CoChat
Sign Up
Cited by 402Explainable Artificial Intelligence (XAI)Human-Automation Interaction and SafetyEthics and Social Impacts of AI

Explanations Can Reduce Overreliance on AI Systems During Decision-Making

Open access
Vasconcelos, Helena, Jörke, Matthew, Grunde-McLaughlin, Madeleine, Gerstenberg, Tobias, Bernstein, Michael S., Krishna, Ranjay·2023·Proceedings of the ACM on Human-Computer Interaction
DOI 10.1145/3579605
Open this paper in CoChat — free accountOpen PDF

Abstract

Prior work has identified a resilient phenomenon that threatens the performance of human-AI decision-making teams: overreliance, when people agree with an AI, even when it is incorrect. Surprisingly, overreliance does not reduce when the AI produces explanations for its predictions, compared to only providing predictions. Some have argued that overreliance results from cognitive biases or uncalibrated trust, attributing overreliance to an inevitability of human cognition. By contrast, our paper argues that people strategically choose whether or not to engage with an AI explanation, demonstrating empirically that there are scenarios where AI explanations reduce overreliance. To achieve this, we formalize this strategic choice in a cost-benefit framework, where the costs and benefits of engaging with the task are weighed against the costs and benefits of relying on the AI. We manipulate the costs and benefits in a maze task, where participants collaborate with a simulated AI to find the exit of a maze. Through 5 studies (N = 731), we find that costs such as task difficulty (Study 1), explanation difficulty (Study 2, 3), and benefits such as monetary compensation (Study 4) affect overreliance. Finally, Study 5 adapts the Cognitive Effort Discounting paradigm to quantify the utility of different explanations, providing further support for our framework. Our results suggest that some of the null effects found in literature could be due in part to the explanation not sufficiently reducing the costs of verifying the AI's prediction.

AI snapshot

Written from the paper’s full text.

TL;DR

The paper argues that overreliance on AI is a strategic cost-benefit choice rather than an inevitable cognitive bias, and tests this in five maze experiments (N = 731). It finds explanations can reduce overreliance when tasks are hard, explanations are easy to verify, or rewards for accuracy are high.

Key findings

  • In Study 1 (N = 340), overreliance increased with maze difficulty in prediction-only conditions; explanations did not change overreliance in easy tasks (mean difference 0.164, 95% CI −1.678 to 2.17) or medium tasks (0.916, −0.305 to 2.08), but reduced it in hard tasks (1.74, 0.889 to 2.76).
  • Study 2 (N = 340) found that easier-to-parse highlight explanations produced lower overreliance than written explanations in medium tasks (mean difference −1.63, 95% CI −2.78 to −0.559) and hard tasks (−1.632, −2.570 to −0.694).
  • Study 3 (N = 286) found that more salient explanations reduced overreliance further; the salient condition had an average overreliance rate of 0%, suggesting no floor effect.
  • Study 4 (N = 114) found that a higher monetary bonus reduced overreliance (low vs. high bonus mean difference 0.93, 95% CI 0.414 to 1.46) and replicated the hard-task explanation benefit (prediction vs. highlight 2.47, 1.72 to 3.36). Study 5 (N = 76) found higher subjective utility for explanations in medium vs. easy tasks (16.96, 5.22 to 28.7) and for highlight vs. written explanations (18.61, 9.51 to 27.79).

Methods

  • Five online experiments with Prolific participants (total N = 731) used a visual maze-solving task with four possible exits; a simulated AI gave predictions at 80% accuracy, with highlight or written explanations depending on condition.
  • Studies manipulated task difficulty (10×10, 25×25, and 50×50 mazes), explanation difficulty/modality, monetary bonus, and used a Cognitive Effort Discounting paradigm to measure subjective utility.
  • Analyses used Bayesian linear mixed-effects models with random intercepts for participants and mazes; overreliance was defined as accepting an incorrect AI prediction.

Limitations

  • Stated: the low-stakes maze task limits generalization to high-stakes, generative, non-game, uncertain, or incomplete-explanation settings; explanations were idealized and explanation quality was not varied; participants were crowdworkers rather than domain practitioners or long-term users; trust was not directly addressed; Need for Cognition results were mixed.
  • Evident from the text: ecological validity is limited because maze solving is artificial and not a typical human-AI decision domain; the simulated AI’s accuracy was fixed at 80%, so real-world model variability is not represented; some Study 3 cells had small new samples (31 additional participants, 15–16 per new condition), and several conditions reused participants from earlier studies.

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

Integrity signals

Registry

Crossref · 10.1145/3579605

Cites

102 of 102 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

44 of 73 cited sources checked · 23 supported · 2 flagged · accuracy not yet measured

  • Premisecitation

    Cognitive forcing functions can reduce overreliance on incorrect model predictions.

    “The authors find that forcing functions successfully reduce overreliance on incorrect model predictions.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [9] 9. Zana Buçinca, Maja Barbara Malaya, and Krzysztof Z Gajos. 2021. To trust or to think: cognitive forcing functions can reduce overreliance on AI in AI-assisted decision-making. Proceedings of the ACM on Human-Computer Interaction 5, CSCW1 (2021), 1-21. source
  • Premisecitation

    Prior empirical work finds that explanations do not reduce overreliance compared to prediction-only.

    “Surprisingly, overreliance does not reduce when the AI produces explanations for its predictions, compared to only providing predictions.”
    • Also stated in: § 1 INTRODUCTION · plain
      “However, empirically, explanations have failed to improve levels of overreliance [4,9].”
    • Also stated in: § 2 BACKGROUND · hedged
      “Unfortunately, they do not find that explanations improve performance above a prediction-only baseline [4,15], offering evidence for the latter theory-that explanations exacerbate overreliance.”
    Quote in the paperSubstantial claim2 of 3 sources checked
    • cites [4] 4. Gagan Bansal, Tongshuang Wu, Joyce Zhou, Raymond Fok, Besmira Nushi, Ece Kamar, Marco Tulio Ribeiro, and Daniel S. Weld. 2021. Does the Whole Exceed its Parts? The Effect of AI Explanations on Complementary Team Performance. arXiv:2006.14779 [cs.AI] source

      The pilot study (Section 3.2, p. 4) reports that explanations increased reliance even when AI recommendations were incorrect, supporting the point that explanations did not reduce overreliance. But the study’s baseline showed recommendations with confidence scores, not predictions alone; Section 4.1 (p. 5) explicitly distinguishes those conditions. The stated prediction-only comparison is therefore not supported. (LLM, conf=0.88)

      Experimental — accuracy not yet measured
    • cites [9] 9. Zana Buçinca, Maja Barbara Malaya, and Krzysztof Z Gajos. 2021. To trust or to think: cognitive forcing functions can reduce overreliance on AI in AI-assisted decision-making. Proceedings of the ACM on Human-Computer Interaction 5, CSCW1 (2021), 1-21. source

      The Introduction (p. 2) says explanations have not had substantial success in reducing overreliance, and Section 2.2 (p. 4) reports that explanations did not help people detect incorrect AI recommendations. But the copy does not establish the claimed comparison of overreliance with a prediction-only condition; Section 2.2 compares those conditions on overall performance instead. (LLM, conf=0.82)

      Experimental — accuracy not yet measured
    • cites [15] 15. Eric Chu, Deb Roy, and Jacob Andreas. 2020. Are Visual Explanations Useful? A Case Study in Model-in-the-Loop Prediction. arXiv:2007.12248 [cs.LG] source
  • Premisecitation

    Explanations anchor humans to the AI's prediction.

