Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 85 of the 86 references it cites were found in a registry or an index; 19 of the 49 sources its claims cite were checked for support.

Explainability in AI: Comparing Human-Like and System-Like Trust Repair Strategies

Open access
Konopka, Björn, Wiesche, Manuel · 2026 · Information Systems Frontiers

Abstract

As AI-based conversational agents (CAs) increasingly automate customer service, inevitable system errors pose a threat to user trust. While eXplainable AI (XAI) techniques are well-established for ex-ante trust formation, their effectiveness for ex-post trust repair remains unexplored. This research investigates whether XAI-based repair strategies (local explanations, counterfactual options) implemented directly by the CA during the interaction can effectively repair trust after errors, compared to CA-implemented human-like strategies (apologies, asking questions). Through a controlled between-subjects online experiment (N = 261), we examined CA repair strategies following a simulated system error, measuring subjective trust and actual continuance decisions. Our findings show that both XAI-based system-like and CASA-aligned human-like strategies repair subjective trust to similar levels, yet XAI-based explanations generate significantly higher rates of actual user continuance decisions following errors. This challenges the human-like-by-default design paradigm for CAs and demonstrates XAI's viability as a post-hoc repair mechanism, extending XAI research beyond trust formation into trust repair contexts.

Overview

CoChat Summary

## TL;DR

This paper tests whether conversational agents can repair trust after an error using either human-like strategies (apology, asking questions) or system-like XAI strategies (local explanations, counterfactual options). In a between-subjects online experiment with 261 participants, both explanation-based and apology-based repair improved subjective trust, but XAI-based repair led to significantly higher actual continuance.

## Key findings

  • The final analyzed sample was 261 participants (59% female, 41% male; mean age = 34.39, SD = 11.41), down from the recruited 300 due to speeding and failed attention checks.
  • The trust violation manipulation worked: post-interaction trust differed strongly between the no-error control and no-TRS condition (p < 0.001, Z = -6.559, η² = 0.495).
  • For subjective trust, apology significantly improved trust versus no repair (H(2) = 15.92, p < 0.001, η² = 0.13), while asking questions did not; for system-like strategies, both local explanations and counterfactual options improved trust versus no repair (H(2) = 11.27, p = 0.004, η² = 0.09).
  • For actual continuance, repair strategy mattered: system-like strategies led to higher continuance than human-like strategies (78.4% vs. 54.7%; χ²(1, N = 174) = 10.89, p < 0.001, V = 0.25). At the treatment level, local explanations: 83.3% continue (35/42) and counterfactual options: 78.3% (36/46) outperformed apology: 62.8% (27/43) and asking questions: 46.5% (20/43).

## Methods

  • Design: Controlled between-subjects online experiment with six groups: five trust-repair treatments (apology, asking questions, local explanations, counterfactual options, no TRS) plus a no-error control group.
  • Sample/context: Recruited on Prolific from adults in Western Europe with fluent English, at least 95% task approval, and prior online shopping experience; the scenario was an e-commerce customer service interaction about a missing item.
  • Analysis: Used Kruskal-Wallis tests, Dunn’s post-hoc tests with Bonferroni correction, Mann-Whitney U tests, and chi-square tests of independence; normality was not satisfied, so the authors used nonparametric tests.

## Limitations

  • The authors state external validity is limited because the study was conducted in a customer service / e-commerce context, and results may not transfer to higher-stakes domains such as healthcare or finance.
  • The design captures only a single interaction in a controlled setting; the paper says this is a limitation because real trust is shaped by cumulative interactions and longitudinal or field studies are needed.
  • The study used a single LLM (Gemini 2.0 Flash) and text-only interactions, which the authors say limits generalizability and does not capture the role of visual or multimodal explanations.
  • Evident from the text: the continuance decision was based on a deceptive fictitious second task, which may limit ecological validity; the paper itself notes that actual continuance could be influenced by the artificial constraints of the experimental setting.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1007/s10796-026-10751-1

Cites

85 of 86 references verified0 citation errors · 1 not checkable

Statistics

Figures

Fig. 1 Overview of the research model
Fig. 1 Overview of the research modelCC-BY · From the PDF
Image 2
CC-BY · From the PDF
Fig. 3 Screenshot of the experiment web page and the CA interface for the local explanations treatment
Fig. 3 Screenshot of the experiment web page and the CA interface for the local explanations treatmentCC-BY · From the PDF

Verification

Claims the paper makes

14 Valid · 3 Not valid · 10 Not checked · 84 Own results and interpretations

66 more claims · 6 with their cited sources checked

Create a free account to see all 111 claims and to verify the 3 references still open.

See all claims — free account
  • It may state two separate claims as one. 19
  • Probably not a substantial claim. 24
  • 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.

  • Consistentχ²(5) = 30.33reported p < .001 · recomputed p < .001

    “χ2 (5, N = 261) = 30.33,p < .001”

  • Consistentχ²(1) = 10.89reported p < .001 · recomputed p < .001

    “χ2 (1, N = 174) = 10.89,p < .001”

2 of 2

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Self-check results appear here once an independent check has confirmed them.

Citations

Cites

Show all 76 references

Research works printed but not linked: 3

  1. Benner, D [D.], Elshan, E., Sofia, S., & Janson, A [A.]. (2021). What do you mean? A review on recovery strategies to overcome conversational breakdowns of conversational agents. In ICIS 2021 Proceedings.
  2. Diederich, S., Janßen-Müller, M., Brendel, A. B., & Morana, S. (2019). Emulating empathetic behavior in online service encounters with sentiment-adaptive responses: Insights from an experiment with a conversational agent. In International Conference on Information Systems (ICIS).
  3. Tomczak, M., & Tomczak, E. (2014). The need to report effect size estimates revisited. An overview of some recommended measures of effect size. Trends in Sport Sciences, 21(1), 19–25.

Metadata

Authors
Konopka, B., & Wiesche, M.
Affiliations
TU Dortmund University (DE)
Corresponding author
Konopka, Björn
Venue
Information Systems Frontiers
Published
2026-06-03
ISSN
  • 1387-3326 (ISSN-L)
  • 1572-9419
OpenAlex
W7163354529
Cited by
1
References
76
Citations by year
2026: 1
FWCI
3.46
Type
journal article
Language
en
Open access
Licence
Locations
Information Systems Frontiers (publishedVersion, cc-by)
Funders
Technische Universität Dortmund
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Explainable Artificial Intelligence (XAI)
  • Social Sciences › Business, Management and Accounting › Marketing › AI in Service Interactions
  • Social Sciences › Business, Management and Accounting › Marketing › Customer Service Quality and Loyalty
Keywords
explainable AI, trust repair, chatbots, customer service automation, system errors, local explanations, counterfactual explanations, apology, user trust

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 2026-10-11, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-11 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

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: Explainability in AI: Comparing Human-Like and System-Like Trust Repair Strategies. https://app.cochat.ai/papers/10.1007/s10796-026-10751-1
The paper
Konopka, B., & Wiesche, M. (2026). Explainability in AI: Comparing Human-Like and System-Like Trust Repair Strategies. Information Systems Frontiers. https://doi.org/10.1007/s10796-026-10751-1

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.