CoChat
Sign Up
Cited by 59Human-Automation Interaction and SafetyAI and HR TechnologiesEthics and Social Impacts of AI

Check the box! How to deal with automation bias in AI-based personnel selection

Open access
Kupfer, Cordula, Prassl, Rita, Fleiß, Jürgen, Malin, Christine, Thalmann, Stefan, Kubicek, Bettina·2023·Frontiers in Psychology
DOI 10.3389/fpsyg.2023.1118723PMC PMC10113449
Open this paper in CoChat — free account

Abstract

Artificial Intelligence (AI) as decision support for personnel preselection, e.g., in the form of a dashboard, promises a more effective and fairer selection process. However, AI-based decision support systems might prompt decision makers to thoughtlessly accept the system’s recommendation. As this so-called automation bias contradicts ethical and legal requirements of human oversight for the use of AI-based recommendations in personnel preselection, the present study investigates strategies to reduce automation bias and increase decision quality. Based on the Elaboration Likelihood Model, we assume that instructing decision makers about the possibility of system errors and their responsibility for the decision, as well as providing an appropriate level of data aggregation should encourage decision makers to process information systematically instead of heuristically. We conducted a 3 (general information, information about system errors, information about responsibility) x 2 (low vs. high aggregated data) experiment to investigate which strategy can reduce automation bias and enhance decision quality. We found that less automation bias in terms of higher scores on verification intensity indicators correlated with higher objective decision quality, i.e., more suitable applicants selected. Decision makers who received information about system errors scored higher on verification intensity indicators and rated subjective decision quality higher, but decision makers who were informed about their responsibility, unexpectedly, did not. Regarding aggregation level of data, decision makers of the highly aggregated data group spent less time on the level of the dashboard where highly aggregated data were presented. Our results show that it is important to inform decision makers who interact with AI-based decision-support systems about potential system errors and provide them with less aggregated data to reduce automation bias and enhance decision quality.

AI snapshot

Written from the paper’s full text.

TL;DR

In a 3 × 2 experiment on AI-based personnel preselection, decision makers who were warned about possible system errors showed more verification behavior and rated their subjective decision quality higher, while responsibility instructions did not reduce automation bias. Highly aggregated dashboard data led users to spend less time on the overview level, but did not improve or harm objective decision quality.

Key findings

  • Verification intensity was positively associated with objective decision quality: e.g., pages visited at level 2 detail correlated with objective decision quality in task 1 (r = 0.49, p < 0.01) and task 2 (r = 0.34, p < 0.01). Subjective decision quality correlated only with level 2 detail time (r = 0.23, p < 0.05), clicks (r = 0.21, p < 0.05), and pages visited (r = 0.21, p < 0.05).
  • The error-awareness group showed higher verification intensity than controls: in task 1, they visited more level 2 detail pages (M = 15.18, SD = 11.06 vs. M = 8.32, SD = 9.05; p = 0.03, d = 0.67); in task 2, they spent more time at level 3 (Mdiff = 17.00 s, p = 0.04, d = 0.63), had more level 3 clicks (Mdiff = 1.59, p = 0.04, d = 0.64), and visited more level 3 pages (Mdiff = 1.12, p = 0.04, d = 0.55).
  • The error-awareness group rated subjective decision quality higher than controls (M = 4.26, SD = 0.67 vs. M = 3.72, SD = 1.06; p = 0.05, d = 0.62), but no significant instruction-group differences were found for objective decision quality in task 1 (F(2,90) = 1.22, p = 0.30) or task 2 (F(2,90) = 1.20, p = 0.32).
  • Responsibility instructions did not significantly increase verification intensity or decision quality, and highly aggregated data led to less time on level 1 (task 1: 106.28 s vs. 180.93 s, p < 0.01, d = 0.57; task 2: 62.96 s vs. 105.78 s, p = 0.04, d = 0.44) but no other verification-intensity or decision-quality differences.