    “Potential cognitive mechanisms that underlie these results suggest that the mere presence of an explanation increases trust [96,97], and that explanations anchor humans to the prediction [9,91].”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [9] 9. Zana Buçinca, Maja Barbara Malaya, and Krzysztof Z Gajos. 2021. To trust or to think: cognitive forcing functions can reduce overreliance on AI in AI-assisted decision-making. Proceedings of the ACM on Human-Computer Interaction 5, CSCW1 (2021), 1-21. source

      The Discussion (p. 16) reports that participants using simple explainable AI interfaces followed AI suggestions even when the predictions were incorrect. But the paper's discussion of anchoring (Section 2.2, p. 4) attributes it to seeing the AI recommendation before making one's own decision, not specifically to explanations. The source supports overreliance in an explanation-equipped interface, but not the claimed explanation-driven anchoring mechanism. (LLM, conf=0.84)

      Experimental — accuracy not yet measured
    • cites [91] 91. Danding Wang, Qian Yang, Ashraf Abdul, and Brian Y. Lim. 2019. Designing Theory-Driven User-Centric Explainable AI. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems (Glasgow, Scotland Uk) (CHI '19). Association for Computing Machinery, New York, NY, USA, 1-15. https://doi.org/10.1145/3290605.3300831 source
  • Premisecitation

    Overreliance is the most frequent outcome among possible error types in empirical AI studies.

    “Of the possible error types in human-AI decision-making (Figure 1), overreliance is the most frequent outcome in empirical AI studies [4,8,9,53,97].”
    Quote in the paperNot a substantial claim5 of 5 sources checked
    • cites [4] 4. Gagan Bansal, Tongshuang Wu, Joyce Zhou, Raymond Fok, Besmira Nushi, Ece Kamar, Marco Tulio Ribeiro, and Daniel S. Weld. 2021. Does the Whole Exceed its Parts? The Effect of AI Explanations on Complementary Team Performance. arXiv:2006.14779 [cs.AI] source

      The Introduction (p. 2) reports that explanations increased reliance on AI recommendations even when the AI was incorrect, and Section 6.2 (p. 12) describes this as increased blind trust. The paper does not establish that overreliance is the most frequent error type across empirical AI studies. (LLM, conf=0.85)

      Experimental — accuracy not yet measured
    • cites [8] 8. Zana Buçinca, Phoebe Lin, Krzysztof Z. Gajos, and Elena L. Glassman. 2020. Proxy Tasks and Subjective Measures Can Be Misleading in Evaluating Explainable AI Systems. In Proceedings of the 25th International Conference on Intelligent User Interfaces (Cagliari, Italy) (IUI '20). Association for Computing Machinery, New York, NY, USA, 454-464. https://doi.org/10.1145/3377325.3377498 source

      The Introduction (p. 1) identifies over-reliance and self-reliance as possible problems in human–AI decision-making, and Results (p. 8) discusses errors and over-reliance in this study. Neither passage compares the frequency of error types across empirical AI studies, so the claim that overreliance is the most frequent outcome is unsupported. (LLM, conf=0.82)

      Experimental — accuracy not yet measured
    • cites [9] 9. Zana Buçinca, Maja Barbara Malaya, and Krzysztof Z Gajos. 2021. To trust or to think: cognitive forcing functions can reduce overreliance on AI in AI-assisted decision-making. Proceedings of the ACM on Human-Computer Interaction 5, CSCW1 (2021), 1-21. source

      The Introduction (p. 1) says people frequently overrely on AI, and the Objective measures section (p. 11) distinguishes overreliance from other outcomes when AI predictions are wrong. The paper does not establish that overreliance is the most frequent error type across empirical AI studies; its comparisons concern outcomes in its own experiment. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
    • cites [53] 53. Vivian Lai, Han Liu, and Chenhao Tan. 2020. "Why is 'Chicago' Deceptive?" Towards Building Model-Driven Tutorials for Humans. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems (Honolulu, HI, USA) (CHI '20). Association for Computing Machinery, New York, NY, USA, 1-13. https://doi.org/10.1145/3313831.3376873 source

      The Trust Analysis (p. 16) compares trust in correct and incorrect machine predictions, and the Results (p. 8) report how often participants outperformed the model. Neither passage compares the frequency of overreliance with other error types across empirical AI studies. Searches for overreliance and error frequencies found no such comparison in our copy. (LLM, conf=0.72)

      Experimental — accuracy not yet measured
    • cites [97] 97. Yunfeng Zhang, Q. Vera Liao, and Rachel K. E. Bellamy. 2020. Effect of Confidence and Explanation on Accuracy and Trust Calibration in AI-Assisted Decision Making. In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency (Barcelona, Spain) (FAT* '20). Association for Computing Machinery, New York, NY, USA, 295-305. https://doi.org/10.1145/3351095.3372852 source

      The Introduction (page unavailable) describes the risk of mistakenly following AI recommendations when the model is likely to err. Related Work (page unavailable) reports increased acceptance of AI predictions regardless of their correctness. These passages address overreliance, but neither compares error types nor says it is the most frequent outcome in empirical AI studies. (LLM, conf=0.83)

      Experimental — accuracy not yet measured
  • Premisecitation

    Written explanations increase the cost of engaging with the task relative to highlight explanations because participants must parse the writing and map it onto the maze.

    “Our written explanations are more analogous to logical written arguments explaining the process of arriving at the answer [4], and in our context, increase the cost of engaging with the task (using the explanations) relative to the highlight explanation, as participants have to parse the writing and map it onto the maze.”
    • Also stated in: § 6.1 Conditions · plain
      “These easy explanations are contrasted with written explanations that incur more cognitive effort, since participants need to follow along these written explanations that are on the side of the maze (see Figure 8).”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [4] 4. Gagan Bansal, Tongshuang Wu, Joyce Zhou, Raymond Fok, Besmira Nushi, Ece Kamar, Marco Tulio Ribeiro, and Daniel S. Weld. 2021. Does the Whole Exceed its Parts? The Effect of AI Explanations on Complementary Team Performance. arXiv:2006.14779 [cs.AI] source
  • Classified as probably not a substantial claim. 72
  • May combine two claims. 31
  • The quote only partly states this claim. 3

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The mere presence of an explanation increases trust.

    “Potential cognitive mechanisms that underlie these results suggest that the mere presence of an explanation increases trust [96,97], and that explanations anchor humans to the prediction [9,91].”
    Quote in the paperNot a substantial claim2 of 2 sources checked
    • cites [96] 96. Kun Yu, Shlomo Berkovsky, Ronnie Taib, Jianlong Zhou, and Fang Chen. 2019. Do I Trust My Machine Teammate? An Investigation from Perception to Decision. In Proceedings of the 24th International Conference on Intelligent User Interfaces (Marina del Ray, California) (IUI '19). Association for Computing Machinery, New York, NY, USA, 460-468. https://doi.org/10.1145/3301275.3302277 source

      Section 4.2 reports that users relied more on the system when they trusted it, but does not test whether an explanation increases trust. A search for “explanation” found no passage addressing the effect of an explanation’s presence; the task description in Section 3.2 describes recommendations and feedback instead. Page numbers were not available in our copy. (LLM, conf=0.76)

      Experimental — accuracy not yet measured
    • cites [97] 97. Yunfeng Zhang, Q. Vera Liao, and Rachel K. E. Bellamy. 2020. Effect of Confidence and Explanation on Accuracy and Trust Calibration in AI-Assisted Decision Making. In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency (Barcelona, Spain) (FAT* '20). Association for Computing Machinery, New York, NY, USA, 295-305. https://doi.org/10.1145/3351095.3372852 source

      In 4.2.1 (Results), the explanation condition was not significantly different from the no-information baseline on a measure of trust in model predictions (p = .66). The paper therefore does not find that merely showing an explanation increases trust in its experiment. (LLM, conf=0.89)

      Checked by a reasoning model — may be wrong.

      Experimental — accuracy not yet measured

  • Classified as probably not a substantial claim. May combine two claims.

    Premisecitation

    Cognitive forcing functions also reduce reliance on correct model predictions, yielding no significant differences when humans are not aided by an AI.

