CoChat
Sign Up
Cited by 32Explainable Artificial Intelligence (XAI)Human-Automation Interaction and SafetyEthics and Social Impacts of AI

Cognitive Forcing for Better Decision-Making: Reducing Overreliance on AI Systems Through Partial Explanations

Open access
de Jong, Sander, Paananen, Ville, Tag, Benjamin, van Berkel, Niels·2025·Proceedings of the ACM on Human-Computer Interaction
DOI 10.1145/3710946
Open this paper in CoChat — free accountOpen PDF

Abstract

In AI-assisted decision-making, explanations aim to enhance transparency and user trust but can also lead to negligence. In two separate studies, we explore the use of partial explanations to activate cognitive forcing and increase user engagement. In Study I (N = 264), we present participants with weighted graphs and ask them to identify the shortest paths. In Study II (N = 210), participants correct spelling and grammar mistakes in short text segments. In both studies, we provide a solution suggestion accompanied by either no explanation, a full explanation, or a partial explanation. Our results show that partial explanations reduce overreliance on incorrect AI suggestions, performing significantly better than the baseline but not as well as full explanations. Individuals with a high need for cognition benefit more from AI explanations and consequently perform better. Our work suggests that partial explanations can be valuable in domains where reducing overreliance on AI is critical, like medical diagnosis. It also underscores the need to consider explanation effectiveness across different task difficulties, a factor often overlooked in contemporary human-AI studies.

Integrity signals

Registry

Crossref · 10.1145/3710946

Cites

82 of 82 references verified0 citation errors · 0 not checkable

Statistics

32 tests recomputed · 2 with the reported and recomputed p on opposite sides of .05

Claims the paper makes

9 of 19 cited sources checked · 5 supported · 0 flagged · accuracy not yet measured

  • Premisecitation

    As people experience positive outcomes when following AI explanations, they are less likely to detect subsequent AI mistakes.

    “Additionally, as people experience positive outcomes when following the explanations provided by an AI system, they are less likely to detect any subsequent mistakes by the AI system [2].”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [2] 2. Gagan Bansal, Tongshuang Wu, Joyce Zhou, Raymond Fok, Besmira Nushi, Ece Kamar, Marco Tulio Ribeiro, and Daniel Weld. 2021. Does the Whole Exceed its Parts? The Effect of AI Explanations on Complementary Team Performance. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems (Yokohama, Japan) (CHI '21). Association for Computing Machinery, New York, NY, USA, Article 81, 16 pages. doi:10.1145/3411764.3445717 source

      The Introduction (p. 2) reports that explanations increased reliance on AI recommendations even when the AI was wrong, and the Conclusions (p. 14) report lower accuracy on examples where the AI erred. The paper does not establish that experiencing positive outcomes from following explanations makes people less likely to detect subsequent mistakes; that sequence and causal link go beyond its findings. (LLM, conf=0.87)

      Experimental — accuracy not yet measured
  • Premisecitation

    When presented with an AI suggestion, people tend to anchor their prediction to the suggestion.

    “When presented with an AI suggestion, people tend to anchor their prediction to the suggestion [1,74]”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites [1] 1. Anne Kathrine Petersen Bach, Trine Munch Nørgaard, Jens Christian Brok, and Niels van Berkel. 2023. "If I Had All the Time in the World": Ophthalmologists' Perceptions of Anchoring Bias Mitigation in Clinical AI Support. In Proceedings of ACM SIGCHI Conference on Human Factors in Computing Systems (CHI'23). Article 16, 14 pages. doi:10.1145/3544548.3581513 source
    • cites [74] 74. Helena Vasconcelos, Matthew Jörke, Madeleine Grunde-McLaughlin, Tobias Gerstenberg, Michael S. Bernstein, and Ranjay Krishna. 2023. Explanations Can Reduce Overreliance on AI Systems During Decision-Making. Proc. ACM Hum.-Comput. Interact. 7, CSCW1, Article 129 (2023), 38 pages. doi:10.1145/3579605 source
  • Findingown data

    The differences between the explanation conditions compared to the baseline all significantly affect the participant performance prediction.