Methods

  • A 3 × 2 between-subjects experiment manipulated system instruction (control vs. error-awareness vs. responsibility) and data visualization (highly aggregated matching score vs. less aggregated 5-point rating of three key indicators).
  • Final sample was N = 93 (68% female, 2% diverse; mean age 23 years, SD = 3.89; 90% full-time students), recruited after excluding 3 participants for technical problems and 4 for very short response times. Participants completed two personnel preselection tasks, selecting and ranking 5 of 10 applicants per task, with three embedded system errors per task.
  • Verification intensity was measured via time spent, clicks, and pages visited at dashboard levels; objective decision quality was the number of correctly selected applicants (0–5), and subjective decision quality used four self-developed items (Cronbach α = 0.73). Analyses included Pearson correlations, MANOVA/ANOVA, t-tests, and ANCOVA controlling for technical affinity and conscientiousness.

Limitations

  • The paper states that participants were students rather than actual HR professionals, the task was isolated rather than embedded in real stressful HR work, and the preselection task was relatively simple (mean objective decision quality above 4 out of 5 in all groups, only 10 applicants per task).
  • The paper also notes high standard deviations in verification intensity within groups and calls for future research on individual differences such as need for cognition, information-seeking styles, and visualization preferences.
  • Evident from the text: the manipulation-check analyses report degrees of freedom smaller than the full sample (e.g., F(2,57), F(2,59)), suggesting missing data or exclusions not fully explained in that section.

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

Integrity signals

Registry

Crossref · 10.3389/fpsyg.2023.1118723

Cites

71 of 71 references verified0 citation errors · 0 not checkable

Statistics

19 tests recomputed · 0 with the reported and recomputed p on opposite sides of .05

Claims the paper makes

14 of 73 cited sources checked · 7 supported · 0 flagged · accuracy not yet measured

  • Premisecitation

    Unawareness of a system's capacities can lead to overestimation of its credibility and heuristic reliance, enhancing automation bias.

    “In terms of using technology, the unawareness of the system’s capacities, i.e., its reliability, can lead to an overestimation of the systems’ credibility as users might heuristically decide to trust a system without systematically evaluating its capacities”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites Buçinca et al., 2021 source

      The Introduction says people may use heuristics about an AI’s competence rather than evaluate individual explanations, and that explanations can increase trust and overreliance. Section 2.2 also describes reliance on incorrect AI predictions. The paper does not specifically establish that unawareness of the system’s reliability causes users to overestimate its credibility. (LLM, conf=0.84)

      Experimental — accuracy not yet measured
  • Premisecitation

    Diffusion of responsibility occurs in tasks humans share with automatic systems.

    “Diffusion of responsibility also occurs in tasks humans share with automatic systems”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites Skitka et al., 2000a

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

    • cites Zerilli et al., 2019 source
  • Premisecitation

    When sharing the selection task with AI, decision makers might not attribute the decision outcome to their own effort.

    “When sharing the selection task with an AI-based decision support system, decision makers might not attribute the decision outcome to their own effort”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites Nolan et al., 2016 source

      The “Attributions and Employee Selection” section says standardized practices can reduce the extent to which people attribute a hiring outcome to the practitioner’s judgment. The paper’s computer-program hypothesis concerns stakeholders’ perceptions, and its discussion of practitioners concerns what they believe stakeholders think. It does not establish that decision makers themselves attribute less of the outcome to their own effort. Page numbers were not available in our copy. (LLM, conf=0.85)

      Experimental — accuracy not yet measured
  • Premisecitation

    Automation bias has been found to lead to adverse effects on decision outcomes in several other contexts.

    “Automation bias has been found to lead to adverse effects on decision outcomes in several other contexts before (e.g., Parasuraman and Manzey, 2010; Goddard et al., 2012).”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites Parasuraman and Manzey, 2010 source
    • cites Goddard et al., 2012 source
  • Premisecitation

    Systems to aid decision-making might especially be considered for novice HR professionals as these systems tend to improve the decision-making quality of less experienced decision makers.

    “However, systems to aid decision-making might especially be considered for novice HR professionals as these systems tend to improve the decision-making quality of less experienced decision makers (Goddard et al., 2012; Langer et al., 2021).”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites Goddard et al., 2012 source
    • cites Langer et al., 2021 source
  • Classified as probably not a substantial claim. 24
  • May combine two claims. 38
  • The quote only partly states this claim. 1

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

    Study design and conductcitation

    The sample size was determined by an a priori power analysis using G*Power, assuming power 0.80, alpha 0.05, required N=90 to detect medium-sized effects of d=0.4.

