Employees evaluate identical performance reviews less favorably when told they were produced by AI: fully AI-generated feedback lowers affective, relational and motivational responses compared with human-written feedback, while hybrid human–AI workflows fall in between and AI-related anxiety plays only a small role.

Main Finding

When employees read identical performance feedback, the disclosed source matters: feedback labeled as fully AI-generated produced significantly worse affective, motivational, relational, and performance-related reactions than feedback labeled as fully human-written. Hybrid workflows generally fell between these extremes; notably, feedback described as human-written then refined by AI produced responses comparable to fully human feedback. Individual differences in AI anxiety had only limited influence on these effects.

Key Points

• Experimental design held feedback content constant and manipulated only the disclosed production workflow: fully human (H), fully AI (AI), AI-generated then human-refined (AIxH), and human-written then AI-refined (HxAI).
• Fully AI-generated disclosure reduced positive affect, willingness to correct mistakes, work motivation, perceived relationship quality with the supervisor, expected performance, and organizational commitment; it also increased psychological distance.
• Hybrid conditions tended to be intermediate; HxAI (human-origin, AI-refined) was closest to purely human feedback in recipient reactions.
• AI anxiety (measured with the 21-item AI Anxiety Scale, α = 0.94) showed only limited covariate effects and did not explain the main disclosure differences.
• Manipulation checks were effective; psychological distance scale reliability α = 0.89.
• Limitations: vignette-based (hypothetical, immediate reactions), convenience sample of Hungarian employees (N = 192, Mage ≈ 32), many outcomes measured with single items.

Data & Methods

• Sample: 192 employed participants (convenience sample recruited in Hungary; after exclusions for unemployment and attention-check failures).
• Experimental manipulation: random assignment to one of four disclosed feedback-production workflows (H n=42; AI n=56; AIxH n=44; HxAI n=50). All participants read the same feedback excerpt (mixed positive and developmental content).
• Outcomes: affective response, rumination, willingness to correct mistakes, openness to seek help from supervisor, organizational commitment, relationship with supervisor, expected performance, work motivation (most single-item 7-point scales), plus a 4-item psychological distance scale (5-point scale).
• Individual difference: AI Anxiety Scale (AIAS; 21 items, 7-point; used as covariate).
• Analysis: manipulation check (chi-square), multivariate analysis of variance (MANOVA) across outcomes with AI anxiety as covariate; post hoc pairwise comparisons using Bonferroni-adjusted estimated marginal means.
• Key psychometrics: psychological distance α = 0.89; AIAS total α = 0.94.

Implications for AI Economics

• Adoption trade-offs: Pure automation of managerial feedback (fully AI-generated) may reduce direct labor costs and increase throughput, but can produce negative relational and motivational externalities that undermine employee performance, commitment, and potentially long-run productivity. Economic models of AI adoption in HR should internalize these relational costs.
• Value of hybrid workflows: Human-in-the-loop approaches, especially those framed as human-originated and AI-assisted (HxAI), preserve relational capital while retaining AI efficiency/quality gains. Cost–benefit analyses should model hybrid architectures as distinct adoption options with different payoff and risk profiles.
• Signaling and disclosure effects: The mere disclosure of AI involvement functions as a signal that changes perceived credibility, warmth, and psychological proximity. Economists modeling firm announcements or transparency policies must account for signaling effects that can alter worker responses independent of objective quality improvements.
• Limited role for heterogeneity in AI anxiety: Since AI anxiety had limited explanatory power, workforce-level acceptance may hinge more on process design and communicated human involvement than on screening for individual-level AI attitudes—shifting managerial focus from employee selection to workflow design and messaging.
• Key outcomes to include in evaluations: Beyond immediate efficiency gains, empirical economic assessments should measure turnover, help-seeking behavior, learning/adaptation rates, morale, and long-term productivity to capture downstream effects of feedback-source choices.
• Policy and governance implications: Regulatory or governance guidance requiring disclosure of algorithmic use may have unintended costs by lowering employee receptivity; policymakers should weigh transparency benefits against potential relational harms and consider recommending best practices (e.g., emphasize human oversight, train managers in disclosure framing).
• Research priorities for informed economic modeling: field experiments linking disclosed source, real feedback interactions, and downstream performance/turnover; cross-cultural replication; heterogeneous worker preferences; and dynamic models of adoption where reputational and morale externalities evolve over time.