Two employees read identical performance feedback โ same criticism, same framing, same closing line. One is told their supervisor wrote it; the other, that an AI generated it. On the item measuring how close they feel to that supervisor, the first scored 2.89 and the second 1.67 โ the disclosed source explained 24% of the variance (ฮทยฒ = 0.24). On whether they kept turning the criticism over afterwards, it explained nothing (p = 0.64) (Marรณt, Palcsรณ & Szabรณ, Frontiers in Psychology, 2026).
The message landed. The reporting line didn't โ and that is the AI performance review's real price.
That asymmetry is the part worth your attention before the next cycle opens, because this is not a 2027 question for mid-market teams. Among managers who already use AI to manage people, 91% use it to assess performance and 88% use it to draft performance improvement plans (ResumeBuilder.com, 2025). The drafting is happening. What has not been priced is what the disclosure does.
The experiment held the text constant and moved only the label
Marรณt, Palcsรณ and Szabรณ ran a vignette experiment with 192 employed adults (230 recruited; 18 excluded as unemployed, 20 for failing an attention check). Participants imagined receiving performance feedback from a supervisor. The feedback text was held completely identical across every condition. The single manipulation was the disclosed production workflow, in four variants: fully human-written, fully AI-generated, AI-generated then refined by a human, and human-written then refined by AI.
This design matters more than the sample size suggests. Most studies of AI in people management confound two things โ what the AI produced and the fact that an AI was involved. Holding the text constant isolates the second. Whatever moved here moved because of the label, not because the AI wrote something worse.
A great deal moved. The overall multivariate effect of disclosed source was significant (Pillai's V = 0.37, F(27, 546) = 2.86, p < 0.001), and eight of the nine outcomes shifted:
| Outcome | Human | AI | AI drafted, human refined | Human drafted, AI refined | ฮทยฒ |
|---|---|---|---|---|---|
| Psychological distance (closeness) | 2.89 | 1.67 | 2.15 | 2.54 | 0.24 |
| Relationship with the supervisor | 4.69 | 3.23 | 3.98 | 4.22 | 0.21 |
| Openness to ask for help | 5.31 | 3.43 | 4.57 | 4.82 | 0.17 |
| Organisational commitment | 4.57 | 3.46 | 4.00 | 4.38 | 0.16 |
| Willingness to correct the mistake | 6.02 | 4.43 | 5.18 | 5.56 | 0.14 |
| Work motivation | 4.79 | 3.66 | 4.30 | 4.68 | 0.13 |
| Expected performance | 5.19 | 4.29 | 4.68 | 4.82 | 0.10 |
| Affective response | 4.76 | 3.86 | 4.43 | 4.48 | 0.08 |
| Rumination | 3.64 | 3.45 | 3.20 | 3.48 | 0.01 |
Read the column order, not just the first two columns. The ranking is the same on every row.
Where the AI performance review actually costs you
Rank those effect sizes and the top four are psychological distance (0.24), relationship with the supervisor (0.21), openness to ask the supervisor for help (0.17) and organisational commitment (0.16). The informational outcomes sit lower: expected performance (0.10), affective response (0.08), rumination (0.01, not significant).
A performance conversation is supposed to do two jobs. Transmit an accurate signal about the work, and keep open the channel through which that employee brings you the next problem early. The disclosure tax falls almost entirely on the second job. Openness to ask for help dropped from 5.31 to 3.43 on a seven-point scale โ the largest absolute gap in the study.
That is the expensive one. A manager who has lost the "come to me when it's going wrong" channel has lost the only early-warning system a 50โ500 FTE operation has. You will not see it in the engagement survey as a feedback problem; you will see it as surprises arriving late.
Who writes the first draft matters more than who does the last pass
This is the practically useful finding, and it is the one the abstract understates. The two hybrid workflows are not interchangeable.
Human-written then AI-refined tracked close to fully human throughout: 4.68 against 4.79 on motivation, 4.38 against 4.57 on commitment, 4.82 against 5.31 on openness to ask for help. It was not significantly different from fully human feedback on any of the nine outcomes.
AI-generated then human-refined did worse, and more importantly it failed to separate itself from fully AI feedback. Against the fully AI condition, it reached significance on only three of nine outcomes โ openness to ask for help (4.57 vs 3.43, p = 0.002), relationship with the supervisor (3.98 vs 3.23, p = 0.003) and psychological distance (2.15 vs 1.67, p = 0.023). Human-written-then-AI-refined, by contrast, beat fully AI feedback on eight of nine.
So the sequence carries the effect, not the presence of a human somewhere in the chain. Having a manager edit an AI draft bought back a fraction of the relational cost. Having a manager write first and use AI to tighten the prose bought back nearly all of it.