    “The differences between the explanation conditions compared to the baseline all significantly affect the participant performance prediction.”
    Quote in the paperSubstantial claim
  • Findingown data

    Reliance significantly positively influences performance (p < 0.001).

    “Reliance also significantly positively influences performance (𝑝 < 0.001).”
    Quote in the paperNot a substantial claim
  • Findingown data

    Explanation conditions increase average reliance (M = 60%, SD = 19%) from baseline (M = 44%, SD = 19%).

    “The explanation conditions also increase the average reliance (𝑀 = 60%, 𝑆𝐷 = 19%) from the baseline level (𝑀 = 44%, 𝑆𝐷 = 19%).”
    Quote in the paperSubstantial claim
  • Classified as probably not a substantial claim. 11
  • May combine two claims. 24

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

    Premisecitation

    Data quality of samples recruited on Prolific is higher than on other labour markets and student samples, with participants paying more attention and providing more meaningful answers.

    “Douglas et al. found that the data quality of samples recruited on Prolific is higher than on other labour markets and student samples, with participants paying more attention to the tasks and providing more meaningful answers [22].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [22] 22. Benjamin D. Douglas, Patrick J. Ewell, and Markus Brauer. 2023. Data quality in online human-subjects research: Comparisons between MTurk, Prolific, CloudResearch, Qualtrics, and SONA. PLOS ONE 18, 3 (2023), 1-17. doi:10. 1371/journal.pone.0279720 source

      The Results and Discussion (page not specified) report that Prolific and CloudResearch provided higher-quality data than MTurk, Qualtrics, and the SONA student sample. The Results describe attention checks and meaningful answers as measures of that quality. The claim overstates the comparison by implying Prolific was better than every other platform; the paper groups it with CloudResearch rather than finding it superior to CloudResearch. (LLM, conf=0.88)

      Experimental — accuracy not yet measured

  • Classified as probably not a substantial claim.

    Premisecitation

    AI systems have been shown to outperform humans in various decision-making tasks across many domains.

    “Artificial Intelligence (AI) systems have been shown to outperform humans in various decisionmaking tasks across many domains.”
    Quote in the paperSubstantial claim2 of 3 sources checked
    • cites [42] 42. 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. doi:10.1145/3287560.3287590 source
    • cites [79] 79. Kun Yu, Shlomo Berkovsky, Ronnie Taib, Dan Conway, Jianlong Zhou, and Fang Chen. 2017. User Trust Dynamics: An Investigation Driven by Differences in System Performance. In Proceedings of the 22nd International Conference on Intelligent User Interfaces (Limassol, Cyprus) (IUI '17). Association for Computing Machinery, New York, NY, USA, 307-317. doi:10.1145/3025171.3025219 source
    • cites [49] 49. Scott Lundberg, Bala Nair, Monica Vavilala, Mayumi Horibe, Michael Eisses, Trevor Adams, David Liston, Daniel Low, Shu-Fang Newman, Jerry Kim, and Su-In Lee. 2018. Explainable machine-learning predictions for the prevention of hypoxaemia during surgery. Nature Biomedical Engineering 2 (2018). doi:10.1038/s41551-018-0304-0 source

      The Results section (“Prescience improves anaesthesiologist’s ability to predict hypoxemia”; page not supplied) reports that Prescience was considerably more accurate than anaesthesiologists on initial and real-time hypoxemia prediction. It does not support the broader claim that AI systems outperform humans across many domains and various decision-making tasks. (LLM, conf=0.88)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Premisecitation

    Cognitive forcing has shown promising signs of reducing overreliance at the cost of increased cognitive load.