    “The sample size for the study was determined by an a priori power analysis using G*Power (Faul et al., 2009). Assuming a power of 1–β = 0.80, and an α-error of 0.05, we calculated a required sample size of 90 participants to be able to consistently detect medium-sized effects of d = 0.4.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Faul et al., 2009 source

  • Classified as probably not a substantial claim.

    Premisecitation

    Ethical standards call for human oversight of automation to address potential risks associated with AI use.

    “First, ethical standards call for human oversight of automation to address potential risks associated with AI use, e.g., system shortcomings”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites Hunkenschroer and Luetge, 2022 source

  • Classified as probably not a substantial claim.

    Interpretationcitation

    Deliberately programming errors into AI-based decision support systems and pointing them out when they are overlooked can support the attention of decision makers by varying reliability over time.

    “This way, the design of the system can support the attention of decision makers by varying reliability over time (Goddard et al., 2012; Zerilli et al., 2019).”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites Goddard et al., 2012 source

      The “Design of DSS” section says varying reliability rather than keeping it constant increased vigilant behavior and improved appropriate reliance, supporting the citing sentence’s point about attention. It does not say to deliberately program errors into a system or point out errors that decision makers overlook. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
    • cites Zerilli et al., 2019 source

      Section 5.1 (System Reliability; page unavailable) describes catch trials that deliberately generate system errors to keep human monitors attentive, and says they may engage people more actively by making a mostly dependable system less reliable. It does not describe pointing out errors when decision makers overlook them, and it cautions that more research is needed on catch trials’ efficacy. (LLM, conf=0.84)

      Experimental — accuracy not yet measured

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

    Study design and conductown data

    Conscientiousness was measured with the extra-short form of the Big-Five-Inventory-2 (Rammstedt et al., 2013) with Cronbach α = 0.76.

    “Conscientiousness was measured with the extra-short form of the Big-Five-Inventory-2 (Rammstedt et al., 2013). The scale had an acceptable internal consistency (Cronbach α = 0.76).”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Rammstedt et al., 2013 source

      The instrument description (page unavailable) identifies the ultra-short BFI-10 as a measure of the five personality dimensions, which supports its use to assess conscientiousness. It explicitly calls the BFI-10 a short version of the Big Five Inventory (BFI), not the Big Five Inventory-2. The cited paper does not establish the citing study’s reported Cronbach α of 0.76. (LLM, conf=0.88)

      Experimental — accuracy not yet measured

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

    Premisecitation

    HR professionals might utilize a system for decision support differently, as prior experience with personnel selection tasks is related to higher confidence in one's own decisions, resulting in a lower reliance on the system and less automation bias.

    “HR professionals might utilize a system for decision support differently, as prior experience with personnel selection tasks is related to higher confidence in one’s own decisions, resulting in a lower reliance on the system and less automation bias (Goddard et al., 2012; Langer et al., 2021).”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites Goddard et al., 2012 source

      The Experience and Confidence and trust sections report that experience may increase users’ confidence, that greater confidence in one’s own decisions decreases reliance on decision support, and that less experienced users can be more prone to automation bias. The review does not establish this relationship specifically for HR professionals or personnel selection tasks. (LLM, conf=0.85)

      Experimental — accuracy not yet measured
    • cites Langer et al., 2021 source

  • May combine two claims.

    Premisecitation

    Automation bias tends to occur especially when the system is perceived as highly reliable and has been working error-free for a long time (low failure rates).

    “Since automation bias tends to occur especially when the system is perceived as highly reliable (Zerilli et al., 2019) and has been working error-free for a long time, i.e., has low failure rates (Parasuraman and Manzey, 2010), companies should make decision makers aware of possible system failures not once, but on a regular base.”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites Zerilli et al., 2019 source
    • cites Parasuraman and Manzey, 2010 source

  • May combine two claims.

    Premisecitation

    Automation bias violates ethical recommendations and legal requirements like Article 22 of the GDPR or the EU AI act that call for human oversight.

