The DOI matches its Crossref record; 90 of the 90 references it cites were found in a registry or an index; none of the 83 sources its claims cite is checked for support yet.
Are numerical or verbal explanations of AI the key to appropriate user reliance and error detection?
Open accessAbstract
Advances in AI and our limited human capabilities have made AI decision-making opaque to humans. One prerequisite for enhancing the transparency of AI recommendations is improving AI explainability as humans need to be enabled to take responsibility for their actions even with AI support. Our study aims to tackle this issue by investigating two basic approaches to explainability: We evaluate numerical explanations, such as certainty measures, against verbal explanations, such as those provided by LLM as explanatory agents. Specifically, we examine whether verbal or numerical (or both) explanations in tasks of high uncertainty lure users into false beliefs or, on the contrary, promote appropriate reliance. Drawing on an experiment with 441 participants, we explore the dynamics of non-expert users' interactions with AI under varying explanatory conditions. Results show that explanations significantly improve reliance and decision accuracy. Numerical explanations aid in identifying uncertainties and errors, but the users' reliance on the advice falls far behind the given numerical certainty. Verbal explanations foster higher reliance while increasing the risk of over-reliance. Combining both explanation types enhances reliance but further amplifies blind trust in AI.
Overview
CoChat Summary
TL;DR
In an online experiment with 441 participants, the authors compared numerical certainty values, verbal (LLM-style) explanations, both, or neither as explanations for AI recommendations on ten classification tasks. Explanations increased reliance and accuracy overall, but only numerical explanations helped users detect an AI error, while verbal and combined explanations fostered over-reliance.
Key findings
- AI support improved accuracy: participants solved 47.69% of tasks correctly without AI, and every explanation condition significantly improved correct classifications; tasks where the AI reported 100% certainty were answered correctly 70.21% of the time even without AI, versus 36.28% (66% tasks) and 39.24% (83% tasks).
- Reliance did not track stated certainty: 64.49% followed a 66%-certainty recommendation and 67.61% followed an 83% one (difference not significant, Fisher's exact p > .100), rising to 82.65% at 100% certainty (p = .000); about 20% of 100%-certainty tasks were answered without relying on the AI.
- All explanation types increased reliance relative to the no-explanation control, but only the numerical and combined groups showed a significant positive effect on correct classifications; the AI's error (task 4) was repeated significantly less often in the numerical-only group than in the combined group (Fisher's exact p < .005).
- 4.66% of participants (15) relied on every recommendation (8 in combined, 5 verbal, 2 numerical), and 0.31% (1 participant) rejected every recommendation; the AI's error had no lasting effect on reliance in any group.
Methods
- Online between-subjects experiment: 441 participants (119 baseline with no AI, 322 in the AI experiment), randomly assigned in a 2×2 design varying numerical certainty (yes/no) and verbal explanation (yes/no), plus a control receiving advice without explanation. Participants completed ten timed (30-second) classification tasks with monetary incentives (0.4 € per correct answer, max 4 €); one AI recommendation (task 4) was intentionally wrong.
- The AI was a white-box class expression learner (Drill) applied to DBpedia using description logic, producing a prediction-accuracy score used as the numerical explanation (66%, 83%, or 100%) and manually written factual verbal explanations (Wizard-of-Oz).
- Analysis used Fisher exact tests for group comparisons, ordered probit models for between-subject variation, and probit models for within-subject (task-level) variation, with 3,105 decisions analysed (115 timeouts treated as missing).
Limitations
- Stated by the authors: a convenience-based, self-selected online sample means external validity cannot be assumed; findings may be artifact-specific; only local explanations were studied, leaving global explanations unobserved; verbal explanations were created manually rather than by a real LLM; and the built-in error in the fourth of ten tasks is artificial in procedure and timing.
- Evident from the text: the AI was correct in 90% of recommendations, so verbal explanations' association with more correct classifications is partly a function of that high base accuracy; the authors also note that "forced reliance" (users having no better option on difficult tasks) may be mistaken for trust, which their design cannot fully disentangle.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
0 Valid · 0 Not valid · 55 Not checked · 106 Own results and interpretations
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Average time per task varies between treatments as treatments with verbal explanations need significantly longer than the control group and also the group with numerical explanations.”
plainstates a causen = 441- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Our analysis shows that, in contrast to models I and II, only the treatment groups Num. & Verb. Expl. and Num. Expl. (III, IV) have a significant positive effect on the number of correct classifications.”