    “This method, known as 'cognitive forcing', has recently been adopted in AI systems, where it has shown promising signs of reducing overreliance at the cost of increased cognitive load [6,57].”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [6] 6. 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. Proc. ACM Hum.-Comput. Interact. 5, CSCW1, Article 188 (2021), 21 pages. doi:10.1145/3449287 source
    • cites [57] 57. Saumya Pareek, Niels van Berkel, Eduardo Velloso, and Jorge Goncalves. 2024. Effect of Explanation Conceptualisations on Trust in AI-assisted Credibility Assessment. Proceedings of the ACM on Human-Computer Interaction -CSCW (2024), 1-31. doi:10.1145/3686922 source

  • May combine two claims.

    Premisecitation

    Explanations have been shown to increase trust in AI systems but can also result in over- or underreliance.

    “While explanations have been shown to increase trust in AI systems, they can also result in over-or underreliance [8,34].”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites [8] 8. 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. 160-169. doi:10.1109/ ICHI.2015.26 source
    • cites [34] 34. M. Jacobs, M. Pradier, T. McCoy, P. Roy, F. Doshi-Velez, and G. Krzysztof. 2021. How machine learning recommendations influence clinician treatment selections: example of antidepressant selection. Translational Psychiatry 1 (2021), 1-9. doi:10.1038/s41398-021-01224-x source

      The Discussion (page not provided) says explanations may increase trust and can lead to overreliance; the paper also reports reduced treatment-selection accuracy when feature-based explanations accompanied incorrect recommendations. It does not link explanations to underreliance, so that part of the citing sentence is unsupported. (LLM, conf=0.87)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Findingown data

    For the single edge explanation, there is a strong tendency towards better performance than the baseline, albeit not statistically significant (p = 0.053).

    “For the single edge explanation, we see a strong tendency towards better performance than the baseline, albeit not statistically significant (𝑝 = 0.053); see Table 2 for a detailed overview.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Baseline performance is 56% (SD = 26%), showing that explanations increase participant performance.

    “The baseline (i.e., showing suggestions but no explanations) performance is 56% (𝑆𝐷 = 26%), showing that explanations increase participant performance.”
    Quote in the paperSubstantial claim

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

    Findingown data

    The performance model explains 45% of the performance variance.

    “The model is able to explain 45% of the performance variance.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Splitting the data between correct and incorrect AI suggestions shows a positive correlation between NFC and performance on tasks with incorrect suggestions (r(262) = 0.15, p < 0.018) but not on correct decisions, while the effect of SNS does not change.

    “Splitting the data between correct and incorrect AI suggestions shows a positive correlation between NFC and performance on tasks with incorrect suggestions (r(262) = 0.15, p < 0.018) but not on correct decisions, while the effect of SNS does not change.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Overall, partial explanations resulted in significantly higher performance than the baseline but worse performance than full explanations for both studies, although this difference was only statistically significant in Study II.

    “Overall, partial explanations resulted in significantly higher performance than the baseline but worse performance than full explanations for both studies, although this difference was only statistically significant in Study II.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Partial explanations activate cognitive forcing and increase user engagement with the task.

  • May combine two claims.

    Findingown data

    Tasks with medium and hard difficulty lead to significantly worse performance compared to the easy condition (both p < 0.001).

    “The tasks with medium and hard difficulty lead to significantly worse performance compared to the easy condition (both 𝑝 < 0.001, see Table 1 for all details).”
    Quote in the paperNot a substantial claim

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

    Findingown data

    The reliance model explains 28% of the variance in reliance.

    “The model is able to explain 28% of the variance in reliance.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Participants in the ends (p = 0.016) and full (p < 0.001) explanation conditions perform significantly better compared to the baseline.

    “Post hoc tests indicate that participants in the ends (𝑝 = 0.016) and full (𝑝 < 0.001) explanation conditions perform significantly better compared to the baseline.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Participants in the two partial explanation conditions score similarly (p = 0.979) and not significantly different from participants in the full explanation condition (Single edge-Full p = 0.086; Ends-Full p = 0.192).

