aceModerate evidence
More
Meta-analysis of 90 studies finds AI explanations only weakly tied to user trust
Study title: Is Trust Correlated With Explainability in AI? A Meta-Analysis
- Evidence type
- Systematic review or meta-analysis
- Sample
- n = 90
- Sources
- 1 sourcePeer-reviewed
- Published
- 2 min read
- Truvace record
- TRV-2026-2788cite or check it
Study title: Is Trust Correlated With Explainability in AI? A Meta-Analysis
Researchers at Ontario Tech University pooled 90 studies published between 2017 and 2024 to test whether making AI systems explainable raises user trust. They found a positive, statistically significant link, but a small one: a random-effects correlation of 0.194, which falls in the range the authors class as weak (0.1 to 0.3).
Read the study abstract
This study critically examines the commonly held assumption that explicability in artificial intelligence (AI) systems inherently boosts user trust. Utilizing a meta-analytical approach, we conducted a comprehensive examination of the existing literature to explore the relationship between AI explainability and trust. Our analysis, incorporating data from 90 studies, reveals a statistically significant but moderate positive correlation between the explainability of AI systems and the trust they engender among us…
The problem
A meta-analysis of 90 studies found that explainability explains little of user trust in AI (random-effects correlation 0.194, 95% CI 0.174 to 0.210), weaker than some earlier studies had suggested.
In brief
The result challenges a common assumption in explainable AI research that explanations automatically build trust. The authors conclude that explainability contributes but is not the main driver, and that ethical safeguards, bias management and cultural context also matter for trust.
The studies varied a lot in design (heterogeneity I² of 78.3%), and the review focused on non-generative AI, so the figure may not carry over to chatbots and other generative models.
Main points
- The authors searched Scopus, IEEE and Web of Science for 2017 to 2024 and screened 1,305 records (IEEE 302, Scopus 510, Web of Science 493), keeping 90 that reported adequate statistics.
- The fixed-effects model gave a correlation of 0.204 (95% CI 0.194 to 0.214); the random-effects model gave about 0.196, and the text reports 0.194 (95% CI 0.174 to 0.210).
- Heterogeneity across studies was high (I² 78.337, tau² 0.009), so the authors relied on the random-effects model.
- A funnel plot looked symmetrical, which the authors read as minimal publication bias.
- Of the 90 studies, 60 measured trust with surveys or questionnaires and 30 used interviews or observation; 40 presented explanations through interactive dashboards and 25 through narrative explanations.
- Explanation methods included model-agnostic tools such as SHAP and LIME (30 studies), intrinsically interpretable models (25) and visualization techniques (20).
- The authors say designers may do better to improve a system's actual trustworthiness than to try to manipulate subjective trust.
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.
Correlation between AI explainability and user trust in AI systems (fixed-effects model)
Correlation (r) 0.204 (95% CI 0.194 to 0.214)
90 studies
The review’s own words
the correlation coefficient is 0.204 with a 95% confidence interval ranging from 0.194 to 0.214
Correlation between AI explainability and user trust in AI systems (random-effects model)
Correlation (r) 0.194 (95% CI 0.174 to 0.210)
90 studies
The review’s own words
The results of the random-effects model analysis indicate that the correlation between trust and XAI is 0.194, indicating a weak positive correlation. The 95% confidence interval for this correlation ranges from 0.174 to 0.210
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 review followed the PRISMA protocol. Studies were excluded if they did not report main statistical indicators such as sample size, reliability or correlation coefficients, or if they used incorrect statistical methods. Qualitative and mixed studies without statistics were also dropped, leaving 90 of 1,305 records.
The included studies covered 17 on data governance and ethics, 22 on algorithmic transparency and bias mitigation, 18 on user interface design for explainability, and 33 on human-AI interaction, decision-making and trust dynamics. The AI models were mostly deep neural networks (25 studies), decision trees (20), support vector machines (15) and ensemble methods (10).
In the discussion, the authors argue that overemphasis on explainability may obscure deeper sources of distrust, and that diversity, inclusion, bias reduction and global inequalities affect trust. They recommend combining explainability with value-sensitive design and proactive risk management, and say the findings can guide policymakers and system designers on risk management and bias mitigation.
Sources
1 on file- IEEE Transactions on Technology and Societydoi:10.1109/tts.2025.3558448
Talk it over
Who carries this cost, and what would have to change to prevent it?
Meta-analysis of 90 studies finds AI explanations only weakly tied to user trusthttps://truvace.com/item/is-trust-correlated-with-explainability-in-ai-a-meta-analysis
Who carries this cost, and what would have to change to prevent it?
The problem A meta-analysis of 90 studies found that explainability explains little of user trust in AI (random-effects correlation 0.194, 95% CI 0.174 to 0.210), weaker than some earlier studies had suggested.
Answer in the threadHow 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 90, from IEEE Transactions on Technology and Society.
Answer in the threadWhat does it leave open? What else would you want to know before acting on it?
The record’s own caveat The review focused on non-generative AI, and the pooled studies differed substantially from one another (I² of 78.337), so the single correlation summarizes mixed designs and may not apply to generative models.
Answer in the threadHave 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