The transparency dilemma: How AI disclosure erodes trust

Synthesis: Thirteen preregistered experiments, most of them run with United States online panelists and pooled in a within-paper meta-analysis of 4,093 observations, test whether actors who disclose their use of generative AI are trusted less than those who stay silent. They are. The penalty appears across writing, analytical, and design tasks and across individual and organizational actors; it survives six different disclosure framings, prior knowledge that AI was used, and mandatory-versus-voluntary disclosure regimes, and it is smaller than the damage done when a third party exposes undisclosed AI use. Drawing on micro-institutional theory, the authors argue the mechanism is perceived legitimacy: disclosure signals a deviation from the expectation that work reflects human judgment, and lower legitimacy in turn lowers Trust. A human who discloses AI use is trusted less even than an autonomous AI agent doing the same task. The moderators are modest — favorable technology attitudes and beliefs that AI is accurate attenuate but do not remove the effect, and neither AI familiarity nor one's own AI use does. The samples are working adults and university students in non-education settings; professors appear as one actor category among several, not as an education finding.

Key Findings

1. Across all 13 experiments, actors who disclosed AI use were trusted less than those who did not; the within-paper meta-analysis of 4,093 observations put the pooled effect at θ = 0.81, z = 10.52, 95 % CI [0.66, 0.96], p < 0.001.
2. Study 1 (195 undergraduate business students) found students trusted a professor less when she disclosed AI grading (M = 2.48, SD = 0.56) than when a human teaching assistant was disclosed (M = 2.87) or nothing was disclosed (M = 2.96), F(2, 192) = 15.92, p < 0.001, η2 = 0.14.
3. The penalty held when the trustor held more power: Study 2 (85 supervisors with hiring experience) found M = 2.91 for AI disclosure versus M = 4.40 for no disclosure, d = 1.42; Study 3 (345 investors) found M = 4.18 versus M = 5.08 for an investment fund's advertisement, d = 0.69.
4. Legitimacy carried the effect. In Study 6 (427 panel members) AI disclosure reduced perceived legitimacy and legitimacy in turn raised trust, giving an indirect effect of ab = −0.56, SE = 0.06, 95 % CI [−0.68, −0.43].
5. Study 10 (753 panel members) showed the disclosure penalty exceeded algorithm aversion: a human who disclosed AI use was trusted least (M = 4.94) compared with an autonomous AI agent (M = 5.61) and a non-disclosing human (M = 5.72), F(2, 750) = 46.07, p < 0.001.
6. Study 13 (195 panel members) found third-party exposure worse than self-disclosure: exposure M = 2.49, disclosure M = 3.15, no disclosure M = 4.02, F(2, 192) = 37.41, p < 0.001.
7. The moderators were weak: technology attitude (b = 0.19, z = 5.38) and perceived AI accuracy (b = 0.17, z = 5.35) attenuated the negative effect, while AI familiarity did not (b = 0.00, p = 0.89).

Who the studies actually sampled

This is an organizational-behavior and marketing paper, and its evidence base reflects that. Study 1 drew 195 undergraduate business students in a behavioral lab; Study 8 ran in a classroom with 93 students; every other study drew United States online panelists from the CloudResearch Connect platform. Among those panelists the authors deliberately recruited professional cohorts: Study 2 sampled 85 supervisors with hiring experience, Study 3 sampled 345 investors, and Study 4 sampled 90 legal analysts. Studies 5 through 7 and 9 through 13 drew general panel members (597, 427, 426, 518, 753, 1,048, 348, and 195 respectively), and the within-paper meta-analysis pooled 4,093 observations. In each experiment participants took the role of a trustor judging another actor, individual or firm. The education connection is narrow: Study 1 has students rating a professor's disclosed AI grading, and professors are simply one actor category alongside supervisors, analysts, founders, and funds. No study measured learning, teaching, or classroom outcomes.

