Exploring the effects of human-centered AI explanations on trust and reliance
Open accessThe DOI matches its Crossref record; 99 of the 99 references it cites were found in a registry or an index; 23 of the 59 sources its claims cite were checked for support.
Abstract
Transparency is widely regarded as crucial for the responsible real-world deployment of artificial intelligence (AI) and is considered an essential prerequisite to establishing trust in AI. There are several approaches to enabling transparency, with one promising attempt being human-centered explanations. However, there is little research into the effectiveness of human-centered explanations on end-users' trust. What complicates the comparison of existing empirical work is that trust is measured in different ways. Some researchers measure subjective trust using questionnaires, while others measure objective trust-related behavior such as reliance. To bridge these gaps, we investigated the effects of two promising human-centered post-hoc explanations, feature importance and counterfactuals , on trust and reliance. We compared these two explanations with a control condition in a decision-making experiment ( N = 380). Results showed that human-centered explanations can significantly increase reliance but the type of decision-making (increasing a price vs. decreasing a price) had an even greater influence. This challenges the presumed importance of transparency over other factors in human decision-making involving AI, such as potential heuristics and biases. We conclude that trust does not necessarily equate to reliance and emphasize the importance of appropriate, validated, and agreed-upon metrics to design and evaluate human-centered AI.
Overview
CoChat Summary
TL;DR
This online experiment (N = 380) tested feature importance and counterfactual explanations against a control for effects on trust and reliance in an apartment-pricing task. Counterfactual explanations increased reliance only when the AI recommended decreasing the price, while explanations did not affect trust and recommendation direction had a larger effect on reliance.
Key findings
- Human-centered explanations did not significantly affect trust: TPA overall H(2) = 1.54, p = 0.46; trust H(2) = 5.06, p = 0.08; distrust H(2) = 2.03, p = 0.36. H4–H6 were rejected.
- Reliance was significantly higher for counterfactual explanations vs. control only when the AI recommended decreasing the initial price: β = 0.04, 95% CI [0.01, 0.08], t(378) = 2.31, p = 0.02, an approximately 4% increase in WOA. Feature importance increased WOA by 2% vs. control but was not significant (p = 0.27).
- Recommendation type had a larger effect than explanations: increase vs. decrease recommendations was highly significant for WOA, β = 0.05, 95% CI [0.03, 0.07], t(1,899) = 3.62, p < 0.001, and improved model fit, χ²(1) = 13.05, p < 0.001. Participants relied more on recommendations to decrease prices.
- Overall WOA was positive (M = 0.69, SD = 0.36), and TPA ratings were average overall (M = 5.01, SD = 0.86), with higher trust (M = 4.98, SD = 1.05) than distrust (M = 2.95, SD = 1.62).
Methods
- Online decision-making experiment on Amazon Mechanical Turk; 913 recruited, 798 completed, and 380 remained after exclusions. Sample was predominantly male (61%), mean age 37.03 years (SD = 10.15), and 68% had a higher-education degree.
- Participants estimated subleasing prices for six real Zillow apartments, received an alleged AI price recommendation (random 10–20% deviation from their initial estimate), and then gave a final price. The design was 3 (explanation: feature importance vs. counterfactual vs. control) × 2 (recommendation: increase vs. decrease price), with explanation between-subjects and recommendation within-subjects.
- Reliance was measured by Weight of Advice (WOA); trust was measured by the Trust between People and Automation Scale (TPA, 12 items, 7-point). WOA was analyzed with linear mixed-effects models; TPA was analyzed with Kruskal-Wallis tests after exploratory factor analysis indicated a two-factor trust/distrust structure.
Limitations
- The paper states the study was domain-specific to real estate, used low-stakes MTurk decisions, and did not test untrustworthy recommendations or larger deviations from initial estimates; future work should broaden domains and examine high-stakes or poorer-performing AI.