§ 5. Results > 5.2. Between-treatment variation
plainstates a 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.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Compared to the control group, receiving a Num. & Verb. explanation increases the likelihood of relying on AI advice the most.”
§ 5. Results > 5.3. Within-treatment variation
plainstates a 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.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Verbal explanations foster higher reliance while increasing the risk of over-reliance.”
- Also stated in: § 1. Introduction · plain
“On the one hand, users show higher reliance rates with verbal explanations.”
plainstates a causen = 441- 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.
- Rests on an assumption not shown here.
- Also stated in: § 1. Introduction · plain
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Observation 1: Even with numerical feedback of 100% certainty given, about 20% of these tasks are answered without relying on the AI recommendation.”
§ 5. Results > 5.1. Analysis: AI-support
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.
65 more claims
Create a free account to see all 161 claims and every source verdict.
- Probably not a substantial claim. 60
- It may state two separate claims as one. 50
- The quote only partly states this claim. 1
Citations
Cites
Peeking Inside the Black-Box: A Survey on Explainable Artificial Intelligence (XAI)
Adadi, A., & Berrada, M. (2018) · IEEE Access · doi:10.1109/access.2018.2870052 · Cited by 6,294 · cited 1×
“…complexity of algorithms makes their rationale harder to interpret, driving the need to break down the black-box nature of the system to build trust in or reliance on an AI recommendation through an explanation (Adadi and Berrada 2018; Schemmer et al. 2022; Schoeffer, De-Arteaga, and Kuehl 2022).”
The Mythos of Model Interpretability
Lipton, Z. C. (2016) · arXiv · doi:10.48550/arxiv.1606.03490 · Cited by 5,422 · cited 1×
“First, several studies show that causality is a fundamental attribute for explainability (Gregor and Benbasat 1999; Holzinger et al. 2019; Lipton 2016).”
Explanation in artificial intelligence: Insights from the social sciences
Miller, T. (2019) · Artificial Intelligence · doi:10.1016/j.artint.2018.07.007 · Cited by 5,302 · cited 10×
“Contrasting (verbal) explanations do not rely on the interpretation of functional values, but instead directly provide a rationale for the recommendation (Mittelstadt, Russell, and Wachter 2019).”
InfluentialWhat is AI Literacy? Competencies and Design Considerations
Long, D., & Magerko, B. (2020) · Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems · doi:10.1145/3313831.3376727 · Cited by 3,241 · cited 1×
“This highlights that explanations and AI literacy are important requirements for successful human-AI cooperation (Long and Magerko 2020).”
Algorithm aversion: People erroneously avoid algorithms after seeing them err.
Dietvorst, B. J., Simmons, J. P., & Massey, C. (2015) · Journal of Experimental Psychology: General · doi:10.1037/xge0000033 · Cited by 3,232 · cited 1×
“Contrary to previous findings (Dietvorst, Massey, and Simmons 2015; Dzin-dolet et al. 2003; Rebholz, Koop, and Hütter 2024; Spitzer et al. 2024), it appears that AI error does not have a lasting effect on reliance on AI recommendations.”
Towards A Rigorous Science of Interpretable Machine Learning
Doshi-Velez, F., & Kim, B. (2017) · arXiv · doi:10.48550/arxiv.1702.08608 · Cited by 3,121
AN EXACT TEST FOR POPULATION DIFFERENTIATION
Raymond, M., & Rousset, F. (1995) · Evolution · doi:10.1111/j.1558-5646.1995.tb04456.x · Cited by 3,015
Trust in Automation: Integrating Empirical Evidence on Factors That Influence Trust
Hoff, K. A., & Bashir, M. (2015) · Human Factors: The Journal of the Human Factors and Ergonomics Society · doi:10.1177/0018720814547570 · Cited by 2,817
Guidelines for Human-AI Interaction
Amershi, S., Weld, D., Vorvoreanu, M., Fourney, A., Nushi, B., Collisson, P., Suh, J., Iqbal, S., Bennett, P. N., Inkpen, K., et al. (2019) · Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems · doi:10.1145/3290605.3300233 · Cited by 2,439 · cited 1×
“Many individuals know that AI recommendations are usually made under uncertainty, that missteps can occur, and that probabilistic behaviour is based on nuances of tasks and settings (Amershi et al. 2019).”