    “However, they also reduce reliance on correct model predictions, yielding no significant differences when humans are not aided by an AI.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [9] 9. Zana Buçinca, Maja Barbara Malaya, and Krzysztof Z Gajos. 2021. To trust or to think: cognitive forcing functions can reduce overreliance on AI in AI-assisted decision-making. Proceedings of the ACM on Human-Computer Interaction 5, CSCW1 (2021), 1-21. source

      Section 4.1 (p. 10) reports a significant difference from the no-AI group: on questions with incorrect model predictions, unaided participants performed better than participants in either AI-assisted category. The Discussion (p. 16) reports no significant overall performance difference between cognitive forcing and simple explainable AI, not between cognitive forcing and no AI. The cited passages do not establish reduced reliance on correct predictions. (LLM, conf=0.84)

      Checked by a reasoning model — may be wrong.

      Experimental — accuracy not yet measured

  • May combine two claims.

    Premisecitation

    People are influenced by AI and often accept incorrect decisions without verifying whether the AI was correct.

    “Unfortunately, the literature has repeatedly established that people are influenced by the AI and often accept incorrect decisions without verifying whether the AI was correct.This phenomenon is known as overreliance [9].”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [9] 9. Zana Buçinca, Maja Barbara Malaya, and Krzysztof Z Gajos. 2021. To trust or to think: cognitive forcing functions can reduce overreliance on AI in AI-assisted decision-making. Proceedings of the ACM on Human-Computer Interaction 5, CSCW1 (2021), 1-21. source

  • Classified as probably not a substantial claim.

    Premisecitation

    Prior work [4,9] has failed to find differences between prediction and explanation conditions.

    “This finding stands in contrast against prior work [4,9], which has failed to find differences between the conditions.”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites [4] 4. Gagan Bansal, Tongshuang Wu, Joyce Zhou, Raymond Fok, Besmira Nushi, Ece Kamar, Marco Tulio Ribeiro, and Daniel S. Weld. 2021. Does the Whole Exceed its Parts? The Effect of AI Explanations on Complementary Team Performance. arXiv:2006.14779 [cs.AI] source
    • cites [9] 9. Zana Buçinca, Maja Barbara Malaya, and Krzysztof Z Gajos. 2021. To trust or to think: cognitive forcing functions can reduce overreliance on AI in AI-assisted decision-making. Proceedings of the ACM on Human-Computer Interaction 5, CSCW1 (2021), 1-21. source

      The Discussion (p. 16) reports no significant difference in overall performance between cognitive forcing functions and simple explainable AI approaches. But the paper does not identify a prediction condition compared with its explanation condition, and the Objective measures (pp. 10–11) report significant differences between those two categories when AI predictions were incorrect. Thus it supports a narrower no-difference finding, not the stated comparison. (LLM, conf=0.86)

      Experimental — accuracy not yet measured

  • Classified as probably not a substantial claim. May combine two claims.

    Premisecitation

    Referring to AI predictions as suggestions counteracts perfect automation schemas and emphasizes that the suggestion can be rejected.

    “We always refer to AI predictions as “suggestions” [74] to counteract perfect automation schemas [21] and emphasize that this suggestion can be rejected.”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [74] 74. Joon Sung Park, Rick Barber, Alex Kirlik, and Karrie Karahalios. 2019. A Slow Algorithm Improves Users' Assessments of the Algorithm's Accuracy. Proceedings of the ACM on Human-Computer Interaction 3, CSCW (2019), 1-15. source
    • cites [21] 21. Mary T Dzindolet, Linda G Pierce, Hall P Beck, and Lloyd A Dawe. 2002. The perceived utility of human and automated aids in a visual detection task. Human Factors 44, 1 (2002), 79-94. source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    NFC is a stable personality trait measuring people's propensity towards completing cognitively difficult tasks.

    “NFC is a stable personality trait measuring people's propensity towards completing cognitively difficult tasks.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [18] 18. Gabriel Lins de Holanda Coelho, Paul H. P. Hanel, and Lukas J. Wolf. 2020. The Very Efficient Assessment of Need for Cognition: Developing a Six-Item Version. Assessment 27, 8 (2020), 1870-1885. https://doi.org/10.1177/ 1073191118793208 arXiv:https://doi.org/10.1177/1073191118793208 PMID: 30095000. source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Generating human-interpretable explanations continues to be a challenge for XAI.

    “However, generating human-interpretable explanations continues to be a challenge for XAI [52].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [52] 52. Isaac Lage, Emily Chen, Jeffrey He, Menaka Narayanan, Been Kim, Sam Gershman, and Finale Doshi-Velez. 2019. An Evaluation of the Human-Interpretability of Explanation. https://doi.org/10.48550/ARXIV.1902.00006 source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Previous work has suggested that trust increases someone's likelihood of using a machine.

    “Previous work has suggested that trust increases someone's likelihood of using a machine [96,97].”
    Quote in the paperNot a substantial claim2 of 2 sources checked
    • cites [96] 96. Kun Yu, Shlomo Berkovsky, Ronnie Taib, Jianlong Zhou, and Fang Chen. 2019. Do I Trust My Machine Teammate? An Investigation from Perception to Decision. In Proceedings of the 24th International Conference on Intelligent User Interfaces (Marina del Ray, California) (IUI '19). Association for Computing Machinery, New York, NY, USA, 460-468. https://doi.org/10.1145/3301275.3302277 source
    • cites [97] 97. Yunfeng Zhang, Q. Vera Liao, and Rachel K. E. Bellamy. 2020. Effect of Confidence and Explanation on Accuracy and Trust Calibration in AI-Assisted Decision Making. In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency (Barcelona, Spain) (FAT* '20). Association for Computing Machinery, New York, NY, USA, 295-305. https://doi.org/10.1145/3351095.3372852 source

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Some studies found that explanations increase trust, some found no significant effect, while others found varying effects under different conditions.

    “Some studies found that explanations increase trust [4,8], some found no significant effect [15,42,97], while others found varying effects under different conditions [49,51].”
    Quote in the paperNot a substantial claim4 of 7 sources checked
    • cites [4] 4. Gagan Bansal, Tongshuang Wu, Joyce Zhou, Raymond Fok, Besmira Nushi, Ece Kamar, Marco Tulio Ribeiro, and Daniel S. Weld. 2021. Does the Whole Exceed its Parts? The Effect of AI Explanations on Complementary Team Performance. arXiv:2006.14779 [cs.AI] source
    • cites [8] 8. Zana Buçinca, Phoebe Lin, Krzysztof Z. Gajos, and Elena L. Glassman. 2020. Proxy Tasks and Subjective Measures Can Be Misleading in Evaluating Explainable AI Systems. In Proceedings of the 25th International Conference on Intelligent User Interfaces (Cagliari, Italy) (IUI '20). Association for Computing Machinery, New York, NY, USA, 454-464. https://doi.org/10.1145/3377325.3377498 source
    • cites [15] 15. Eric Chu, Deb Roy, and Jacob Andreas. 2020. Are Visual Explanations Useful? A Case Study in Model-in-the-Loop Prediction. arXiv:2007.12248 [cs.LG] source
    • cites [42] 42. Harmanpreet Kaur, Harsha Nori, Samuel Jenkins, Rich Caruana, Hanna Wallach, and Jennifer Wortman Vaughan. 2020. Interpreting Interpretability: Understanding Data Scientists' Use of Interpretability Tools for Machine Learning. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems (Honolulu, HI, USA) (CHI '20). Association for Computing Machinery, New York, NY, USA, 1-14. https://doi.org/10.1145/3313831.3376219 source
    • cites [97] 97. Yunfeng Zhang, Q. Vera Liao, and Rachel K. E. Bellamy. 2020. Effect of Confidence and Explanation on Accuracy and Trust Calibration in AI-Assisted Decision Making. In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency (Barcelona, Spain) (FAT* '20). Association for Computing Machinery, New York, NY, USA, 295-305. https://doi.org/10.1145/3351095.3372852 source
    • cites [49] 49. Todd Kulesza, Simone Stumpf, Margaret Burnett, Sherry Yang, Irwin Kwan, and Weng-Keen Wong. 2013. Too much, too little, or just right? Ways explanations impact end users' mental models. In 2013 IEEE Symposium on visual languages and human centric computing. IEEE, 3-10. source
    • cites [51] 51. Johannes Kunkel, Tim Donkers, Lisa Michael, Catalin-Mihai Barbu, and Jürgen Ziegler. 2019. Let Me Explain: Impact of Personal and Impersonal Explanations on Trust in Recommender Systems. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems (Glasgow, Scotland Uk) (CHI '19). Association for Computing Machinery, New York, NY, USA, 1-12. https://doi.org/10.1145/3290605.3300717 source

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationnone given

    The task in Bansal et al. (sentiment analysis) could have been too easy to complete alone and the explanation modality (logical explanations) was too complicated to understand when participants were placed in a higher-effort LSAT task.