Which is the exact inverse of how the tooling is being used. "Generate a review from these competency scores, then I'll edit it" is the default workflow in every HR suite that shipped an assistant in 2026. On this evidence, that is the worse of the two hybrids โ and it is the one your managers will pick, because it saves more of their time.
The counter-argument, and where it holds
The honest objection: this is a vignette. Participants reported anticipated reactions to a hypothetical scenario, not behaviour after a real review. The sample was a Hungarian convenience sample recruited online, skewed toward subordinate-level employees (66%), with a mean age of 32. There is no behavioural outcome and no follow-up. Do not treat 1.67 versus 2.89 as a forecast of your own numbers.
What survives the objection is the direction and the ordering, and both are reinforced by the study's own robustness check. Restricted to the 123 participants (64%) who correctly identified the disclosed source โ in other words, the people who actually registered the manipulation โ every effect ran the same way and most ran larger, with ฮทยฒ between 0.19 and 0.42. The effect is not an artefact of inattention. If anything, the full-sample figures are diluted by participants who did not notice the label.
"AI literacy" does not buy this back
The tempting internal read is that this is a technology-comfort problem that training solves. The study tested that directly. AI anxiety was measured as a covariate, and its multivariate effect was not significant (V = 0.08, F(9, 179) = 1.73, p = 0.085). At the univariate level it predicted exactly one thing โ rumination โ and nothing else.
Note what that means in combination with the main result. Rumination is the one outcome the disclosed source did not move. AI anxiety moved rumination and nothing else. The two effects are almost perfectly disjoint. Employees are not reacting out of generalised unease about AI. They are reacting to a specific inference: that the person responsible for evaluating them delegated the part of the job that was about them.
There is a second reason to stop treating this as a training problem. Only 16% of organisations are confident their managers can explain AI-related pay and work changes to employees at all (Korn Ferry, 2026, N = 5,512 organisations across 135 countries). The population you would be asking to deliver a careful explanation of the drafting workflow is, by its own leadership's assessment, not currently able to deliver it.
The customer-side disclosure result does not transfer
If you have been following the disclosure literature, you may have the opposite prior. A randomised field experiment with 1,337 customers reached found that disclosing an AI voice agent no longer reduces conversion relative to non-disclosure (Habel, Zawadzki & Freise, SSRN 7013259, 2026). Disclosure, in a commercial interaction, has become close to free.
It does not follow that it is free internally, and the reason is structural rather than cultural. A customer's relationship is with your company; there is no ongoing reporting line to damage, and they largely assume an agent anyway. An employee's relationship is with a named individual who holds their rating, their band and their progression. Disclosure in that setting is information about how that individual spent their attention.
The hiring literature points the same way. In algorithmic evaluation of interviews, transparency corrected candidate behaviour โ it reduced the stress and deceptive impression management that AI scoring induced โ but did nothing on the evaluator side, where humans anchored on AI scores and discounted their own judgement (Lakhiwal et al., Information Systems Research, 2026). Transparency fixes the person being told. It does not fix the delegation it reveals.
What to change before the next review cycle
Five decisions, none of which require new procurement.
- Set the sequence as policy, not preference. Manager writes the narrative first; AI may tighten language, structure or length afterwards. The reverse order is the workflow this study penalised hardest.
- Keep AI out of the relational sentences specifically. The evidence says the damage concentrates in closeness, trust and help-seeking โ the parts of a review that are about the relationship, not the work. Those paragraphs are the ones a manager must own in their own words.
- Decide your disclosure position before someone asks, and write it down. Do not run an undisclosed workflow on the theory that nobody will notice; 64% of participants correctly identified the source from the text alone, and the effects were larger among them.
- Fix the input, not the prose. Most managers reach for AI drafting because they are staring at a blank box with no evidence in front of them. Give them per-cycle competency evidence, 360 divergence already synthesised, and calibration flags, and the narrative becomes a short writing task rather than a generation task โ which is what Scovai's performance and succession layer is built to supply.
- Measure the channel, not the satisfaction score. Add one post-cycle item: "after this review, how likely are you to bring a problem to your manager early?" That is the variable this study says is moving, and it is not in most review surveys.
The budget question answers itself. The constraint on review quality in a 50โ500 FTE company is not writing capacity โ it is that managers write reviews from memory because no system hands them the evidence. AI drafting treats the symptom and charges you the reporting line to do it.
One decision for this quarter
The AI performance review is already in your managers' browsers whether or not it is in your policy. You are not choosing whether AI touches review text. You are choosing the order of operations โ and this study says that is the whole ballgame: human first and AI second was statistically indistinguishable from fully human on all nine outcomes, while AI first and human second was indistinguishable from fully AI on six of nine.
Before the cycle opens, write one line into your review guidance specifying who produces the first draft. It costs nothing, it is enforceable, and it is the only variable here that you control outright.