- The paper states the design could not distinguish dispositional, situational, and learned trust, and trust was measured after task completion, so trial-level effects of bias and heuristics on trust could not be examined.
- Evident from the text: the final sample was predominantly male and highly educated, which may limit generalizability; the task deliberately had no “true price,” and WOA captures reliance behavior rather than trust attitudes.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
21 Valid · 0 Not valid · 19 Not checked · 68 Own results and interpretations
The sources with a verdict support this claim; the rest weren’t checked yet, have no full text we could read, or their verdicts aren’t rated yet.
In the paper
“However, research has shown that subjective trust can be a poor predictor of actual reliance (Dzindolet et al., 2003; Miller et al., 2016; Papenmeier et al., 2022).”
- Also stated in: § 1. Introduction · plain
“research has shown that subjective trust can be a poor predictor of actual reliance (Dzindolet et al., 2003; Miller et al., 2016; Papenmeier et al., 2022).”
What the sources say
- Not checked — no full text of the sourcecites Dzindolet et al., 2003 source
- Not checked — no full text of the sourcecites Miller et al., 2016 source
- Supportedcites Papenmeier et al., 2022 source
Section 4.4 reports a weak correlation between self-reported and behavioural trust (r = 0.2998) and concludes that neither is a good proxy for the other. The discussion characterizes the behavioural measure as a better proxy for reliance than for trust as a whole, supporting the claim that subjective trust can poorly predict actual reliance. (LLM, conf=0.91)
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Also stated in: § 1. Introduction · plain
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“To the best of our knowledge, there is little to no empirical research on the effectiveness of these two humancentered explanations derived from the literature (i.e., Adadi and Berrada, 2018) using Miller's criteria in fostering end-users' trust in AI.”
What the sources say
- Not checked — no full text of the sourcecites Adadi and Berrada, 2018 source
- Partly supportedcites Miller, 2019 sourceExperimental — accuracy not yet measured
Section 4.6.6 identifies criteria for evaluating explanations—simplicity, generality, and coherence—and discusses their potential role in generating trust. It does not establish the claimed lack of empirical research on feature-importance and counterfactual explanations evaluated using those criteria for end-users’ trust. (LLM, conf=0.84)
hedged- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
Every source the paper cites for this claim was checked and supports it.
In the paper
“This deliberate choice by the participants to either rely on the AI recommendation or not makes our experiment an actual decision-making task rather than a proxy task (Buçinca et al., 2020).”
What the sources say
- Supportedcites Buçinca et al., 2020 source
The Discussion (p. 9) distinguishes the proxy task from the actual decision-making task: in the latter, users aim to make accurate nutritional decisions and choose whether and how deeply to engage with AI recommendations. It also reports that participants sometimes ignored incorrect recommendations, supporting the cited distinction. (LLM, conf=0.91)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The sources with a verdict support this claim; the rest weren’t checked yet, have no full text we could read, or their verdicts aren’t rated yet.
In the paper
“Even if an AI system is trusted, reliance must not necessarily follow (Kirlik, 1993; Körber, 2018)”
What the sources say
- Not checked — no full text of the sourcecites Kirlik, 1993 source
- Supportedcites Körber, 2018 source
In the theoretical discussion of trust and reliance, the paper states that even when trust is high and an automated system is perceived as capable, reliance does not necessarily follow. This directly supports the citing sentence’s paraphrase. (LLM, conf=0.95)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
Every source the paper cites for this claim was checked and supports it.
In the paper
“Whether trust translates into reliance is more nuanced than often assumed and depends on an interaction between the operator, the automation, and the situation (Körber, 2018).”
