Truace
Subscribe
70/100

Moderate evidence

Cite
More
Export CSVExport JSONEmailLinkedInX

Telling people AI made it lowers their trust, a 21-experiment analysis finds

Study title: When AI says who made it: A meta-analysis of AI disclosure effects on consumer responses

Evidence type
Systematic review or meta-analysis
Sample
n = 13,582
Sources
1 sourcePeer-reviewed
Published
2 min read
Truvace record
TRV-2026-2705cite or check it

Study title: When AI says who made it: A meta-analysis of AI disclosure effects on consumer responses

A meta-analysis pooled 21 experimental effect sizes from studies involving 13,582 people to test what happens when content is labeled as AI-made. Disclosure lowered trust, attitudes and behavioral intentions, with a pooled effect of Hedges' g = −0.440 (95% CI −0.532 to −0.348), and all 21 effect sizes went in the negative direction.

Read the study abstract

Artificial intelligence (AI) is increasingly embedded in digital communications, yet the behavioral consequences of disclosing AI authorship remain inconsistent across studies. This meta-analysis synthesizes 21 experimental effect sizes (N = 13,582) to estimate the effect of AI disclosure on human behavioral responses—trust, attitudes, and behavioral intentions—and to identify the conditions under which it is strongest. Using a DerSimonian–Laird random-effects model, the pooled effect is significant and negative…

When AI says who made it: A meta-analysis of AI disclosure effects on consumer responses
"Meeting with the Board of Directors of NAITA" by repmobrooks, CC BY-SA 2.0.

The problem

A meta-analysis of 21 experiments with 13,582 participants found that disclosing AI authorship reduced trust, attitudes and behavioral intentions (Hedges' g = −0.440), with every experiment pointing the same way.

In brief

The penalty depended on the kind of task. In subjective settings such as advertising, service and chatbot communication the effect was g = −0.522, compared with g = −0.216 in objective informational settings such as news labeling. That difference was statistically significant (p = .004).

Results varied a lot between studies (I² = 80.5%), though effects within objective contexts were highly consistent (I² = 0.0%). The authors say the findings point to a task-dependent model of AI disclosure, with implications for how platforms design transparency labels and govern AI content.

Main points

  1. The analysis combined 21 experimental effect sizes from studies with 13,582 participants in total.
  2. The pooled effect of AI disclosure on trust, attitudes and behavioral intentions was g = −0.440, with a 95% confidence interval of −0.532 to −0.348.
  3. All 21 effect sizes pointed in the same negative direction.
  4. Effects were stronger in subjective contexts such as advertising, service and chatbots (g = −0.522) than in objective ones such as news labeling (g = −0.216).
  5. Heterogeneity across studies was substantial overall (I² = 80.5%) but 0.0% among objective-context studies.
  6. Larger samples were linked to bigger effects in a single-variable analysis, but this vanished when task context was modeled alongside it, because large studies clustered in objective domains.
  7. The authors tie the results to the persuasion knowledge model, domain ownership threat theory and the machine heuristic.

What the review found when it combined the studies

These are the review’s own pooled results, copied from its abstract. Truvace did not recalculate them.

  • Human trust, attitudes and behavioral intentions when AI authorship is disclosed versus not disclosed

    Hedges' g −0.440 (95% CI −0.532 to −0.348)

    21 studies · 13,582 people

    The review’s own words
    the pooled effect is significant and negative (Hedges' g = −0.440, 95% CI [−0.532, −0.348])

Code checks every number above against the review’s text before it is shown. The short labels are an AI model’s reading of that text, so the review’s own sentence sits under each one. A pooled figure is only as strong as the studies behind it. How Truvace uses pooled results

The full story

The researchers used a DerSimonian–Laird random-effects model to combine 21 experimental effect sizes covering trust, attitudes and behavioral intentions. The pooled result was a significant negative effect of AI disclosure (g = −0.440), and no individual effect size went the other way.

Because results differed widely between studies (I² = 80.5%), the authors ran moderator analyses. Task subjectivity was the main moderator: subjective contexts such as advertising, service and chatbot communication had a stronger penalty (g = −0.522) than objective informational contexts such as news labeling (g = −0.216; Q_between = 8.27, p = .004).

Sample size appeared linked to effect size in a single-variable test, but large-sample studies were concentrated in objective domains. When both factors were modeled together the link was no longer significant, so the authors read it as a sign of contextual differences rather than a methodological artifact. They say the results support a task-dependent approach to transparency design and platform governance.

Sources

1 on file
  1. Computers in Human Behavior Reportsdoi:10.1016/j.chbr.2026.101345
    Peer-reviewed2026-09-23

Talk it over

Four questions to start onWho carries this cost, and what would have to change to prevent it?

Start with a question

Telling people AI made it lowers their trust, a 21-experiment analysis findshttps://truvace.com/item/when-ai-says-who-made-it-a-meta-analysis-of-ai-disclosure-effects-on-consumer-re

  1. Who carries this cost, and what would have to change to prevent it?

    The problem A meta-analysis of 21 experiments with 13,582 participants found that disclosing AI authorship reduced trust, attitudes and behavioral intentions (Hedges' g = −0.440), with every experiment pointing the same way.

    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 systematic review or meta-analysis with a sample of 13,582, from Computers in Human Behavior Reports.

    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 Results varied substantially across the pooled studies (I² = 80.5%), and the apparent link between sample size and effect size was confounded with task context, so the moderator findings rest on a small set of 21 effect sizes.

    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 Business