    “Participants in the two partial explanation conditions score similarly (𝑝 = 0.979) and not significantly different from participants in the full explanation condition (Single edge -Full: 𝑝 = 0.086 and Ends -Full: 𝑝 = 0.192).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Partial explanations can reduce overreliance on incorrect AI suggestions compared to full explanations, albeit at the cost of decreasing reliance on correct AI suggestions.

    “Our results across both studies indicate that partial explanations can reduce overreliance on incorrect AI suggestions compared to full explanations, albeit at the cost of decreasing reliance on correct AI suggestions.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Partial explanations activate cognitive forcing and increase user engagement with the task.

  • May combine two claims.

    Findingown data

    Participants performed best on easy tasks but worse on medium-difficulty tasks compared to hard ones.

    “Participants performed best on the tasks with easy difficulty but interestingly performed worse on the medium-difficulty tasks as compared to the hard ones.”
    Quote in the paperSubstantial claim

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

    Findingown data

    Participants' self-reported difficulty scores across the three difficulty levels (6, 10, and 14 nodes) are respectively 1.90 (SD = 0.92), 2.99 (SD = 1.05), and 3.09 (SD = 1.13) on a five-point scale.

    “Participants' self-reported difficulty scores across the three difficulty levels (i.e., 6 nodes, 10 nodes, and 14 nodes) are respectively 1.90 (𝑆𝐷 = 0.92), 2.99 (𝑆𝐷 = 1.05), and 3.09 (𝑆𝐷 = 1.13) on a five-point scale.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    NFC and SNS do not correlate significantly with task performance (p = 0.126 and p = 0.109).

    “NFC and SNS do not correlate significantly with task performance (𝑝 = 0.126 and 𝑝 = 0.109).”
    Quote in the paperNot a substantial claim

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

    Findingown data

    On average, participants entered 8.33 distinct answers per task, confirming the task yields non-binary answers.

    “On average, participants entered 8.33 distinct answers per task, confirming our suspicion that the task yields non-binary answers.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    There is a significant difference in reliance between both partial conditions and the full explanation condition (both p < 0.001), indicating less reliance on AI suggestions.

    “Contrary to the performance model, there is a significant difference in reliance between both partial conditions and the full explanation condition (both 𝑝 < 0.001), indicating less reliance on AI suggestions.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    In Study II, participants' average correct task completion is 43% (SD = 22%), making it a considerably harder task than Study I.

    “The participants' average correct task completion is 43% (𝑆𝐷 = 22%), making it a considerably harder task than Study I.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationcitation

    The graph-based task is suitable for participants without specific domain knowledge and can act as a proxy for AI-based routing problems.

    “Therefore, we argue that the graph-based task is suitable for participants without specific domain knowledge and can act as a proxy for AI-based routing problems.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [54] 54. Thomas Mussweiler, Fritz Strack, and Tim Pfeiffer. 2000. Overcoming the Inevitable Anchoring Effect: Considering the Opposite Compensates for Selective Accessibility. Personality and Social Psychology Bulletin 26 (2000), 1142-1150. doi:10.1177/01461672002611010 source
    • Rests on an assumption: The shortest path graph task does not require domain knowledge because it is comparable to puzzles that most people are familiar with.

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

    Study design and conductown data

    VIF values for the reliance model range from 1.01 to 1.03, within the chosen threshold of five to ten.

    “The VIF values are well within our chosen threshold of five to ten, ranging from 1.01 to 1.03.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    In Study I, reduced performance on medium-difficulty tasks is congruent with less reliance on AI suggestions compared to easy and hard tasks, consistent across all explanation conditions.

    “In Study I, the reduced performance on the medium-difficulty tasks is congruent with less reliance on the AI suggestions compared to easy and hard tasks, as consistent across all explanation conditions.”
    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

    VIF for performance model parameters range from 1.00 to 1.51, below the commonly used multicollinearity threshold of five to ten.

    “We check the multicollinearity between the model's parameters and find variance inflation factors (VIF) ranging from 1.00 to 1.51.This is well below the commonly used multicollinearity threshold of five to ten [32].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [32] 32. J. Hair, R. Anderson, and B. Babin. 2009. Multivariate Data Analysis (7 ed.). Prentice Hall.