    “we postulate that the task (sentiment analysis) could have been too easy to complete alone and that the explanation modality (logical explanations) was too complicated to understand when participants were placed in a higher-effort LSAT task.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [4] 4. Gagan Bansal, Tongshuang Wu, Joyce Zhou, Raymond Fok, Besmira Nushi, Ece Kamar, Marco Tulio Ribeiro, and Daniel S. Weld. 2021. Does the Whole Exceed its Parts? The Effect of AI Explanations on Complementary Team Performance. arXiv:2006.14779 [cs.AI] source

      The Domain and AI Model sections describe LSAT questions as more challenging and more time-consuming than sentiment classification. Section 5.3 suggests that verifying explanations was costlier for LSAT participants. The paper does not say sentiment analysis was too easy to need help, or that participants found the logical explanations too complicated to understand; those are the citing authors’ postulates. (LLM, conf=0.83)

      Experimental — accuracy not yet measured

  • Classified as probably not a substantial claim.

    Premisecitation

    Bansal et al. found that explanations do not decrease overreliance.

    “Bansal et al. [4] find that explanations do not decrease overreliance”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [4] 4. Gagan Bansal, Tongshuang Wu, Joyce Zhou, Raymond Fok, Besmira Nushi, Ece Kamar, Marco Tulio Ribeiro, and Daniel S. Weld. 2021. Does the Whole Exceed its Parts? The Effect of AI Explanations on Complementary Team Performance. arXiv:2006.14779 [cs.AI] source

      Section 6.2 (page not provided) says explanations increased blind trust rather than appropriate reliance, supporting the claim for some explanations. But the analysis of AI recommendations (page not provided) reports that adaptive explanations reduced blind agreement when the AI was uncertain. The unqualified claim overlooks that result; the paper separately notes that the reduction did not significantly improve team accuracy. (LLM, conf=0.88)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    A recent study found that clinicians simply did not have time to engage with feature-based explanations for individual patient predictions and instead requested evidence that the tool had been clinically validated.

    “a recent study found that clinicians simply did not have time to engage with feature-based explanations for individual patient predictions and instead requested evidence that the tool had been clinically validated [37].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [37] 37. Maia Jacobs, Jeffrey He, Melanie F. Pradier, Barbara Lam, Andrew C Ahn, Thomas H McCoy, Roy H Perlis, Finale Doshi-Velez, and Krzysztof Z Gajos. 2021. Designing AI for Trust and Collaboration in Time-Constrained Medical Decisions: A Sociotechnical Lens. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. 1-14. source

  • Classified as probably not a substantial claim.

    Premisecitation

    Buçinca et al. found that cognitive forcing functions decrease overreliance compared to simple explanations.

    “Buçinca et al. [9] find that cognitive forcing functions decrease overreliance compared to simple explanations.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [9] 9. Zana Buçinca, Maja Barbara Malaya, and Krzysztof Z Gajos. 2021. To trust or to think: cognitive forcing functions can reduce overreliance on AI in AI-assisted decision-making. Proceedings of the ACM on Human-Computer Interaction 5, CSCW1 (2021), 1-21. source

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    The effect of a mental model of AI accuracy is unclear, as some studies found that any accuracy score increases trust, others found high accuracy scores increased trust and low accuracy scores decreased trust, while another found that these effects are mediated by an individual's innate likelihood of trusting automation.

    “The effect of a mental model of AI accuracy is unclear, as some studies found that any accuracy score increases trust [55], others found high accuracy scores increased trust and low accuracy scores decreased trust [96,97], while another found that these effects are mediated by an individual's innate likelihood of trusting automation [77].”
    Quote in the paperNot a substantial claim3 of 4 sources checked
    • cites [55] 55. Vivian Lai and Chenhao Tan. 2019. On Human Predictions with Explanations and Predictions of Machine Learning Models: A Case Study on Deception Detection. In Proceedings of the Conference on Fairness, Accountability, and Transparency (Atlanta, GA, USA) (FAT* '19). Association for Computing Machinery, New York, NY, USA, 29-38. https://doi.org/10.1145/3287560.3287590 source

      Split into 2 atoms — 1 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

      Experimental — accuracy not yet measured
    • cites [96] 96. Kun Yu, Shlomo Berkovsky, Ronnie Taib, Jianlong Zhou, and Fang Chen. 2019. Do I Trust My Machine Teammate? An Investigation from Perception to Decision. In Proceedings of the 24th International Conference on Intelligent User Interfaces (Marina del Ray, California) (IUI '19). Association for Computing Machinery, New York, NY, USA, 460-468. https://doi.org/10.1145/3301275.3302277 source

      Split into 2 atoms — 1 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

      Experimental — accuracy not yet measured
    • cites [97] 97. Yunfeng Zhang, Q. Vera Liao, and Rachel K. E. Bellamy. 2020. Effect of Confidence and Explanation on Accuracy and Trust Calibration in AI-Assisted Decision Making. In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency (Barcelona, Spain) (FAT* '20). Association for Computing Machinery, New York, NY, USA, 295-305. https://doi.org/10.1145/3351095.3372852 source

      Split into 2 atoms — 0 confirmed, 2 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

      Experimental — accuracy not yet measured
    • cites [77] 77. Vlad L Pop, Alex Shrewsbury, and Francis T Durso. 2015. Individual differences in the calibration of trust in automation. Human factors 57, 4 (2015), 545-556. source

  • Classified as probably not a substantial claim.

    Premisecitation

    Bayesian statistics have the advantage of allowing researchers to hypothesize null effects.

    “We adopt Bayesian statistics for its advantages in being able to hypothesize null effects[27,73].”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [27] 27. Andrew Gelman, John B Carlin, Hal S Stern, and Donald B Rubin. 1995. Bayesian data analysis. Chapman and Hall/CRC. source
    • cites [73] 73. Alonso Ortega and Gorka Navarrete. 2017. Bayesian Hypothesis Testing: An Alternative to Null Hypothesis Significance Testing (NHST) in Psychology and Social Sciences. In Bayesian Inference, Javier Prieto Tejedor (Ed.). IntechOpen, Rijeka, Chapter 12. https://doi.org/10.5772/intechopen.70230 source

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Some studies use reliance as a proxy measure of trust, while others subjectively measure an individual's trust in the system and innate disposition to trust.

