Truace
Subscribe
70/100

Moderate evidence

Cite
More
Export CSVExport JSONEmailLinkedInX

Feedback written by AI alone scored lower with 192 employees; AI-polished human drafts held up

Study title: Source matters, AI anxiety less so: comparing employee reactions to human, AI, and hybrid performance feedback and the limited role of AI anxiety

Evidence type
Randomized controlled trial
Sample
n = 192
Sources
1 sourcePeer-reviewed
Published
2 min read
Truvace record
TRV-2026-2532cite or check it

Study title: Source matters, AI anxiety less so: comparing employee reactions to human, AI, and hybrid performance feedback and the limited role of AI anxiety

A vignette experiment with 192 employed Hungarian adults tested how people react to the same performance-review text when told it was written in different ways. Those told the feedback was fully AI-generated gave lower ratings than those told a human wrote it, for example 3.23 versus 4.69 on the effect on the supervisor relationship and 3.43 versus 5.31 on willingness to ask the supervisor for help.

Read the study abstract

Introduction: Large language models (LLMs) are increasingly integrated into human resource management practices and are becoming capable of supporting managerial tasks such as performance feedback. Although these technologies may improve the efficiency and quality of feedback delivery, it remains unclear whether knowledge of AI involvement influences recipients' reactions to it. Methods: = 192) examining employees' reactions to a hypothetical performance feedback scenario. Participants were randomly assigned to…

61Gain side

How the two sides scoreClose call

57Problem side

The gain

In a 192-employee vignette experiment, feedback a human wrote and then refined with AI did not differ significantly from purely human feedback and beat fully AI-generated feedback on all outcomes except rumination.

The problem

In a 192-employee vignette experiment, identical feedback disclosed as fully AI-generated scored lower than human-written feedback on work motivation (3.66 vs 4.79 on a 7-point scale) and on the supervisor relationship (3.23 vs 4.69).

A fixed formula sets both scores from the evidence and its sources. Reader votes don’t move them.

What the scores mean

G is the gain side, out of 100. It goes up when the gain is specific, when it measurably helped people and when the sources behind it are strong.

P is the problem side, out of 100. It goes up when the problem is specific, when it measurably hurt people and when the sources behind it are strong.

In brief

Hybrid workflows landed in between. Feedback a human wrote and then refined with AI was statistically indistinguishable from purely human feedback. AI-generated feedback reviewed by a human beat fully AI-generated feedback only on relational measures: openness to ask for help, supervisor relationship and psychological distance.

AI anxiety, measured with a 21-item scale, mattered little. It was linked only to more rumination after feedback and did not change how people reacted to the different sources. The authors caution that the effects came from a hypothetical scenario and may overstate what would happen with real managers.

Main points

  1. Participants were randomly assigned to four disclosed workflows: fully human-written (42), fully AI-generated (56), AI-generated and human-refined (44), and human-written and AI-refined (50). The feedback text was identical.
  2. Fully AI-generated feedback scored lower than human feedback on willingness to correct the mistake (4.43 vs 6.02) and on organizational commitment (3.46 vs 4.57).
  3. Expected performance was rated 4.29 for AI-generated feedback versus 5.19 for human feedback.
  4. Rumination did not differ by feedback source.
  5. AI-generated feedback refined by a human beat fully AI-generated feedback on openness to ask for help (4.57 vs 3.43), supervisor relationship (3.98 vs 3.23) and psychological distance.
  6. The two hybrid workflows did not differ significantly from each other.
  7. AI anxiety was significantly associated only with rumination (partial eta squared 0.07); its overall multivariate effect was not significant (p = 0.085).
  8. Among 123 participants who correctly identified the disclosed source, effects were larger, with eta squared from 0.19 to 0.42.

The full story

Data were collected in 2025 through an online Qualtrics survey. Of 230 Hungarian participants recruited via social media and professional networks, 18 unemployed people and 20 who failed an attention check were excluded, leaving 192 employees (54.69% female, ages 19 to 65, mean 32.34). Two-thirds worked at subordinate level.

Participants imagined receiving a review excerpt with both praise (meeting deadlines) and a development point (communication in joint discussions). They rated eight single-item outcomes on 7-point scales plus a four-item psychological distance scale. A multivariate analysis showed a significant overall effect of source (Pillai's V = 0.37, p < 0.001), with effect sizes up to eta squared 0.24 for psychological distance.

Only 64.06% (123 participants) correctly identified the disclosed source in the manipulation check. Restricting the analysis to them reproduced the main pattern with larger effects. AI anxiety scores did not differ across conditions.

The authors conclude that managers should not replace their own judgment with AI and that using AI to refine human-written feedback looks viable, though they note that efficiency and quality benefits were not measured in this study.

What this doesn’t fix

The study used a hypothetical vignette with 192 Hungarian convenience-sample participants and single-item measures; the authors say the largest differences (up to d = 1.46) may reflect label salience or demand characteristics and overstate real workplace effects.

Sources

1 on file
  1. Frontiers in Psychologydoi:10.3389/fpsyg.2026.1924487
    Peer-reviewed2026-08-17

Talk it over

Four questions to start onThis record found a gain and a problem in the same place. Which matters more to you, and who might answer differently?

Start with a question

Feedback written by AI alone scored lower with 192 employees; AI-polished human drafts held uphttps://truvace.com/item/source-matters-ai-anxiety-less-so-comparing-employee-reactions-to-human-ai-and-h

  1. This record found a gain and a problem in the same place. Which matters more to you, and who might answer differently?

    The gain In a 192-employee vignette experiment, feedback a human wrote and then refined with AI did not differ significantly from purely human feedback and beat fully AI-generated feedback on all outcomes except rumination.

    The problem In a 192-employee vignette experiment, identical feedback disclosed as fully AI-generated scored lower than human-written feedback on work motivation (3.66 vs 4.79 on a 7-point scale) and on the supervisor relationship (3.23 vs 4.69).

    Answer in the thread

  2. How do we know? What can this kind of evidence show, and what can’t it?

    The evidence This rests on a randomized controlled trial with a sample of 192, from Frontiers in Psychology.

    Answer in the thread

  3. What does it leave open? What else would you want to know before acting on it?

    The record’s own caveat The study used a hypothetical vignette with 192 Hungarian convenience-sample participants and single-item measures; the authors say the largest differences (up to d = 1.46) may reflect label salience or demand characteristics and overstate real workplace…

    Answer in the thread

  4. Have you seen AI used this way where you study, work or live? Did it go the same way?

    Answer in the thread

The questions are the same on every record. The details under them are taken from this record as written.

What readers say

More from Labor