    “and violates ethical recommendations (Hunkenschroer and Luetge, 2022) as well as legal requirements like Article 22 of the GDPR (The European Parliament and the Council of the European Union, 2016) or the EU AI act (European Commission, 2021) that call for human oversight.”
    Quote in the paperNot a substantial claim1 of 3 sources checked
    • cites Hunkenschroer and Luetge, 2022 source
    • cites The European Parliament and the Council of the European Union, 2016

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

    • cites European Commission, 2021

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

  • May combine two claims.

    Findingown data

    Error-awareness group spent significantly more time, had significantly more clicks, and visited significantly more pages at level 3 than the control group.

    “The error-awareness group spent significantly more time at level 3 (Mdiff = 17.00, 95% [0.88, 33.12], p = 0.04, d = 0.63), had significantly more clicks at level 3 (Mdiff = 1.59, 95% [0.11, 3.07], p = 0.04, d = 0.64) and visited significantly more pages at level 3 (Mdiff = 1.12, 95% [0.07, 3.10], p = 0.04, d = 0.55) than the control group.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The behavioral indicators (time spent, number of clicks, number of pages visited) validly measure verification intensity and automation bias.

  • May combine two claims.

    Findingown data

    For the first task, time spent at each level and number of clicks did not significantly differ between instruction groups.

    “For the first task, time spent at each level (F(2,87) = 1.20, p = 0.18) and the number of clicks (F(2,87) = 2.69, p = 0.07) did not significantly differ between the instruction groups.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The behavioral indicators (time spent, number of clicks, number of pages visited) validly measure verification intensity and automation bias.

  • May combine two claims.

    Findingown data

    Only pages visited at level 2 detail in the first task and all indicators at level 3 in the second task differed significantly between the groups.

    “While there was a tendency for all indicators at all levels, only pages visited of the detail interface of level 2 during the first task and all indicators of level 3 in the second task differed significantly between the groups.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The behavioral indicators (time spent, number of clicks, number of pages visited) validly measure verification intensity and automation bias.

  • May combine two claims. The quote only partly states this claim.

    Findingown data

    For the second task, significant differences between instruction groups were found for time spent at each level, number of clicks, and number of pages visited.

    “For the second task, significant differences between the instruction groups were found for time spent at each level (F(2,59) = 3.35, p = 0.04, partial η2 = 0.09), the number of clicks (F(2,60) = 3.32, p = 0.04, η2 = 0.07), and the number of pages visited (F(6, 172) = 2.03, p = 0.06, η2 = 0.07).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The behavioral indicators (time spent, number of clicks, number of pages visited) validly measure verification intensity and automation bias.

  • May combine two claims.

    Findingreasoning

    The same experimental group (informed about system errors) rated their subjective decision quality higher, which reflects the increased effort they put into decision-making.

    “Accordingly, the same experimental group, rated their subjective decision quality higher, which reflects the increased effort they put into decision-making.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The highly aggregated design did not lead to peripheral, heuristic information processing and, thus, less verification intensity.

    “As described above, the highly aggregated design did not lead to peripheral, heuristic information processing and, thus, less verification intensity.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning

    In our study, decision makers of both groups were able to change levels and actively access more detailed information, so both groups were able to achieve the same decision quality.

    “In our study, decision makers of both groups were able to change levels and actively access more detailed information, so both groups were able to achieve the same decision quality.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Information about system errors influences decision makers' verification intensity in the expected direction, with decision makers knowing about possible system errors seeking out more detailed information about candidates.

    “We found evidence that information about system errors influences decision makers’ verification intensity in the expected direction, with decision makers knowing about possible system errors seeking out more detailed information about candidates.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Longer time spent, greater number of clicks, and greater number of pages visited were associated with better objective decision quality.

    “In general, the longer the time spent, the greater the number of clicks and the greater the number of pages visited, the better the objective decision quality.”
    • Also stated in: § Discussion · boosted
      “Our experimental study showed that decision makers who scored lower on verification intensity indicators (i.e., less time spent on pages, lower number of clicks and pages visited), and thus had higher automation bias, selected fewer correct applicants.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The behavioral indicators (time spent, number of clicks, number of pages visited) validly measure verification intensity and automation bias.
    • Rests on an assumption: The objective decision quality score (number of correctly selected applicants) validly measures decision quality.