    “Some studies use reliance as a proxy measure of trust [4,15,55,97], while others subjectively measure an individual's trust in the system and innate disposition to trust [8-10, 42, 51, 77].”
    Quote in the paperNot a substantial claim6 of 10 sources checked
    • cites [4] 4. Gagan Bansal, Tongshuang Wu, Joyce Zhou, Raymond Fok, Besmira Nushi, Ece Kamar, Marco Tulio Ribeiro, and Daniel S. Weld. 2021. Does the Whole Exceed its Parts? The Effect of AI Explanations on Complementary Team Performance. arXiv:2006.14779 [cs.AI] source

      Section 5.3 (p. 11) reports participants’ subjective accounts of when they trusted or distrusted the AI, which supports the trust-in-the-system portion. But the post-task survey described in sections 4.3 (p. 7) and 5.2 (p. 10) asked about usefulness and strategies, and our copy does not report measuring an individual’s innate disposition to trust. (LLM, conf=0.82)

      Experimental — accuracy not yet measured
    • cites [15] 15. Eric Chu, Deb Roy, and Jacob Andreas. 2020. Are Visual Explanations Useful? A Case Study in Model-in-the-Loop Prediction. arXiv:2007.12248 [cs.LG] source
    • cites [55] 55. Vivian Lai and Chenhao Tan. 2019. On Human Predictions with Explanations and Predictions of Machine Learning Models: A Case Study on Deception Detection. In Proceedings of the Conference on Fairness, Accountability, and Transparency (Atlanta, GA, USA) (FAT* '19). Association for Computing Machinery, New York, NY, USA, 29-38. https://doi.org/10.1145/3287560.3287590 source
    • cites [97] 97. Yunfeng Zhang, Q. Vera Liao, and Rachel K. E. Bellamy. 2020. Effect of Confidence and Explanation on Accuracy and Trust Calibration in AI-Assisted Decision Making. In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency (Barcelona, Spain) (FAT* '20). Association for Computing Machinery, New York, NY, USA, 295-305. https://doi.org/10.1145/3351095.3372852 source
    • cites [8] 8. Zana Buçinca, Phoebe Lin, Krzysztof Z. Gajos, and Elena L. Glassman. 2020. Proxy Tasks and Subjective Measures Can Be Misleading in Evaluating Explainable AI Systems. In Proceedings of the 25th International Conference on Intelligent User Interfaces (Cagliari, Italy) (IUI '20). Association for Computing Machinery, New York, NY, USA, 454-464. https://doi.org/10.1145/3377325.3377498 source

      The Methods sections (3.1, p. 4; 3.2, p. 6) describe Likert-scale questions asking participants whether they trusted the AI, supporting the part about subjectively measuring trust in the system. They do not describe measuring participants’ innate disposition to trust. The reliance-as-proxy clause has a separate citation group and is not attributed to this source here. (LLM, conf=0.87)

      Experimental — accuracy not yet measured
    • cites [9] 9. Zana Buçinca, Maja Barbara Malaya, and Krzysztof Z Gajos. 2021. To trust or to think: cognitive forcing functions can reduce overreliance on AI in AI-assisted decision-making. Proceedings of the ACM on Human-Computer Interaction 5, CSCW1 (2021), 1-21. source
    • cites [10] 10. Adrian Bussone, Simone Stumpf, and Dympna O'Sullivan. 2015. The role of explanations on trust and reliance in clinical decision support systems. In 2015 international conference on healthcare informatics. IEEE, 160-169. source

      The Data Collection and Analysis section (p. 6) says participants rated their trust in the CDSS on a seven-point scale before and after use, supporting the claim that this study subjectively measured trust in the system. It does not say the study measured participants’ innate disposition to trust; a pre-use rating of this particular system is not the same measure. (LLM, conf=0.87)

      Experimental — accuracy not yet measured
    • cites [42] 42. Harmanpreet Kaur, Harsha Nori, Samuel Jenkins, Rich Caruana, Hanna Wallach, and Jennifer Wortman Vaughan. 2020. Interpreting Interpretability: Understanding Data Scientists' Use of Interpretability Tools for Machine Learning. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems (Honolulu, HI, USA) (CHI '20). Association for Computing Machinery, New York, NY, USA, 1-14. https://doi.org/10.1145/3313831.3376219 source
    • cites [51] 51. Johannes Kunkel, Tim Donkers, Lisa Michael, Catalin-Mihai Barbu, and Jürgen Ziegler. 2019. Let Me Explain: Impact of Personal and Impersonal Explanations on Trust in Recommender Systems. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems (Glasgow, Scotland Uk) (CHI '19). Association for Computing Machinery, New York, NY, USA, 1-12. https://doi.org/10.1145/3290605.3300717 source
    • cites [77] 77. Vlad L Pop, Alex Shrewsbury, and Francis T Durso. 2015. Individual differences in the calibration of trust in automation. Human factors 57, 4 (2015), 545-556. source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Explanations intend to improve performance in part by helping a user calibrate their trust according to a more accurate mental model of the AI's ability and accuracy.

    “Explanations intend to improve performance in part by helping a user calibrate their trust according to a more accurate mental model of the AI's ability and accuracy [4].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [4] 4. Gagan Bansal, Tongshuang Wu, Joyce Zhou, Raymond Fok, Besmira Nushi, Ece Kamar, Marco Tulio Ribeiro, and Daniel S. Weld. 2021. Does the Whole Exceed its Parts? The Effect of AI Explanations on Complementary Team Performance. arXiv:2006.14779 [cs.AI] source

  • May combine two claims.

    Findingown data

    In Study 3, the incomplete explanation condition has less overreliance than the prediction only and written explanation conditions.

    “We also find that the incomplete explanation condition has less overreliance than the prediction only and written explanation conditions”
    Quote in the paperNot a substantial claim

  • The quote only partly states this claim.

    Findingown data

    In Study 3, there is no difference in overreliance between the incomplete and highlight explanation conditions.

    “finds no difference in overreliance in the highlight explanation condition.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    The more bonus provided for completing the task properly, the less overreliance there is.

    “We hypothesize and validate that, the more bonus provided for completing the task properly, the less overreliance there is.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    There is no floor effect on overreliance; people will not overrely if they know the AI is wrong.

    “there is no “floor” effect on overreliance; that is, we observe that people will not overrely if they know the AI is wrong.”
    • Also stated in: § 7.6 Summary · plain
      “we do not find a “floor” effect, i.e. an effect in which we cannot reduce overreliance to 0%.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Findingown data

    Through pilot studies, small changes in the interface affected the results: including a progress bar, the total number of questions, naming the AI, providing feedback on what answers were correct, and explicitly stating the model's accuracy.

    “Through pilot studies, we found that small changes in the interface affected the results: including a progress bar, the total number of questions, naming the AI, providing feedback on what answers were correct, and explicitly stating the model's accuracy.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Overreliance is reduced when people see a benefit to not overrelying: when people are given a larger monetary benefit to complete the task, they overrely less.

    “We also find that overreliance is reduced when people see a benefit to not overrelying: when people are given a larger monetary benefit to complete the task, they overrely less.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    When the cost of engaging with the task is increased (via decreasing the understandability of an explanation), people are less likely to utilize the explanation, which leads to an increase in overreliance.

    “We also find that when we increase the cost of engaging with the task (via decreasing the understandability of an explanation), people are less likely to utilize the explanation, which leads to an increase in overreliance.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    In the prior pre-registered Wikipedia Question Answering Task studies, the hypotheses did not appear to have a large effect.

    “Our hypotheses with the prior pre-registration did not appear to have a large effect, so we ended the study early at a number of participants far lower than pre-registered; therefore, we do not report these values.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    People explicitly forego monetary rewards for an AI in a harder task or for an AI that gives more understandable explanations.

    “We find that people explicitly forego monetary rewards for an AI in a harder task or for an AI that gives more understandable explanations.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisereasoning

    Existing literature finds that explanations exacerbate or do not alter overreliance unless forcing functions compel people to engage with the task.

    “Taken together, existing literature puts forth a resilient finding: that explanations exacerbate or do not alter overreliance unless intervention methods in the form of forcing functions compel people to engage with the task.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationreasoning

    People are likely to overrely if engaging with the task is not the optimal strategy.

    “People are likely to overrely if engaging with the task is not the optimal strategy.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    In prior work, receiving the AI's explanation does not reduce the costs of engaging with the task.

    “In particular, receiving the AI's explanation does not reduce the costs of engaging with the task.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    If the task were more difficult and the AI's explanation reduced the cost of engaging with the task, engaging with the task would become relatively much more attractive.

    “Consider the following example: if the task were more difficult and if the AI's explanation reduced the cost of engaging with the task, the strategy to engage with the task would seem relatively much more attractive.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    If the explanation were to become substantially easier to check, it would become preferable and overreliance would reduce.

    “Likewise, if the explanation were to become substantially easier to check than it was before, it would become preferable and overreliance would reduce.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    In prior work, the cost of engaging with the task using the AI's explanation is roughly equal to engaging via using the prediction or by completing the task alone.

