CoChat

Exploring the effects of human-centered AI explanations on trust and reliance

Open access

The 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

Crossref · 10.3389/fcomp.2023.1151150

Cites

99 of 99 references verified0 citation errors · 0 not checkable

Statistics

Verification

Claims the paper makes

21 Valid · 0 Not valid · 19 Not checked · 68 Own results and interpretations

53 more claims · 1 with their cited sources checked

Create a free account to see all 108 claims and every source verdict.

See all claims — free account
  • Probably not a substantial claim. 29
  • It may state two separate claims as one. 26
  • The quote only partly states this claim. 1

Create a free account to see these claims.

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”

  • Consistentχ²(1) = 13.05reported p < .001 · recomputed p < .001

    “χ 2 (1) = 13.05, p < .001”

  • Consistentt(378) = 2.31reported p = .02 · recomputed p = .021

    “t (378) = 2.31, p = .02”

5 of 6 · the rest aren’t loaded

1 more recomputed test

Create a free account to see every test the Stats check recomputed.

See all tests — free account

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.

1 of 1

Citations

Cites

Show all 99 references

Research works printed but not linked: 5

  1. 2022, “ACM conference on fairness, accountability, and transparency 2022 (ACM FAccT 2022) call for papers,”
  2. Biran, 2017, “Explanation and justification in machine learning: a survey,”, 8
  3. 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)
  4. Hartmann, 2020, Vertrauen - Die unsichtbare Macht
  5. Solso, 2005, Cognitive Psychology

Cited by

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
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
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

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.