  • May combine two claims.

    Findingown data

    No significant differences in time spent at other levels, number of clicks, and number of visited pages were found between the data visualization groups.

    “However, no significant differences in time spent at other levels, the number of clicks and the number of visited pages were found between the data visualization groups (F(3,85) = 0.23, p = 0.87), indicating no difference in verification intensity and thus the tendency of automation bias.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The behavioral indicators (time spent, number of clicks, number of pages visited) validly measure verification intensity and automation bias.

  • Classified as probably not a substantial claim.

    Premisecitation

    ELM describes information processing as either peripheral (fast, uncritical, heuristic) or central (thorough, systematic).

    “ELM describes how information processing occurs either via the peripheral route, which is characterized by fast, uncritical and heuristic information processing, or the central route, which describes thorough and systematic information processing.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Petty and Cacioppo, 1986 source

  • Classified as probably not a substantial claim.

    Interpretationcitation

    For personnel preselection, automatically processed information and AI-based decision-making must be reviewed by humans (human-in-the-loop) unless applicants voluntarily renounce their right.

    “For personnel preselection, this means that automatically processed information and AI-based decision-making must be reviewed by humans, a so-called human-in-the-loop, unless applicants voluntarily renounce their right.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Automation bias can lead to system errors being overlooked and result in poor decision quality.

    “Automation bias can lead to system errors being overlooked and thus result in poor decision quality”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites Mosier and Skitka, 1999 source
    • cites Bahner et al., 2008 source

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

    Premisecitation

    Human oversight needs to be an active assessment and verification, not passive approval.

    “However, a passive approval of automated processing falls too short; human oversight rather needs to be an active assessment and verification”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Malgieri and Comandé, 2017 source

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

    Premisecitation

    Responsibility for overall performance or decision accuracy can reduce automation bias in flight simulations.

    “It was found that responsibility for overall performance or decision accuracy can reduce automation bias in flight simulations”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Skitka et al., 2000a

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

  • May combine two claims.

    Interpretationcitation

    People who share decision-making with AI feel less responsible and reduce cognitive effort, leading to heuristic processing and increased automation bias.

    “Consequently, people who share a decision-making task with an AI-based decision support system, feel less responsible for the decision and reduce their cognitive effort. This leads to a more heuristic and peripheral information processing, which increases automation bias”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites Skitka et al., 2000a

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

    • cites Parasuraman and Manzey, 2010 source

  • May combine two claims.

    Premisecitation

    Presenting too little or highly aggregated information can decrease transparency and limit critical elaboration of the data.

    “Conversely, presenting too little information, or information that is highly aggregated, can decrease transparency and limit critical elaboration of the data”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites Endsley, 2017 source
    • cites Sarikaya et al., 2019 source

  • May combine two claims.

    Premisecitation

    Automation bias can lead users to follow a recommendation even if it is not the best choice, resulting in poor decision quality.

    “Therefore, automation bias, including not seeking out confirmatory or contradictory information, can lead the user to follow a recommendation, even if it is not the best choice (Mosier and Skitka, 1999; Bahner et al., 2008), resulting in poor decision quality.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites Mosier and Skitka, 1999 source
    • cites Bahner et al., 2008 source

  • May combine two claims.

    Premisecitation

    A high decision quality relies on critical analysis of all applicants and selection of the best match.

    “A high decision quality relies on the critical analysis of all applicants and the selection of the applicant, who best matches the job requirements”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites Kowalczyk and Gerlach, 2015 source
    • cites Langer et al., 2021 source

  • May combine two claims.

    Premisecitation

    The format of data visualization affects how users interpret the underlying data and thereby influences their decision-making process.

    “The format of data visualization affects how users interpret the underlying data and thereby influences their decision-making process (Endsley, 2017; Sosulski, 2018).”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites Endsley, 2017 source
    • cites Sosulski, 2018 source

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

    Premisecitation

    Users in a simulated process control task reduced the verification of additional parameters over time, but further controlled for necessary parameters.