    “This is to say that, in these cases, the cost of engaging with the task using the AI's explanation is roughly equal to engaging via using the prediction or by completing the task alone.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationreasoning

    The cost-benefit framework from behavioral economics applies to the way people collaborate with AI predictions and explanations.

    “We propose that this cost-benefit framework from behavioral economics also applies to the way people collaborate with AI predictions and explanations.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Study design and conductreasoning

    Evaluating with accurate explanations ensures that results are not confounded by faulty or uninformative AI-generated explanations.

    “Evaluating with “accurate” explanations ensures that our results are not confounded by faulty or uninformative AI-generated explanations.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). The quote only partly states this claim.

    Study design and conductown data

    Study 4's sample size came from a power analysis done on pilot data.

    “Our sample size comes from a power analysis done on pilot data.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    The paper's highlight explanations constitute an explanation under the definition from [34].

    “We argue that our highlight explanations constitute as an explanation given by the definition from [34].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [34] 34. Carl G. Hempel and Paul Oppenheim. 1948. Studies in the Logic of Explanation. Philosophy of Science 15, 2 (1948), 135-175. http://www.jstor.org/stable/185169 source

  • Probably not a substantial claim: may not pass S3 (load-bearing). The quote only partly states this claim.

    Study design and conductown data

    Study 2 sample size comes from a power analysis done on pilot data.

    “Our sample size comes from a power analysis done on pilot data.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationreasoning

    Including easy-to-understand explanations reduces the cost of engaging with the task.

    “Therefore, including (easy to understand) explanations reduces the cost of engaging with the task.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Highlight explanations are easier to check than written explanations.

  • May combine two claims.

    Interpretationown data

    Overreliance can be mitigated by explanations, and this occurs precisely when tasks are substantially complex and effortful that reviewing the explanation yields a substantial cognitive effort benefit.

    “This result demonstrates that overreliance can in fact be mitigated by explanations-and that this occurs precisely when tasks are substantially complex and effortful that reviewing the explanation yields a substantial cognitive effort benefit.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    Written explanations constitute a viable explanation because they describe, in written format, the premise that logically leads to the conclusion.

    “We argue that our written explanations also constitute a viable explanation, since it describes, in a written format, the premise (the steps taken to an exit in a maze) which logically leads to the conclusion (the exit to take to get out of the maze).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    When humans and AIs perform roughly equally for a given task, there is no additional incentive for the human decision-maker to over- or underrely on the AI's predictions.

    “When humans and AIs perform roughly equally for a given task, there is no additional incentive for the human decision-maker to over-or underrely on the AI's predictions.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Increasing the benefit of completing the task properly makes the utility of engaging with the task higher than relying on the AI, because relying on the AI is likely to lead to more mistakes.

    “increasing the benefit of completing the task properly makes the utility of the strategy to engage with the task higher than the strategy to rely on the AI, since the latter strategy is likely to lead to more mistakes.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    Since the strategy to rely on the AI inherently results in overreliance, overreliance will increase as task difficulty increases.

    “Since the strategy to rely on the AI inherently results in overreliance, overreliance will increase, as seen here.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    Study 4 was pre-registered.

    “We pre-registered this study (https://osf.io/hgz2x).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Overreliance is responsive not just to the cognitive effort of completing the task, but also to the effort to understand an explanation.

    “In summary, Study 2 demonstrates that overreliance is responsive not just to the cognitive effort of completing the task, but also to the effort to understand an explanation.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data

    Study 1 sample size comes from a power analysis done on pilot data.

    “Our sample size comes from a power analysis done on pilot data.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data

    Study 2 was pre-registered.

    “We pre-registered this study (osf.io/4dbqp) .”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductcitation

    Study 4 used a blocked, within-subjects design in which participants were randomly placed in one bonus condition for the first half and the other condition for the second half.

    “we adopted a blocked, within-subjects study design in which participants are randomly placed in one bonus condition for the first half of the study and the other condition for the second half 4 .”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S1 (asserted).

    Study design and conductreasoning

    The study design neutralizes effects of trust so that differences in overreliance are attributable to cost-benefit manipulations and not to individual differences in trust.

    “we aim to neutralize any effects of trust in our task such that any differences in overreliance are attributable to our cost-benefit manipulations and not to individual differences in trust.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Premisecitation

    Prior work shows that people calculate their judgements of monetary gains and losses relative to a certain reference point.

    “prior work [86] shows that people calculate their judgements of monetary gains and losses relative to a certain reference point”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [86] 86. Amos Tversky and Daniel Kahneman. 1974. Judgment under uncertainty: Heuristics and biases. science 185, 4157 (1974), 1124-1131. source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductreasoning

    Study 3 did not have pre-registration.

    “This experiment is exploratory and therefore does not have pre-registration.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductreasoning

    Cutting the number of trials in the hard condition did not affect how hard people perceived the task to be.

    “This, however, does not affect how hard people perceived the task to be, since (1) people did not have access to how many trials they had left, and (2) qualitative data shows people found the hard task to be more difficult than the other two tasks.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    Removing feedback on answer correctness prevented participants' judgements of the AI's ability from changing across the collaboration phase.

    “We removed feedback on answer correctness as we did not want their judgements of the AI's ability to change across the collaboration phase.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationcitation

    The time pressures doctors face when treating patients inflated the costs of engaging with an explanation.

    “In this example, the time pressures doctors face when treating patients inflated the costs of engaging with an explanation.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given

    Highlight explanations that identify patches of images or phrases provide incomplete information about a model's underlying behavior, unlike the procedural highlight explanations used in the maze solving tasks.

    “These types of highlight explanations provide incomplete information about a model's underlying behavior, unlike the procedural highlight explanations we used in the maze solving tasks.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Interpretationreasoning

    Monetary reward aimed to target increasing benefits, which could also be targeted by increasing the stakes of a task.

    “monetary reward aimed to target increasing benefits, which could also be targeted by increasing the stakes of a task.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Explanations that are harder to understand have higher costs (cognitive effort and time) than those that are easier to understand, and therefore have lower subjective utility.

    “Since explanations that are harder to understand have higher costs (in the form of cognitive effort and time) than those that are easier to understand, they have lower subjective utility.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationreasoning

    High-stakes domains may be a case where the benefits of doing tasks properly is really high, thereby discouraging people from relying on the AI.

    “High-stakes domains, for example, may be a case where the benefits of doing tasks properly is really high, thereby discouraging people from relying on the AI.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    As the task gets harder, engaging with the AI allows people to be more likely correct with less cost (time and cognitive effort).

    “as the task gets harder, this strategy allows people to (1) be more likely to be correct with (2) less cost (via time and cognitive effort).”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    Findings may differ when providing a bonus or not in a crowd-sourced study.

    “These findings also have implications on the designs of crowd-sourced studies, for example, as findings may differ when providing a bonus or not in a study.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Prior work has established ample evidence of cognitive biases in high-stakes domains.

    “While prior work has established ample evidence of cognitive biases in high-stakes domain [16,17,31,56,58,89]”
    Quote in the paperNot a substantial claim0 of 6 sources checked
    • cites [16] 16. Pat Croskerry. 2009. Clinical cognition and diagnostic error: applications of a dual process model of reasoning. Advances in health sciences education 14, 1 (2009), 27-35. source
    • cites [17] 17. Pat Croskerry. 2009. A universal model of diagnostic reasoning. Academic medicine 84, 8 (2009), 1022-1028. source
    • cites [31] 31. Chris Guthrie and Andrew J Wistrich. 2007. Blinking on the Bench: How Judges Decide Cases. Cornell Law Review 93 (2007), 1.
    • cites [56] 56. Kathryn Ann Lambe, Gary O'Reilly, Brendan D Kelly, and Sarah Curristan. 2016. Dual-process cognitive interventions to enhance diagnostic reasoning: a systematic review. BMJ quality & safety 25, 10 (2016), 808-820. source
    • cites [58] 58. James H. Lebovic. 2014. National Security Through a Cockeyed Lens: How Cognitive Bias Impacts U.S. Foreign Policy by Steve A. Yetiv. Baltimore, MD, Johns Hopkins University Press, 2013. 168 pp. $24.95. Political Science Quarterly 129, 3 (2014), 534-536. https://doi.org/10.1002/polq.12235 arXiv:https://onlinelibrary.wiley.com/doi/pdf/10.1002/polq.12235 source
    • cites [89] 89. Milica Vasiljevic, Mario Weick, Peter Taylor-Gooby, Dominic Abrams, and Tim Hopthrow. 2013. Reasoning about extreme events: A review of behavioural biases in relation to catastrophe risks. (2013). source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    XAI studies found conflicting results about the effect of explanations on trust.

