The risk in AI-written feedback isn't that it's wrong. It's that no one believes you meant it.

It's nine at night and you're writing a performance review for someone who has been sliding for two quarters. You know what you want to say, but you don't trust yourself to say it without sounding harsh, so you paste your notes into a chatbot and ask it to clean up the tone. The words that come back are better than yours were. You send them. As far as you're concerned, nothing has changed — the judgment is still yours, the AI just helped you phrase it.

A new study says your employee may not experience it that way, even when the sentences are identical to the ones you would have written.

The paper is "Source matters, AI anxiety less so: comparing employee reactions to human, AI, and hybrid performance feedback and the limited role of AI anxiety," by Janka Laura Marót, Tamara Palcsó, and Zsolt Péter Szabó, of Eötvös Loránd University and Corvinus University of Budapest, published in Frontiers in Psychology on August 17. It's a vignette experiment, not a workplace field study: 192 Hungarian employees, recruited through social media and the researchers' own networks, were randomly told that a piece of performance feedback had been written fully by a human, fully by AI, by AI and then edited by a human, or by a human and then edited by AI. The feedback text itself was held identical across all four conditions — only the disclosed authorship changed. Participants then rated how they felt about it: their emotional reaction, their motivation, their willingness to ask for help, how close or distant they felt from their supervisor, how likely they were to correct the mistake being flagged.

The label alone moved the numbers. Feedback disclosed as fully AI-generated produced worse reactions than the identical text disclosed as fully human-written on eight of the nine measures — less willingness to ask for help, more perceived distance from the supervisor, lower motivation, weaker commitment. The one measure that did not budge was rumination: people chewed the feedback over just as much either way. Feedback that AI had drafted and a human had then edited fell in between. But feedback a human had written and AI had only polished for wording looked statistically the same as feedback that was entirely human, start to finish.

These findings support the disclosure effect: once the source of feedback is known, it alters its effect. — from the Discussion section, Marót, Palcsó & Szabó, Frontiers in Psychology, read in full at frontiersin.org

The authors are careful about what they can claim. This is a hypothetical scenario, not an actual manager delivering actual feedback to someone whose job depends on it — participants reported an anticipated reaction to a disclosure, not behavior that followed real feedback from a real boss. Some differences were large — up to a full standard deviation and a half by their own measure — and the authors flag that a gap that size may partly reflect the novelty of seeing the label at all, and could shrink in a real workplace. The outcomes were single survey items, not validated multi-item scales, and the sample was a Hungarian convenience sample recruited online, not a representative slice of any workforce, including yours.

Here's why that gap between "hypothetical" and "your actual Tuesday" matters more to you than to almost anyone else reading this. In a company of five thousand employees, a review runs through a manager, then HR, then a template — the source is already diffuse before the employee sees it. In a company you run yourself, you are frequently the entire chain. Your employee has spent years learning what your name on a piece of feedback means: that you looked at their work, formed a view, and were willing to put it in writing. That's the value of feedback from you rather than a form. The study's finding is that once someone believes a machine did the looking, that value doesn't survive intact — even when the sentences are exactly the ones you'd have chosen.

The part the research doesn't say, but that I'll say plainly: this almost certainly isn't limited to formal reviews. If your team ever learns — and in a company like yours, they eventually do — that the thoughtful Slack message or the personal-sounding email came out of a prompt, they will start reading everything you send through that lens, including what you actually wrote yourself. The study didn't test that spillover. I'm inferring it from the mechanism: it's not the words that carry weight, it's the belief that you stood behind them first.

One thing you can do this week: keep using AI to help you write, if it helps you write better — the study found no penalty for that when a human did the thinking and AI only refined the wording. The difference was in outsourcing the judgment itself, not the sentence structure. So don't let a tool generate the assessment of someone's performance from your notes and a prompt; write the judgment, then let it help with tone. And if anyone ever asks whether you used AI on something with your name on it, answer as if the honest answer were the plan all along, because it should be.

If the harder problem is that you don't fully trust your own read on people under you — who's actually struggling, who's coasting, who's ready for more — that's not a wording problem, and no tool fixes it. The Leadership Profile is built for exactly that gap.