Trust fell in 13 experiments on AI disclosure: say it first anyway
The question comes up at the coffee break in most AI trainings I run, and usually in a lower voice: should I tell my manager I used a model for this? Three research teams have now measured both sides of that choice. Saying it costs you something. Hiding it costs you more when it surfaces.
Observers rate AI users as lazier.
Jessica Reif, Richard Larrick and Jack Soll ran four preregistered experiments with 4,439 participants. People who received help from an AI tool expected to be judged lazier, less competent and less diligent than people who received the same help from a colleague. Observers then delivered the judgment the first group had predicted 1.
The third experiment put 1,718 participants in a manager's seat, choosing between a candidate who used AI daily and one who never did. Managers who rarely used AI themselves preferred the candidate who did not use it. Managers who used AI often preferred the daily user 1. One CV reads as resourceful or as lazy depending on who opens it.
Saying it out loud moves the trust.
Oliver Schilke and Martin Reimann ran 13 experiments across communication and creative tasks, with actors who were supervisors, professors, analysts and investment funds. Actors who disclosed their AI use were trusted less than actors who stayed quiet. Perceived legitimacy carried the effect, isolated in their studies 6 to 8 2.
Three of their results decide the practical question. Disclosure lowered trust even when the evaluator already knew AI had been involved, so the drop comes from the act of saying it rather than from the news itself. It lowered trust whether the disclosure was voluntary or required by a rule. And evaluators with favorable attitudes toward technology applied a smaller penalty without dropping it 2.
Their studies 9 to 13 tested the alternative. Trust after voluntary disclosure stayed higher than trust after a third party exposed the same AI use 2. Getting caught is the worse branch.
Readers grade the byline before the writing.
Yunhao Zhang and Renée Gosline had people rate persuasive ads, half written by professional creators at a consulting firm and half by GPT-4, each capped at 100 words. Under blind evaluation the AI content scored higher on quality than the human experts' work. Revealing the author narrowed the gap by lifting the ratings of the human work, and the gap never reversed 3. Readers paid for the byline.
Chiara Longoni, Andrey Fradkin, Luca Cian and Gordon Pennycook found the same reflex in news. Across two preregistered experiments with 4,034 US participants, headlines labeled as AI-generated were rated less accurate than identical headlines labeled human-written, for true headlines and false ones alike. The pattern survived separate and joint evaluation, varied news items and respondent characteristics 4.
Jess Hohenstein's team at Cornell watched it play out inside conversations. In two randomized experiments, AI smart replies made exchanges faster and shifted the language more positive. Participants whose partners suspected them of using smart replies were evaluated more negatively than partners believed to be typing, and a partner seen as using more AI suggestions was rated less cooperative 5.
The protocol.
1. Disclose before anyone can expose you. Voluntary disclosure beat third-party exposure on trust in Schilke and Reimann's final five studies 2. Treat a disclosure as a cost you choose rather than a risk you carry, and pay it in the same message that delivers the work.
2. Describe the division of labor. "Used AI" leaves the reader to imagine the laziest version. Write the sentence that fixes the picture: "I set the structure, the model drafted sections 2 and 3, I checked the four figures against the filings and rewrote the conclusion." Legitimacy is what carries the trust effect 2, and a verified step is the legitimate part.
3. Collect the verdict before you name the process. Send the draft, ask what lands and what breaks, then add the provenance line. The label changes the rating of identical content 34, so take the unlabeled read first.
4. Read the evaluator's own habits. A manager who uses AI daily prefers candidates who do 1. In a room that has adopted the tools, lead with the workflow. With a reviewer who avoids them, lead with what you verified by hand.
5. Write the human messages yourself. Condolences, feedback to a direct report, an apology, a reference. Suspicion of smart replies lowered perceived cooperation and affiliation 5, and the whole point of those messages is the time you spent.
6. Put a tools line at the bottom of recurring deliverables. A fixed footer on a weekly report ("Drafted with Claude, figures checked against the source system, reviewed by me") removes the per-message decision. The trust cost does not vanish under a standing rule 2, and the exposure risk does.
7. Run the numbers on one deliverable this month. Send two versions of a short report to two comparable reviewers, one with the provenance line and one without, and ask each for a quality rating out of 10. Two readers will not settle anything statistically. They will tell you whether your audience looks like Reif's AI-using managers or the other kind.
Take the document you are sending this week. Add two sentences at the end: which tool touched which section, and what you checked before sending. Then ask one colleague whether those two sentences changed how they read the rest.
Sources.
- Reif, J., Larrick, R. P., Soll, J. B. (2025). Evidence of a social evaluation penalty for using AI. Proceedings of the National Academy of Sciences, 122, e2426766122. doi:10.1073/pnas.2426766122
- Schilke, O., Reimann, M. (2025). The transparency dilemma: How AI disclosure erodes trust. Organizational Behavior and Human Decision Processes, 188, 104405. doi:10.1016/j.obhdp.2025.104405
- Zhang, Y., Gosline, R. (2023). Human favoritism, not AI aversion: People's perceptions (and bias) toward generative AI, human experts, and human-GAI collaboration in persuasive content generation. Judgment and Decision Making, 18, e41. doi:10.1017/jdm.2023.37
- Longoni, C., Fradkin, A., Cian, L., Pennycook, G. (2022). News from Generative Artificial Intelligence Is Believed Less. FAccT '22: Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, 97-106. doi:10.1145/3531146.3533077
- Hohenstein, J., Kizilcec, R. F., DiFranzo, D., Aghajari, Z., Mieczkowski, H., Levy, K., Naaman, M., Hancock, J. T., Jung, M. F. (2023). Artificial intelligence in communication impacts language and social relationships. Scientific Reports, 13. doi:10.1038/s41598-023-30938-9