What the sources say
- Supportedcites Körber, 2018 source
Körber's theoretical section states the claim almost verbatim: 'Whether trust translates into reliance behavior depends on a dynamic interaction of operator, automation, situational factors, and interface (Lee & See, 2004)', and surrounds it with the nuance the citing sentence asserts — trust and reliance are 'closely related' yet 'distinct constructs', attitude/intention/behaviour stand in a 'probabilistic' rather than deterministic relationship (Ajzen & Fishbein, 1980), and 'even if trust is at a high level ... reliance does not necessarily follow' (Kirlik, 1993), with environmental and cognitive constraints determining whether intention becomes actual reliance. Körber adopts and relies on this point to build his own model and justify a self-report questionnaire, so citing him for it is appropriate even though the four-factor formulation is credited onward to Lee & See (2004); the citing paraphrase drops 'interface' from the list of interacting factors but asserts nothing the source contradicts. (LLM, conf=0.82)
plainstates a possible cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
53 more claims · 1 with their cited sources checked
Create a free account to see all 108 claims and every source verdict.
- Probably not a substantial claim. 29
- It may state two separate claims as one. 26
- The quote only partly states this claim. 1
Statistics recomputed
Every reported test result, recomputed from its test statistic and degrees of freedom. A result shows as not adding up only once an independent check has confirmed it.
- Consistentt(378) = 0.64reported p = .53 · recomputed p = .523
“t (378) = 0.64, p = .53”
- Consistentt(378) = 1.33reported p = .19 · recomputed p = .184
“t (378) = 1.33, p = .19”
- Consistentt(378) = 0.63reported p = .53 · recomputed p = .529
“t (378) = 0.63, p = .53”
- Consistentt(378) = 2.31reported p = .02 · recomputed p = .021
“t (378) = 2.31, p = .02”
- Consistentt(378) = 1.1reported p = .27 · recomputed p = .272
“t (378) = 1.10, p = .27”
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 6. Discussion
“A WOA of 0.70 signifies that participants adopted 70% of the AI recommendations when updating their prior beliefs to form their final estimate.”
§ 4.1. Measures
“WOA is equal to 1 if the participant's final prediction matches the AI recommendation and equal to 0.5 if they average their initial prediction with the AI recommendation.”
Citations
Cites
Fitting Linear Mixed-Effects Models Using lme4
Bates, D., Mächler, M., Bolker, B., & Walker, S. (2015) · Journal of Statistical Software · doi:10.18637/jss.v067.i01 · Cited by 86,835
LeCun, Y., Bengio, Y., & Hinton, G. (2015) · Nature · doi:10.1038/nature14539 · Cited by 84,126
Perceived usefulness, perceived ease of use, and user acceptance of information technology
Davis, F. D. (1989) · MIS Quarterly · doi:10.2307/249008 · Cited by 68,483
Tversky, A., & Kahneman, D. (1974) · Science · doi:10.1126/science.185.4157.1124 · Cited by 28,810
(1980) · Medical Entomology and Zoology · Cited by 25,810
The Framing of Decisions and the Psychology of Choice
Tversky, A., & Kahneman, D. (1981) · Science · doi:10.1126/science.7455683 · Cited by 18,497
Use of Ranks in One-Criterion Variance Analysis
Kruskal, W. H., & Wallis, W. A. (1952) · Journal of the American Statistical Association · doi:10.1080/01621459.1952.10483441 · Cited by 13,912
Arrieta, A. B., Díaz-Rodríguez, N., Del Ser, J., Bennetot, A., Tabik, S., Barbado, A., García, S., Gil-López, S., Molina, D., Benjamins, R., et al. (2019) · arXiv · doi:10.48550/arxiv.1910.10045 · Cited by 11,321
An integrative model of organizational trust
Mayer, R. C., Davis, J. H., & Schoorman, F. D. (1995) · The Academy of Management Review · doi:10.2307/258792 · Cited by 10,736
Barredo Arrieta, A., Díaz-Rodríguez, N., Del Ser, J., Bennetot, A., Tabik, S., Barbado, A., Garcia, S., Gil-Lopez, S., Molina, D., Benjamins, R., et al. (2020) · Information Fusion · doi:10.1016/j.inffus.2019.12.012 · Cited by 10,423
Research works printed but not linked: 5
- 2022, “ACM conference on fairness, accountability, and transparency 2022 (ACM FAccT 2022) call for papers,”
- Biran, 2017, “Explanation and justification in machine learning: a survey,”, 8
- 2016, Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 on the Protection of Natural Persons With Regard to the Processing of Personal Data and on the Free Movement of Such Data, and Repealing Directive 95/46/EC (General Data Protection Regulation)
- Hartmann, 2020, Vertrauen - Die unsichtbare Macht
- Solso, 2005, Cognitive Psychology
Cited by
Patient Consent and The Right to Notice and Explanation of AI Systems Used in Health Care
Hurley, M. E., Lang, B. H., Kostick-Quenet, K. M., Smith, J. N., & Blumenthal-Barby, J. (2025) · The American Journal of Bioethics · doi:10.1080/15265161.2024.2399828 · Cited by 63
“Patient trust in an AI system may also differ due to its particular application and context, regardless of the quality of the explanation given (Scharowski et al. 2023).”
