores by about one point.
Evidence & limitations
- Evidence comes from a preregistered field experiment on a large online labor market using a controlled image-categorization task with ground-truth labels and AI recommendations, plus a second-stage profile-evaluation task and platform-feedback elicitation.
- The setting gives strong experimental control and real labor-market incentives, but the focal task is short-cycle remote image categorization; generalization to richer professional work, team settings, and long-term organizational adoption remains an empirical question.
- The paper is a March 26, 2026 job-market-paper version, with no DOI or peer-reviewed publication record found during ingest; the PDF links to a public latest-version Google Drive file with a matching MD5 checksum.