    “This is in line with previous research, where users in a simulated process control task reduced the verification of additional parameters over time, but further controlled for necessary parameters (Manzey et al., 2012).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Manzey et al., 2012 source

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

    Interpretationcitation

    Evaluation concerns might have raised feelings of responsibility in addition to those elicited in the experimental group and were thus not captured by our manipulation check.

    “Such evaluation concerns might have raised feelings of responsibility in addition to those elicited in the experimental group (that was informed about legal requirements due to the GDPR) and that were thus not captured by our manipulation check.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Study design and conductnone given

    The task was an isolated experimental task and not integrated into the stressful work situation of HR professionals.

    “Second, the task was an isolated experimental task and not integrated into the stressful work situation of HR professionals.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    Participants were randomly assigned to one of the six conditions.

    “Participants were randomly assigned to one of the six conditions.”
    Quote in the paperSubstantial claim

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

    Study design and conductreasoning

    Technical affinity and conscientiousness were measured to control for confounding factors.

    “To control for confounding factors, we measured participants’ technical affinity and conscientiousness.”
    Quote in the paperNot a substantial claim

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

    Study design and conductnone given

    A manipulation check was included after the experimental tasks to verify if the experimental manipulation was effective.

    “We included a manipulation check after the experimental tasks to verify if the experimental manipulation was effective.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    Responsibility group did not significantly differ from the control group in verification behavior; hypothesis 4 was rejected.

    “However, the responsibility group did not significantly differ from the control group in their verification behavior. Thus, hypothesis 4 had to be rejected.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Langer et al. (2021) found that decision makers who received an automated ranking before processing candidate information were less satisfied with their decision and had lower self-efficacy than those who processed information first.

    “Langer et al. (2021) also found, that decision makers who received an automated ranking of candidates before they even could process candidate information themselves, were less satisfied with their decision and had a lower feeling of self-efficacy compared to those who first processed candidate information and received an automated ranking later on.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Langer et al. (2021) source

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

    Premisecitation

    According to Yigitbasioglu and Velcu (2012), a good fit between data visualization and users' information needs, as well as a balance between complexity and utility of the information visualization, are required to enable effective information processing by dashboard users.

    “However, according to Yigitbasioglu and Velcu (2012) a good fit between data visualization and users’ information needs, as well as a balance between complexity and utility of the information visualization, are required to enable effective information processing by dashboard users.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Yigitbasioglu and Velcu (2012) source

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

    Premisecitation

    Previous studies demonstrated that decision makers who encountered system errors subsequently showed more verification behavior.

    “Previous studies demonstrated that decision makers who encountered system errors subsequently showed more verification behavior (Manzey et al., 2006; Bahner et al., 2008).”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites Manzey et al., 2006 source
    • cites Bahner et al., 2008 source

  • May combine two claims.

    Interpretationcitation

    Knowing that the system is not 100% reliable encourages users to critically check recommendations and to use the central, systematic route of information processing instead of the peripheral, heuristic route.

    “Knowing that the system is not 100% reliable encourages users to critically check recommendations and to use the central, systematic route of information processing instead of the peripheral, heuristic route (Petty and Cacioppo, 1986).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Petty and Cacioppo, 1986 source

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

    Premisecitation

    Bankins (2021) proposes that organizations must align actual AI use with its intended use by instructing employees on how to interact and rely on AI.

    “Bankins (2021) in her ethical framework for AI in human resource management proposes that organizations must align actual AI use with its intended use by instructing employees on how to interact and rely on AI.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Bankins (2021) source

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

    Study design and conductown data

    Three participants were excluded due to technical problems; four additional participants were removed for response times 1.5 SD below average in both tasks.

    “Three participants were excluded from further analysis because they reported technical problems with the dashboard at the end of the study. After carefully checking the data, we removed an additional four participants as they had response times 1.5 SD below average in both experimental tasks.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Study design and conductown data

    We controlled all analyses for consciousness and technical affinity, but found no differences.

    “We controlled all analyses for consciousness and technical affinity, but found no differences.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    Automation bias often occurs in a multitasking setting and under a high workload, as it serves as a decision-making heuristic, which saves time and cognitive effort.