Causability and explainability of artificial intelligence in medicine
Holzinger, A., Langs, G., Denk, H., Zatloukal, K., & Müller, H. (2019) · WIREs Data Mining and Knowledge Discovery · doi:10.1002/widm.1312 · Cited by 1,885 · cited 1×
“First, several studies show that causality is a fundamental attribute for explainability (Gregor and Benbasat 1999; Holzinger et al. 2019; Lipton 2016).”
Research works printed but not linked: 6
- 10.1002/for.v1:3 · doi:10.1002/for.v1:3
- De Graaf M. M. and B. F. Malle. 2017. “How People Explain Action (and Autonomous Intelligent Systems Should Too).” In 2017 AAAI Fall Symposium Series. Arlington Virginia USA: AI Access Foundation.
- 10.1111/risk.2005.25.issue-3 · doi:10.1111/risk.2005.25.issue-3
- 10.1146/statistics.2013.1.issue-1 · doi:10.1146/statistics.2013.1.issue-1
- Jussupow E. I. Benbasat and A. Heinzl. 2020. “Why Are We Averse towards Algorithms? A Comprehensive Literature Review on Algorithm Aversion.” In Proceedings of the 28th European Conference on Information Systems (ECIS) 1–16. Marrakech Morocco: AISeL.
- van Harmelen F. V. Lifschitz and B. W. Porter eds. 2008. Handbook of Knowledge Representation Vol. 3 of Foundations of Artificial Intelligence. San Diego Californien USA: Elsevier.
Cited by
Lund, B. D., Alizai, K., Amoje, E., Chandrasekaran, A., Darvischi, S., Denmark, J., PAULRAJ, N. J., Nallamothula, L., Paes, A., & Vemulapalli, B. S. (2026) · AI in Education · doi:10.3390/aieduc2030032
Zhu, Z., & Li, Y. (2026) · Telematics and Informatics · doi:10.1016/j.tele.2026.102464
Metadata
- Authors
- Papenkordt, J., Ngonga Ngomo, A.-C., & Thommes, K.
- Affiliations
- Paderborn University (DE)
- Paderborn University, Paderborn
- Corresponding author
- Papenkordt, Jörg
- Venue
- Behaviour & Information Technology
- Publisher
- Informa UK Limited
- Published
- 2025-10-13
- Volume
- 45
- Issue
- 10
- Pages
- 2094-2115
- ISSN
- 0144-929X (ISSN-L)
- 1362-3001
- OpenAlex
- W4415102320
- Cited by
- 2
- References
- 82
- Citations by year
- 2026: 2
- FWCI
- 0.80
- Type
- journal article
- Language
- en
- Open access
- hybrid
- doi.org
- Licence
- Locations
- Behaviour and Information Technology (publishedVersion, cc-by)
- Funders
- German Federal Ministry of Research, Technology and Space (02L19C115)
- German Research Foundation (438445824)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Explainable Artificial Intelligence (XAI)
- Social Sciences › Human Factors and Ergonomics › Human-Automation Interaction and Safety
- Physical Sciences › Computer Science › Artificial Intelligence › Ethics and Social Impacts of AI
- Keywords
- explainable AI, verbal explanation, large language models, error detection, trust in AI, non-expert users, decision accuracy, AI transparency
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.
A language model drew the claims from the paper on 2026-10-27, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-27 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
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: Are numerical or verbal explanations of AI the key to appropriate user reliance and error detection? https://app.cochat.ai/papers/10.1080/0144929x.2025.2568928
- The paper
- Papenkordt, J., Ngonga Ngomo, A.-C., & Thommes, K. (2026). Are numerical or verbal explanations of AI the key to appropriate user reliance and error detection? Behaviour & Information Technology. https://doi.org/10.1080/0144929x.2025.2568928
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.