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

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

    Premisecitation

    The effect of explanations on overreliance depends on the user's cost-benefit analysis weighing costs such as task and explanation difficulty against benefits such as performance increase or monetary compensation.

    “Vasconcelos et al. argue that the effect of explanations on overreliance depends on the user's cost-benefit analysis by weighing costs such as task and explanation difficulty against possible benefits such as a performance increase or monetary compensation [73].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [73] 73. Amos Tversky and Daniel Kahneman. 1974. Judgment under Uncertainty: Heuristics and Biases. Science 185, 4157 (1974), 1124-1131. doi:10.1126/science.185.4157.1124 source

  • May combine two claims.

    Interpretationown data

    Individuals with high need for cognition benefit more from AI explanations and consequently perform better.

    “Individuals with a high need for cognition benefit more from AI explanations and consequently perform better.”
    Quote in the paperNot a substantial claim

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

    Interpretationown data

    Partial explanations can be valuable in domains where reducing overreliance on AI is critical, like medical diagnosis.

    “Our work suggests that partial explanations can be valuable in domains where reducing overreliance on AI is critical, like medical diagnosis.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Partial explanations activate cognitive forcing and increase user engagement with the task.

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

    Study design and conductown data

    Of the recruited participants, 16 failed the attention check, resulting in N = 264 participants for analysis.

    “Of the recruited participants, 16 failed the attention check, resulting in 𝑁 = 264 participants for analysis (131 male, 132 female, and one participant not disclosing their gender) distributed over the explanation styles (Baseline = 66, Ends = 67, Single node = 64, and Full = 67).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    Taken together, full explanations outperform the partial explanation conditions, although the difference is not statistically significant in Study I.

    “Taken together, full explanations outperform the partial explanation conditions, although the difference is not statistically significant in Study I.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    Our results show that partial explanations reduce overreliance on AI suggestions compared to full explanations and reduce users' reliance across the board.

    “Our results show that partial explanations reduce overreliance on AI suggestions compared to full explanations and reduce users' reliance across the board.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The AI suggestions used in the studies are representative of AI systems in real-world decision-making contexts.
30 of 30

  • The paper relies on this without support.

    Partial explanations activate cognitive forcing and increase user engagement with the task.

    “Revealing only part of an explanation prevents users from relying fully on the explanation, thereby activating cognitive forcing, which we hypothesise will challenge users to consider alternative solutions.”

  • The paper relies on this without support.

    Participants believe that the suggestions they receive come from an AI system.

    “Participants are led to believe they had access to an AI system's suggestions to aid them in their tasks.”

  • The paper relies on this without support.

    The shortest path graph task does not require domain knowledge because it is comparable to puzzles that most people are familiar with.

    “The task does not require domain knowledge as it is comparable to puzzles that most people are familiar with.”

  • The paper relies on this without support.

    The AI suggestions used in the studies are representative of AI systems in real-world decision-making contexts.

    “As AI support systems are far from perfect, this can be crucial in critical contexts in which incorrect decisions significantly impact people's lives.”

  • The paper relies on this without support.

    Participants' self-reported perceptions accurately reflect their actual cognitive states.

    “Mean scores for perceived task difficulty (𝑀 = 2.66, 𝑆𝐷 = 0.74), confidence in the AI (𝑀 = 3.26, 𝑆𝐷 = 0.79), and perceived usefulness of the AI (𝑀 = 3.22, 𝑆𝐷 = 0.78) indicate that participants generally appreciate the AI assistance.”

  • The paper relies on this without support.

    It is crucial for users to detect incorrect AI suggestions in AI-assisted decision-making.

    “As AI-assisted decision-making is increasingly common, it is crucial for users to detect incorrect AI suggestions.”
6 of 6

Statistics that don’t add up

Each reported p-value recomputed from its test statistic and degrees of freedom.