    “Automation bias often occurs in a multitasking setting and under a high workload, as it serves as a decision-making heuristic, which saves time and cognitive effort (Parasuraman and Manzey, 2010; Cummings, 2017).”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites Parasuraman and Manzey, 2010 source
    • cites Cummings, 2017 source

  • May combine two claims.

    Interpretationown data

    Highly aggregated data oftentimes do not convey sufficient information to reach a decision and can lead to oversimplification and automation bias.

    “To present information as parsimoniously as possible, we suggest that highly aggregated data should be avoided because they oftentimes do not convey sufficient information to reach a decision and can lead to oversimplification and automation bias.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    Decision makers of the error-awareness group tended to score higher on verification intensity indicators, meaning they expressed less automation bias.

    “To sum up, decision makers of the error-awareness group tended to score higher on verification intensity indicators which means they expressed less automation bias.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The behavioral indicators (time spent, number of clicks, number of pages visited) validly measure verification intensity and automation bias.

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

    Study design and conductown data

    Technical affinity was measured with the Affinity for Technology Interaction Scale (Franke et al., 2019) with Cronbach α = 0.93.

    “Technical affinity was measured with the Affinity for Technology Interaction Scale (Franke et al., 2019). The questionnaire was answered on a 6-point scale ranging from (1) strongly disagree to (6) strongly agree (Cronbach α = 0.93).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Franke et al., 2019 source

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

    Premisecitation

    Possible impacts on the decision quality due to an unreflected use of the system as well as the uncritical acceptance of its recommendations might only arise under a higher workload.

    “Possible impacts on the decision quality due to an unreflected use of the system as well as the uncritical acceptance of its recommendations might only arise under a higher workload (Parasuraman and Manzey, 2010).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Parasuraman and Manzey, 2010 source

  • May combine two claims.

    Premisecitation

    In a real-life personnel preselection task, a higher number of applicants can be expected, which increases the effort for information processing and makes decision-making more difficult.

    “In a real-life personnel preselection task, a higher number of applicants can be expected, which increases the effort for information processing and makes decision-making more difficult (Black and van Esch, 2020).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Black and van Esch, 2020 source

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

    Premisecitation

    Burton et al. (2020) attribute the misuse of AI-based systems partly to unaddressed psychological needs of decision makers, like agency, autonomy or control.

    “Burton et al. (2020) attribute the misuse of AI-based systems partly to unaddressed psychological needs of decision makers, like agency, autonomy or control.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Burton et al. (2020) source

  • May combine two claims.

    Study design and conductreasoning

    The dashboard was designed so that integrated system errors could only be discovered at level 2, and level 2 must be accessed to verify the dashboard's assessment.

    “While level 1 provided an overview of all applicants, only level 2 provided enough information to thoroughly evaluate the applicants’ suitability for the position. Moreover, the integrated system errors could only be discovered at level 2. Therefore, level 2 must be accessed to verify the dashboard’s assessment.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    The highly aggregated score did not convey sufficient information to fulfil the tasks, because participants of this group quickly switched to the other levels that presented more detailed information.

    “This finding suggests that the highly aggregated score did not convey sufficient information to fulfil the tasks, because participants of this group quickly switched to the other levels that presented more detailed information.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    Informing decision makers about the actual capabilities of AI-based systems and raising their awareness of system errors encourages high verification intensity.

    “Based on our results we suggest that organizations inform decision makers about the actual capabilities of AI-based systems and raise their awareness of system errors to encourage high verification intensity.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Study design and conductnone given

    Participants were students and not actual HR professionals.

    “First, participants were students and not actual HR professionals.”
    Quote in the paperSubstantial claim

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

    Premisecitation

    Prior research shows that performance improves when decision makers are responsible for the decision-making process, but not when they are responsible for the decision-making outcome.

    “One potential explanation comes from prior research showing that performance improves when decision makers are responsible for the decision-making process, but not when they are responsible for the decision-making outcome (Doney and Armstrong, 1995; Siegel-Jacobs and Yates, 1996).”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites Doney and Armstrong, 1995 source
    • cites Siegel-Jacobs and Yates, 1996 source

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

    Premisecitation

    ELM points to individual differences, such as the need for cognition, that influence information processing.

