Truace
Subscribe
70/100

Moderate evidence

Cite
More
Export CSVExport JSONEmailLinkedInX

Review of 76,142 people in AI experiments shows mild aversion, strongest for robots

Study title: Not all AI is created equal: A meta-analysis revealing drivers of AI resistance across markets, methods, and time

Evidence type
Systematic review or meta-analysis
Sample
n = 76,142
Sources
1 sourcePeer-reviewed
Published
3 min read
Truvace record
TRV-2026-2750cite or check it

Study title: Not all AI is created equal: A meta-analysis revealing drivers of AI resistance across markets, methods, and time

Researchers at the University of St.Gallen, ESADE and Baruch College pooled two decades of experiments that compared people's reactions to AI with their reactions to humans performing identical tasks. Across 440 effect sizes from 76,142 participants, reactions to AI were on average slightly negative (Cohen's d = -0.21), and 289 effects were negative against 146 positive.

Read the study abstract

While artificial intelligence (AI) is used by billions of consumers daily through tools like ChatGPT, prior research often documents that consumers are resistant to it. The current research proposes that such resistance is strongly context-dependent, rapidly evolving, and often an artifact of how researchers study it. We provide a comprehensive synthesis of consumer responses to AI by analyzing 440 effect sizes from 76,142 unique participants across two decades of experimental research. Our meta-analysis reveals…

Not all AI is created equal: A meta-analysis revealing drivers of AI resistance across markets, methods, and time
"Container Port Cranes, Southampton Docks - 2" by Terry Robinson, CC BY-SA 2.0.

The problem

A meta-analysis of 440 effect sizes from 76,142 participants found consumers responded more negatively to AI than to humans doing the same task (Cohen's d = -0.21), and far more negatively to robots (d = -0.83).

In brief

How AI was described mattered. Robots drew the strongest negative reaction (d = -0.83), while algorithms (-0.24), assistants (-0.22) and general AI systems (-0.17) clustered together. Transportation (-0.80), legal and public-safety uses (-0.45) and social welfare (-0.40) provoked more resistance than operations and management (-0.12).

The authors report that aversion has faded over time, particularly for judgements of competence and performance, and is approaching zero in the most recent years. The studies vary widely, with a 95 percent prediction interval from -1.20 to 0.78, so future experiments could show aversion, no effect or a preference for AI.

Main points

  1. The meta-analysis covered 72 publications and 172 studies published between 2002 and 2023, and pooled 440 effect sizes from 76,142 participants using a three-level random-effects model.
  2. Of the effects, 289 were negative, 146 positive and 5 null, so 66 percent of studies showed aversion; the pooled effect was small (d = -0.21, 95% CI -0.29 to -0.14).
  3. Robots (13 effects) triggered d = -0.83, significantly worse than algorithms, AI systems or AI assistants, which did not differ significantly from one another.
  4. By outcome type, affective responses such as trust and fairness were most negative (d = -0.31), then behavioral responses such as choice and advice acceptance (-0.22), then cognitive judgements of competence (-0.11).
  5. Extensive AI autonomy produced aversion (d = -0.28), while limited autonomy produced none (d = -0.05); AI described as inferior to humans drew d = -0.53, while superior AI showed no significant aversion.
  6. Studies from 2020 to 2025 (300 effects) averaged d = -0.12, versus -0.47 for 2010 to 2014 and -0.42 for 2015 to 2019; the time trend was positive (0.09).
  7. When the human comparison was a layperson, there was no aversion (d = 0.01); against a human expert it was d = -0.27.

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.

  • Overall consumer responses to AI compared with humans across experimental studies

    Cohen's d -0.21 (95% CI -0.29 to −0.14)

    72 studies · 76,142 people

    The review’s own words
    As illustrated in Fig. 3, a multi-level random effects model confirmed a statistically significant overall negative effect (Cohen’s d = -0.21, 95 % CI[-0.29, −0.14], t(171) = 5.51, p < 0.001).

  • Consumer responses to AI after removing influential outliers

    Cohen's d -0.20 (95% CI -0.26 to −0.13)

    72 studies · 76,142 people

    The review’s own words
    (d = -0.20, 95 % CI [-0.26, −0.13], t(168) = 5.82, p < 0.001)

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 team searched PsycInfo, Business Source Complete and Web of Science up to January 10, 2024, and also scanned reference lists and expert tips. Of 1,096 initial records, an AI-based screener removed non-experimental work, and the first author handled duplicates and full-text screening. Eligible studies compared differently labeled AI with humans doing identical tasks, and participants had to know whether they were evaluating an AI or a human.

Domain results ranged widely. Transportation (7 effects) averaged d = -0.80 and legal and public safety (28 effects) -0.45, while healthcare (55 effects) was -0.19, entertainment and lifestyle (81) -0.18 and investing and finance (60) -0.26. Operations and management (167 effects) was the mildest at -0.12 and was not statistically significant on its own.

Most studies used text-only descriptions (397 of 440 effects) and online samples (256), and only 21 effects came from field settings. Age and gender did not significantly predict aversion, and the geographic split (America -0.24, Europe -0.16, Asia -0.27) was not significant.

Publication-bias checks gave mixed signals: Egger's test and rank correlation found no asymmetry, while trim-and-fill and selection models suggested the true aversion could be larger, with bias-adjusted estimates between -0.13 and -0.54. The authors judged the risk of systematic publication bias to be low.

Sources

1 on file
  1. International Journal of Research in Marketingdoi:10.1016/j.ijresmar.2025.02.005
    Peer-reviewed2025-02-13

Talk it over

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

Start with a question

Review of 76,142 people in AI experiments shows mild aversion, strongest for robotshttps://truvace.com/item/not-all-ai-is-created-equal-a-meta-analysis-revealing-drivers-of-ai-resistance-a

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

    The problem A meta-analysis of 440 effect sizes from 76,142 participants found consumers responded more negatively to AI than to humans doing the same task (Cohen's d = -0.21), and far more negatively to robots (d = -0.83).

    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 76,142, from International Journal of Research in Marketing.

    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 studies differ widely (I2 = 96.15%, effects from -2.43 to 1.66), so the single pooled effect of d = -0.21 describes an average rather than what any one AI product or setting will produce.

    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