Do people fear judgment when they use AI at work?

Reif, Larrick, and Soll argue that using AI at work carries a social cost that users both foresee and, in part, pay. In the excerpt's results, participants in the AI Tool condition expected others to judge them less competent (M = 4.72, SD = 1.46, against M = 5.45, SD = 1.22 in the Non-AI Tool condition) and less diligent (M = 4.66, SD = 1.44, against M = 5.25, SD = 1.28). The 95% intervals for both differences exclude zero. The same participants reported less willingness to disclose the tool to managers (4.91 against 5.25, P = 0.016) and to colleagues (4.85 against 5.17, P = 0.026). The introduction goes further: it says observers "perceive people who get help from AI as lazier, less competent, and less diligent" (Study 2), that managers who do not use AI may act on those assumptions in a hiring task (Study 3), and that perceived laziness mediates assessments of poor task fit (Study 4). The excerpt does not show those later results.

The mechanism is attribution applied to help. Assistance "creates attributional ambiguity," so observers must weigh "how much credit is due to the person versus the assistance," and the authors argue that accepting help can read as "a signal that the recipient is not willing or able to perform the task themselves." AI should be read more harshly than earlier tools because it "may be perceived as more agentic," learning from experience and operating "more autonomously." This is theory in the excerpt. Its data are the expectation and rating results above.

The social cost sits beside the performance gains the library already tracks. The authors frame it as a dilemma: AI "can enhance human performance on a variety of tasks," yet its use may cost the user reputation. Can AI narrow the education performance gap? measures the performance side. The excerpt's measures are expectations and ratings, not performance, so Can self-ratings replace objective performance scores for AI competence? supplies the caution: these results describe what people believe others think, not what AI does to their work. The disclosure gap is a measured case of the provenance management described in Which workplace cues survive AI mediation and which disappear?, though the excerpt measures stated willingness, not actual disclosure.

The excerpt leaves several things open: the scale anchors, the wording of the vignettes and of the AI versus non-AI manipulation, the sample size of each study, and any effect size beyond the intervals. It describes a rating study of 1,215 online participants who read about an employee but reports none of its results, so the claim that observers penalize AI users rests on the authors' summary. The disclosure result is stated willingness with no behavioral measure. At the strength the evidence allows, the anticipated penalty and the lower stated disclosure are measured, which makes a social barrier to adoption plausible. That observers actually apply the penalty is the authors' claim, resting on studies this excerpt does not show.