AIdeation: Designing a Human-AI Collaborative Ideation System for Concept Designers
Wang, W.-F., Lu, C.-T., Ponsa i Campanyà, N., Chen, B.-Y., & Chen, M. Y. (2025) · Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems · doi:10.1145/3706598.3714148 · Cited by 50
“Through effective communication, iterative feedback, and user control, these systems create dynamic collaborations to enhance workflows [45, 91, 105].”
Baron, S. (2025) · Philosophy & Technology · doi:10.1007/s13347-024-00837-6 · Cited by 29
“As in the more general case of reliance, explainability can promote reliance on AI (see, e.g., Scharowski et al., 2023; Schemmer et al., 2023; Schoeffer et al., 2024).”
Torkamaan, H., Steinert, S., Pera, M. S., Kudina, O., Freire, S. K., Verma, H., Kelly, S., Sekwenz, M.-T., Yang, J., van Nunen, K., et al. (2024) · Behaviour and Information Technology · doi:10.1080/0144929x.2024.2431068 · Cited by 42
Eckhardt, S., Kühl, N., Dolata, M., & Schwabe, G. (2025) · ACM Computing Surveys · doi:10.1145/3776528 · Cited by 44
Metadata
- Authors
- Scharowski, N., Perrig, S. A. C., Svab, M., Opwis, K., & Brühlmann, F.
- Affiliations
- University of Basel (CH)
- Corresponding author
- Scharowski, Nicolas
- Venue
- Frontiers in Computer Science
- Publisher
- Frontiers Media SA
- Published
- 2023-07-17
- Volume
- 5
- ISSN
- 2624-9898 (ISSN-L)
- OpenAlex
- W4384524844
- Cited by
- 57
- References
- 97
- Citations by year
- 2026: 32 · 2025: 19 · 2024: 6
- FWCI
- 5.04
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- Locations
- Frontiers in Computer Science (publishedVersion, cc-by)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
- OSF Preprints (OSF Preprints) (submittedVersion)
- Funders
- Universität Basel
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Explainable Artificial Intelligence (XAI)
- Physical Sciences › Computer Science › Artificial Intelligence › Ethics and Social Impacts of AI
- Social Sciences › Human Factors and Ergonomics › Human-Automation Interaction and Safety
- Keywords
- trust in AI, AI reliance, AI transparency, post-hoc explanation, feature importance, counterfactual explanations, subjective trust, heuristics and biases, trust metrics
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-09-30, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-09-30 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol and qwen3.8-omni-flash).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Exploring the effects of human-centered AI explanations on trust and reliance. https://app.cochat.ai/papers/10.3389/fcomp.2023.1151150
- The paper
- Scharowski, N., Perrig, S. A. C., Svab, M., Opwis, K., & Brühlmann, F. (2023). Exploring the effects of human-centered AI explanations on trust and reliance. Frontiers in Computer Science. https://doi.org/10.3389/fcomp.2023.1151150
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.