    “However, XAI studies found conflicting results about the effect of explanations on trust.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Tasks that are inherently enjoyable also have high benefit attached to them, as people naturally want to engage with the task.

    “Tasks that are inherently enjoyable also have high benefit attached to them, as people naturally want to engage with the task [82].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [82] 82. Paul J. Silvia and Todd B. Kashdan. 2009. Interesting Things and Curious People: Exploration and Engagement as Transient States and Enduring Strengths. Social and Personality Psychology Compass 3, 5 (2009), 785-797. https://doi.org/10.1111/j.1751-9004.2009.00210.x arXiv:https://compass.onlinelibrary.wiley.com/doi/pdf/10.1111/j.1751- 9004.2009.00210.x source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationnone given

    Practitioners in high-stakes domains may have vastly different needs and values, yielding different assessments of costs and benefits.

    “practitioners in these domains may have vastly different needs and values, yielding different assessments of costs and benefits.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data

    Study 5 was pre-registered.

    “We pre-registered this study (https://osf.io/cskvb).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Interpretationreasoning

    High stakes tasks may have different levels of costs and benefits, in which case the findings may not generalize.

    “High stakes tasks may have different levels of costs and benefits, in which case our findings may not generalize.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    Since rewards impact overreliance, there may also be a difference in the findings of work that bonuses crowd-workers versus work that does not.

    “Since we found an impact of rewards on overreliance, there may also be a difference in the findings of work that bonuses crowd-workers versus work that does not.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data

    Study 5's sample size came from a power analysis done on pilot data.

    “Our sample size comes from a power analysis done on pilot data.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationreasoning

    Cognitive forcing functions increase the cost of following the strategy to rely on the AI.

    “We postulate that these forcing functions increase the cost of following the strategy to rely on the AI.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premiseown data

    The studies were primarily done with crowd-sourced workers.

    “Our studies were primarily done with crowd-sourced workers.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Interpretationreasoning

    Anthropomorphizing the AI may lead to an increased sense of trust in the AI's accuracy, which may lead people to falsely assume that such an AI has a greater benefit than it actually does.

    “anthropromorphizing the AI may lead to an increase sense of trust in the AI's accuracy, which may lead people to falsely assume that such an AI has a greater benefit (via overall accuracy) than it actually does.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Interpretationreasoning

    The study in [28] was in open-domain question-answering, a more difficult, open-ended task, where explanations can reduce costs of engaging with the task.

    “The study's task was in open-domain question-answering, which is a more difficult, open-ended task, where explanations can reduce costs of engaging with the task.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Prospect theory explains that people value gains and losses differently, placing more weight on perceived gains versus perceived losses.

    “Prospect theory explains that people value gains and losses differently, placing more weight on perceived gains versus perceived losses.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [87] 87. Amos Tversky and Daniel Kahneman. 1992. Advances in prospect theory: Cumulative representation of uncertainty. Journal of Risk and uncertainty 5, 4 (1992), 297-323. source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given

    The crowd-sourced study lacks long-term user engagement.

    “Our crowd-sourced study lacks long-term user engagement,.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Premisecitation

    [28] found that some explanations do help in error detection.

    “[28] find that some explanations do help in error detection.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [28] 28. Ana Valeria Gonzalez, Gagan Bansal, Angela Fan, Robin Jia, Yashar Mehdad, and Srinivasan Iyer. 2020. Human Evaluation of Spoken vs. Visual Explanations for Open-Domain QA. arXiv:2012.15075 [cs.CL] source

  • Classified as probably not a substantial claim.

    Premisenone given

    The task was in a very controlled, low-stakes, maze-solving environment.

    “our task was in a very controlled, low-stakes, maze-solving environment.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationreasoning

    Cognitive forcing functions could decrease overreliance because they increase the difficulty of doing the task.

    “We postulate that this could be due to cognitive forcing functions increasing the difficulty of doing the task.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Interpretationreasoning

    Task difficulty, explanation difficulty, and monetary reward are metaphorical knobs to modulate levels of overreliance.

    “We identify some of the metaphorical knobs to be turned to modulate levels of overreliance: task difficulty, explanation difficulty, and monetary reward.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Bayesian statistics has been proposed as a better alternative to frequentist statistics for HCI research.

    “Bayesian statistics has also been proposed as a better alternative to frequentist statistics for HCI research[43].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [43] 43. Matthew Kay, Gregory L Nelson, and Eric B Hekler. 2016. Researcher-centered design of statistics: Why Bayesian statistics better fit the culture and incentives of HCI. In Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems. 4521-4532. source

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    Prior work in human-AI decision-making found that explanations do not empirically reduce overreliance.

    “Prior work in human-AI decision-making found that explanations do not empirically reduce overreliance.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Study design and conductnone given

    The pre-registration does not specify a blocked study design.

    “The pre-registration does not specify a blocked study design.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data

    The qualitative analysis provides evidence for the proposed explanation of the prior studies' results.

    “We find evidence for this in our qualitative analysis.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductreasoning

    Random intercepts of participants and mazes were fitted rather than random slopes, because the fixed effects of interest (AI Condition, task difficulty, and explanation condition) do not vary within a group.

    “We fit random intercepts of participants and mazes, as opposed to random slopes, as our fixed effects of interest (AI Condition, task difficulty, and explanation condition) do not vary within a group.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductcitation

    Two prior studies using a Wikipedia Question Answering Task, which made similar hypotheses, were pre-registered.

    “We previously pre-registered two studies 5 that made similar hypotheses using a popular Wikipedia Question Answering Task [79].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [79] 79. Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ Questions for Machine Comprehension of Text. arXiv:1606.05250 [cs.CL] source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given

    When researching the effects of trust on behavior, trust is difficult to measure accurately.

    “when researching the effects of trust on behavior, trust is difficult to measure accurately.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductown data

    The prior Wikipedia Question Answering Task study was ended early at a number of participants far lower than pre-registered, and its values are therefore not reported.

    “Our hypotheses with the prior pre-registration did not appear to have a large effect, so we ended the study early at a number of participants far lower than pre-registered; therefore, we do not report these values.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    A recent study found that placebic, uninformative explanations increased trust in a recommender system nearly as much as real explanations.

    “A recent study found that placebic, uninformative explanations increased trust in a recommender system nearly as much as real explanations [24]”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [24] 24. Malin Eiband, Daniel Buschek, Alexander Kremer, and Heinrich Hussmann. 2019. The impact of placebic explanations on trust in intelligent systems. In Extended Abstracts of the 2019 CHI Conference on Human Factors in Computing Systems. 1-6. source
87 of 87

  • The paper relies on this without support.

    Highlight explanations are easier to check than written explanations.

    “The maze task supports multiple explanation modalities, including highlight explanations, which are easy to check, and written explanations, which are more difficult to check.”

  • The authors state this as an assumption.

    At the converged point of COG-ED, the subjective utility of both tasks are equal.

    “At this converged point, the subjective utility of both tasks are assumed to be equal, allowing the difference in reward to quantify the subjective utility of the low-effort task relative to the high-effort task.”

  • The authors state this as an assumption.

    The indifference point in a choice experiment implies that the decision-maker associates both conditions with equal utilities.

    “The assumption underlying such choice experiments is that the indifference point, i.e. the point at which a decision-maker freely chooses between two conditions at roughly equal rates, implies that the decision-maker associates both conditions with equal utilities [93].”