Legitimacy as the argued mechanism

The authors ground the effect in micro-institutional theory, proposing that disclosing AI use is read as a departure from the taken-for-granted expectation that a person's work reflects human expertise and judgment, which lowers perceived legitimacy and thereby erodes trust. Studies 6 through 8 test that account three ways. Study 6 (427 panel members) measured legitimacy and found the indirect path ab = −0.56, SE = 0.06, 95 % CI [−0.68, −0.43]. Study 7 (426 panel members) manipulated collective validity — showing some participants a Boston Consulting Group article stating that generative AI can improve performance by 40 % on ideation and content-creation tasks — and found the disclosure penalty was stronger in the control condition (d = 1.12) than in the collective-validity-prime condition (d = 0.72). Study 8 (93 students) manipulated legitimacy directly and used a behavioral measure: students invested less when the fund was framed as illegitimate (M = $3.39, SD = 1.86) than as legitimate (M = $4.09, SD = 1.20), t(91) = 2.14, p = 0.03, d = 0.44.

Robustness across framings, knowledge, and regimes

Later studies probe how far the penalty travels. Study 9 (518 panel members) tested six disclosure framings — general, human-reviewed, proofreading-only, intent, error warning, and a transparency pledge — and every framing produced lower trust than no disclosure (all ps < 0.02). Study 10 (753 panel members) contrasted a disclosing human with an autonomous AI agent and a silent human, and the disclosing human was trusted least (F(2, 750) = 46.07, p < 0.001), a result the authors read as going beyond algorithm aversion. Study 11 (1,048 panel members) varied whether the evaluator already knew AI had been used; the effect weakened but persisted (d = 0.40 when AI use was known versus d = 0.60 in the control). Study 12 (348 panel members) varied the disclosure regime and found the penalty in both the mandatory (d = 0.74) and voluntary (d = 0.64) conditions. Study 5 (597 panel members) extended the finding to a mundane scheduling email (t(595) = 4.43, p < 0.001, d = 0.36).

What this means for practice

• Instructors and administrators. Study 1 is the only education-adjacent result, and it is a trust judgment, not a learning outcome: students trusted a professor less when her grading was disclosed as AI-performed. Treat it as evidence about how a disclosure norm may land, not as evidence about teaching or learning.
• Organizations setting disclosure policy. Because the penalty held under both mandatory and voluntary regimes (Study 12) and across six framings (Study 9), a policy that simply requires employees to disclose AI use does not avoid the trust cost the paper documents; the authors suggest building environments where AI use is seen as collectively valid (Study 7).
• Managers and communicators. Third-party exposure was worse than self-disclosure (Study 13), so the choice is not disclosure versus silence but controlled disclosure versus uncontrolled exposure; the authors also advise against treating AI as "just a tool," which Study 9 found did not protect trust.
• Researchers and evaluators. The mechanism is argued rather than fully isolated — legitimacy is one path (Studies 6–8) and other accounts remain untested — and the moderators that did appear, technology attitude and perceived AI accuracy, attenuated the effect without removing it.

Limitations

• The samples do not support an education claim. Study 1 is the only one where students judge a professor, and Study 8's student sample made a financial investment decision; the remaining studies used professional or general adult panels. No outcome is about learning, instruction, or assessment validity.
• Most studies examine trust in previously unknown trustees — what the authors call swift trust — so whether the effect extends to established working relationships is untested, as the paper itself notes.
• The tasks studied are ones that traditionally required considerable manual effort; the authors flag that generalizability to highly automated settings is unknown.
• The paper does not state its data-collection dates. It reports receipt on 20 June 2024 and acceptance on 13 March 2025, and its writing-process statement records that the experimental stimuli were generated with ChatGPT-4 and Dall-E (OpenAI, 2023); the named tools, including Superhuman (Study 5), GrowthBar (Study 11), and Dall-E (Study 12), are tied to that window and have since changed.
• Legitimacy is presented as the mechanism, but the authors acknowledge other mechanisms may operate in parallel or in series, and they do not test alternatives such as ability–benevolence–integrity or warmth–competence accounts.
• The within-paper meta-analysis covers a subset of the data: it excludes Study 8 and the non-disclosure conditions of some studies, and it focuses on perceptual trust, so the pooled estimate θ = 0.81 describes that narrower contrast.

Citation

Schilke, O., & Reimann, M. (2025). The transparency dilemma: How AI disclosure erodes trust. Organizational Behavior and Human Decision Processes, 188, 104405.