  • Inconsistent p-value across alpha = .05: r(262) = 0.48, p = .064

    Major

    The paper reports p = .064 here; recomputing from the test statistic and its degrees of freedom gives p < .001 (two-tailed). The two fall on opposite sides of alpha = .05. The check cannot tell whether the statistic or the p-value is the misprint.

    Reported p = .064 · recomputed p < .001

  • Inconsistent p-value across alpha = .05: r(262) = 0.45, p = .064

    Major

    The paper reports p = .064 here; recomputing from the test statistic and its degrees of freedom gives p < .001 (two-tailed). The two fall on opposite sides of alpha = .05. The check cannot tell whether the statistic or the p-value is the misprint.

    Reported p = .064 · recomputed p < .001

  • Inconsistent p-value: r(208) = 0.15, p = .021

    Minor

    The paper reports p = .021 here; recomputing from the test statistic and its degrees of freedom gives p = .030 (two-tailed). The check cannot tell whether the statistic or the p-value is the misprint.

    Reported p = .021 · recomputed p = .030

3 of 3

Statistics recomputed

Every reported test result, recomputed from its test statistic and degrees of freedom. A result shows as not adding up only once an independent check has confirmed it.

  • Doesn’t add upr(262) = 0.48reported p = .064 · recomputed p < .001The two fall on opposite sides of .05

    “r(262) = 0.48, p = .064”

  • Doesn’t add upr(262) = 0.45reported p = .064 · recomputed p < .001The two fall on opposite sides of .05

    “r(262) = 0.45, p = .064”

  • Doesn’t add upr(208) = 0.15reported p = .021 · recomputed p = .030

    “r(208) = 0.15, p = .021”

  • Consistentr(262) = 0.15reported p < .018 · recomputed p = .015

    “r(262) = 0.15, p < .018”

  • Consistentr(262) = -0.22reported p < .001 · recomputed p < .001

    “r(262) = -0.22,p < .001”

  • Consistentr(262) = 0.86reported p < .001 · recomputed p < .001

    “r(262) = 0.86, p < .001”

  • Consistentr(262) = -0.11reported p = .064 · recomputed p = .074

    “r(262) = -0.11,p = .064”

  • Consistentr(262) = -0.23reported p < .001 · recomputed p < .001

    “r(262) = -0.23,p < .001”

  • Consistentr(262) = -0.16reported p = .007 · recomputed p = .009

    “r(262) = -0.16,p = .007”

  • Consistentr(262) = 0.74reported p < .001 · recomputed p < .001

    “r(262) = 0.74, p < .001”

  • Consistentr(262) = 0.72reported p < .001 · recomputed p < .001

    “r(262) = 0.72, p < .001”

  • Consistentr(262) = 0.42reported p < .001 · recomputed p < .001

    “r(262) = 0.42, p < .001”

  • Consistentr(262) = 0.16reported p = .007 · recomputed p = .009

    “r(262) = 0.16, p = .007”

  • Consistentr(262) = 0.13reported p = .035 · recomputed p = .035

    “r(262) = 0.13, p = .035”

  • Consistentr(262) = 0.16reported p = .01 · recomputed p = .009

    “r(262) = 0.16, p = .010”

  • Consistentr(262) = -0.21reported p < .001 · recomputed p < .001

    “r(262) = -0.21,p < .001”

  • Consistentr(262) = 0.34reported p < .001 · recomputed p < .001

    “r(262) = 0.34, p < .001”

  • Consistentr(208) = 0.14reported p = .042 · recomputed p = .043

    “r(208) = 0.14, p = .042”

  • Consistentr(208) = 0.18reported p = .01 · recomputed p = .009

    “r(208) = 0.18, p = .010”

  • Consistentr(208) = 0.18reported p = .009 · recomputed p = .009

    “r(208) = 0.18, p = .009”

  • Consistentr(208) = 0.17reported p = .016 · recomputed p = .014

    “r(208) = 0.17, p = .016”