  • The authors state this as an assumption.

    Results from comparisons between easy and medium tasks can extrapolate to comparisons between easy and hard, as well as medium and hard.

    “We presume that these results can also extrapolate to comparisons between easy and hard, as well as medium and hard.”

  • The authors state this as an assumption.

    Results from the medium task can extrapolate to the hard task.

    “we presume any results can also extrapolate to the hard task, since in prior studies we found that effect sizes of manipulations were much smaller in the medium task than the hard task.”

  • The paper relies on this without support.

    The easy task was unlikely to find any differences in behavior.

    “the easy task was unlikely to find any differences in behavior”

  • The paper relies on this without support.

    Anyone who engages with the strategy of using explanations would want to minimize the cost to do so.

    “Anyone who engages with this strategy would want to also minimize the cost to do so.”

  • The authors state this as an assumption.

    Monetary reward is a proxy for increasing benefits.

    “These, of course, were proxies for other elements of human-AI interaction; for example, monetary reward aimed to target increasing benefits, which could also be targeted by increasing the stakes of a task.”

  • The paper relies on this without support.

    Participants' free-form text reports that they skimmed, scanned, or searched accurately reflect their actual behavior on the task.

    “At least 7/17 (∼ 41.1%) participants in the hard prediction condition wrote in the free-form text box (which asked people something along the lines of, “how did you use the AI?”) that they skimmed, scanned, or searched”
9 of 9

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Overreliance on incorrect LLM outputs can lead to disastrous outcomes when people take actions based on them.

    “This can lead to disastrous outcomes through overreliance [19, 25, 83, 102], when people take actions based on incorrect outputs.”

    From "I'm Not Sure, But...": Examining the Impact of Large Language Models' Uncertainty Expression on User Reliance and Trust (2024) · cites this paper as [102]

    The Introduction defines overreliance as accepting an incorrect AI prediction and says it is concerning in high-stakes domains. The Limitations and Future Work section says the study used a low-stakes maze task and that consequences of wrong decisions could be much higher in high-stakes settings. The paper does not establish disastrous outcomes from people acting on incorrect LLM outputs specifically. (LLM, conf=0.85)

    Experimental — accuracy not yet measured

  • Task performance, measured as correctness of participants' answers and time on task, has been studied in prior work on AI-assisted decision making and specifically for studying the effect of uncertainty expression.

    “Both of these have been studied in prior work on AI-assisted decision making [54], as well as specifically for studying the effect of uncertainty expression [102].”

    From "I'm Not Sure, But...": Examining the Impact of Large Language Models' Uncertainty Expression on User Reliance and Trust (2024) · cites this paper as [102]

    Section 4.5 (page unavailable) measures overreliance by whether participants accept an incorrect AI prediction, supporting the claim that this paper studies answer correctness in AI-assisted decisions. It does not identify time on task as a performance measure; the time estimates in Section 4.4 concern study planning, not measured task performance. The uncertainty-expression portion of the citing sentence is attached to a different citation. (LLM, conf=0.85)

    Experimental — accuracy not yet measured

  • Students' desirable reliance behaviours should be associated with cognitive engagement.

    “Based on previous studies, students' desirable reliance behaviours should be associated with cognitive engagement (Fok & Weld, 2024; Guo et al., 2024; Vasconcelos et al., 2023).”

    From The role of critical thinking on undergraduates' reliance behaviours on generative <scp>AI</scp> in problem‐solving (2025) · cites this paper as Vasconcelos et al., 2023

    The Framework (Section 3) links engaging with a task to reduced overreliance, and the Introduction (Section 1) defines appropriate reliance as accepting correct AI predictions or correcting incorrect ones. This supports the proposed connection, but the paper studies participants in a maze task, not students; its Limitations (Section 10.4) also report mixed results for whether propensity for effortful thinking influenced overreliance. (LLM, conf=0.84)

    Experimental — accuracy not yet measured

  • People tend to over-rely and over-trust AI systems.

    “other research shows people tend to over-rely and over-trust AI systems [3, 28, 31, 50].”

    From Who Validates the Validators? Aligning LLM-Assisted Evaluation of LLM Outputs with Human Preferences (2024) · cites this paper as [50]

    The Introduction says people often accept incorrect AI decisions without checking them and calls this overreliance. The paper links trust to reliance, but §10.4 says measuring and modulating trust was outside its scope; it does not establish that people generally over-trust AI systems. Page numbers were not provided in our copy. (LLM, conf=0.86)

    Experimental — accuracy not yet measured

  • Post-hoc interaction with single-pass AI outputs is opaque and coarse-grained: experts must reverse-engineer the model's reasoning to correct errors, while novices risk automation bias and over-reliance on machine-generated authority.

    “For humans, post-hoc interaction with such outputs is opaque and coarse-grained: experts must reverseengineer the model's reasoning to correct errors, while novices risk automation bias and over-reliance on machine-generated authority [32, 44, 87].”

    From Criticmate: Stagewise Human–AI Co-Critique in Single-Screen UI Evaluation (2026) · cites this paper as [87]

    The Introduction (page unavailable) says people often accept incorrect AI decisions without verifying them, supporting the over-reliance concern. The paper does not establish that this risk is specific to novices, or that experts must reverse-engineer opaque, coarse-grained, single-pass outputs to correct errors. (LLM, conf=0.85)

    Experimental — accuracy not yet measured

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.

  • Methods incorrectly say task difficulty never varies within participants

    Minor

    The model description says task difficulty does not vary within a group, yet Study 1 assigns some participants both easy and medium tasks.

  • Benefit result is attributed to the wrong study

    Minor

    The Study 4 summary attributes its finding about benefits to Study 3, although the paper identifies Study 4 as the experiment testing monetary benefits.

  • Summary caption reverses the easy-task finding

    Minor

    Table 7's caption says explanations decrease overreliance in easier tasks, while its easy-task row says overreliance stays the same.

  • Study 5 gives conflicting descriptions of its high-effort option

    Minor

    The procedure describes the high-effort choice as prediction-only, but then calls it completing the task alone. Those are different conditions, so the account of what the utility measure compares is inconsistent.

4 of 4

Cites

Flagged first, then by citations

Research works printed but not linked: 4

  1. 10.1145/3287560.3287590 · DOI · Paper page
  2. Michael Scott McClendon et al. 2001. The complexity and difficulty of a maze. In Bridges: Mathematical Connections in Art, Music, and Science. Citeseer, 213--222.
  3. 10.18653/v1 · DOI
  4. 10.1111/j.1751--9004.2009.00210.x · DOI

Cited by

Sorted by citations

Related

Authors
Vasconcelos, H., Jörke, M., Grunde-McLaughlin, M., Gerstenberg, T., Bernstein, M. S., & Krishna, R.
Affiliations
  • Stanford University (US)
  • University of Washington (US)
ORCID
Venue
Proceedings of the ACM on Human-Computer Interaction
Publisher
Association for Computing Machinery (ACM)
Published
2023-04-14
Volume
7
Issue
CSCW1
Pages
1-38
ISSN
2573-0142 (ISSN-L)
DOI
10.1145/3579605
OpenAlex
W4366003124
Cited by
402
References
72
Citations by year
2026: 182 · 2025: 121 · 2024: 81 · 2023: 17 · 2022: 1
FWCI
37.70
Type
journal article
Language
en
Open access
Licence
www.acm.org/publications/policies/copyright_policy#Background (vor)
Locations
  • Proceedings of the ACM on Human-Computer Interaction (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
Funders
Snap Inc.
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Explainable Artificial Intelligence (XAI)
  • Social Sciences › Psychology › Social Psychology › Human-Automation Interaction and Safety
  • Social Sciences › Safety Research › Ethics and Social Impacts of AI
Keywords
overreliance on AI, AI explanations, human-AI decision-making, strategic choice, trust in AI, task difficulty, financial compensation, maze task

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.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-11.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-11.

Cited sources are checked by 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