  • Consistentr(208) = 0.91reported p < .001 · recomputed p < .001

    “r(208) = 0.91, p < .001”

  • Consistentr(208) = -0.2reported p = .0029 · recomputed p = .004

    “r(208) = -0.20,p = .0029”

  • Consistentr(208) = -0.23reported p < .001 · recomputed p < .001

    “r(208) = -0.23,p < .001”

  • Consistentr(208) = 0.3reported p < .001 · recomputed p < .001

    “r(208) = 0.30, p < .001”

  • Consistentr(208) = -0.16reported p = .024 · recomputed p = .020

    “r(208) = -0.16,p = .024”

  • Consistentr(208) = 0.73reported p < .001 · recomputed p < .001

    “r(208) = 0.73, p < .001”

  • Consistentr(208) = 0.68reported p < .001 · recomputed p < .001

    “r(208) = 0.68, p < .001”

  • Consistentr(208) = 0.14reported p = .037 · recomputed p = .043

    “r(208) = 0.14, p = .037”

  • Consistentr(208) = 0.15reported p = .03 · recomputed p = .030

    “r(208) = 0.15, p = .030”

  • Consistentr(208) = 0.15reported p = .033 · recomputed p = .030

    “r(208) = 0.15, p = .033”

  • Consistentr(208) = -0.32reported p < .001 · recomputed p < .001

    “r(208) = -0.32,p < .001”

32 of 32

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.

  • Discussion overstates the overall NFC-performance result in Study I

    Major

    The discussion says higher need for cognition is associated with significantly higher Study I performance, while the Study I results explicitly report no significant association with overall task performance. The significant result reported there is limited to tasks with incorrect AI suggestions.

  • A Study II post hoc z-ratio has the wrong sign

    Minor

    Table 9 gives a negative estimate and a positive standard error for medium minus hard performance, but reports a positive z-ratio. Dividing the printed estimate by the printed standard error gives approximately −4.53, not +4.64; rounding cannot reverse the sign.

    Reported 4.64 · recomputed -4.533333

  • Study I gives conflicting p-values for the two partial explanations

    Minor

    The reliance results report p = 1.000 for comparing the two partial-explanation conditions, whereas Table 5 reports p = 0.100 for that comparison.

  • Study II contradicts itself on whether medium tasks differ from easy tasks

    Minor

    The results text says the post hoc comparison is not significant, but Table 9 reports p = 0.006 for medium versus easy tasks. That comparison is significant at the paper's usual threshold.

4 of 4

Cites

Flagged first, then by citations

Research works printed but not linked: 6

  1. 10.1037/0022--3514.42.1.116 · DOI
  2. 10.1097/00001888--200308000-00003 · DOI
  3. J. Hair R. Anderson and B. Babin. 2009. Multivariate Data Analysis 7 ed.). Prentice Hall.
  4. 10.1037//0033--2909.110.3.499 · DOI
  5. 10.1111/j.1467--9280.2006.01720.x · DOI
  6. 10.1038/s44159-023-00188--7 · DOI

Cited by

Sorted by citations

Related

Authors
de Jong, S., Paananen, V., Tag, B., & van Berkel, N.
Affiliations
  • Aalborg University, Aalborg, Denmark
  • University of Oulu (FI)
  • UNSW Sydney (AU)
ORCID
Venue
Proceedings of the ACM on Human-Computer Interaction
Publisher
Association for Computing Machinery (ACM)
Published
2025-05-02
Volume
9
Issue
2
Pages
1-30
ISSN
2573-0142 (ISSN-L)
DOI
10.1145/3710946
OpenAlex
W4410513150
Cited by
32
References
78
Citations by year
2026: 23 · 2025: 9
FWCI
12.94
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)
  • VBN Forskningsportal (Aalborg Universitet) (publishedVersion, other-oa)
Funders
Carlsberg Foundation (CF21-0159)
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
AI explanations, overreliance on AI, AI-assisted decision-making, need for cognition, user engagement, human-AI interaction, task difficulty, text correction, diagnosis

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