    “ELM also points to individual differences, such as the need for cognition, that influence information processing (Petty and Cacioppo, 1986).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Petty and Cacioppo, 1986 source
54 of 54

  • The authors state this as an assumption.

    Instructing decision makers about system errors and responsibility, and providing an appropriate level of data aggregation, should encourage systematic processing and increase decision quality.

    “We assume that organizational factors, i.e., the instruction about a system regarding errors and responsibility, and technological design factors, i.e., the level of data aggregation, can decrease the heuristic processing that is prone to automation bias and can increase decision quality.”

  • The paper relies on this without support.

    The behavioral indicators (time spent, number of clicks, number of pages visited) validly measure verification intensity and automation bias.

    “Verification intensity indicators were operationalized with three verification behavior variables, i.e., time spent at each level, the number of clicks, and the number of pages visited at each level during the decision-making process.”

  • The paper relies on this without support.

    The self-developed subjective decision quality items validly measure decision quality.

    “In addition, we assessed subjective decision quality by asking participants to rate their performance on the tasks using four self-developed items (see Table 2).”

  • The paper relies on this without support.

    The objective decision quality score (number of correctly selected applicants) validly measures decision quality.

    “Objective decision quality was assessed by the number of correctly selected applicants. Five out of the ten applicants were designed to be better suited for each position than the other five applicants. Participants received one point for each correctly selected applicant, resulting in a possible score from 0 to 5.”

  • The paper relies on this without support.

    The single-item manipulation check items validly measure awareness of system errors and feeling of responsibility.

    “We included a manipulation check after the experimental tasks to verify if the experimental manipulation was effective. One item measured the effect of the warning presented to the error-awareness group. I controlled the dashboard’s assessment because I was aware that system errors might occur. Another item examined the feeling of responsibility.”

  • The paper relies on this without support.

    Student participants can act as proxies for HR professionals in personnel selection.

    “They were told to imagine themselves as HR professionals.”

  • The paper relies on this without support.

    The integrated system errors are realistic representations of errors in AI-based personnel selection systems.

    “The dashboard was designed in such a way that the decision makers could quickly make a decision using the AI-based assessment at level 1, a realistic scenario in personnel selection.”

  • The paper relies on this without support.

    Our study is the first to explore strategies to avoid automation bias in the context of personnel preselection.

    “To the best of our knowledge, our study is the first to explore strategies, that have previously been investigated in other application areas, such as raising responsibility and raising awareness about system errors (Zerilli et al., 2019), in the context of personnel preselection.”

  • The paper relies on this without support.

    Studies that previously examined AI-based personnel preselection tools from the perspective of decision makers have not yet addressed automation bias.

    “Studies that previously examined AI-based personnel preselection tools from the perspective of decision makers have not yet addressed automation bias.”
9 of 9

Statistics that don’t add up

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

  • Inconsistent p-value: F(2,87) = 1.20, p = .18

    Minor

    The paper reports p = .18 here; recomputing from the test statistic and its degrees of freedom gives p = .306. The check cannot tell whether the statistic or the p-value is the misprint.

    Reported p = .18 · recomputed p = .306

  • Inconsistent p-value: F(2,90) = 1.20, p = .32

    Minor

    The paper reports p = .32 here; recomputing from the test statistic and its degrees of freedom gives p = .306. The check cannot tell whether the statistic or the p-value is the misprint.

    Reported p = .32 · recomputed p = .306

2 of 2

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 upF(2, 87) = 1.2reported p = .18 · recomputed p = .306

    “F(2,87) = 1.20, p = .18”

  • Doesn’t add upF(2, 90) = 1.2reported p = .32 · recomputed p = .306

    “F(2,90) = 1.20, p = .32”

  • ConsistentF(2, 57) = 7.47reported p < .01 · recomputed p = .001

    “F(2,57) = 7.47, p < .01”

  • ConsistentF(2, 59) = 5.53reported p = .02 · recomputed p = .006

    “F(2,59) = 5.53, p = .02”

    The paper states a correction for multiple tests

  • ConsistentF(2, 87) = 2.69reported p = .07 · recomputed p = .074

    “F(2,87) = 2.69, p = .07”

  • ConsistentF(2, 60) = 3.59reported p = .03 · recomputed p = .034

    “F(2,60) = 3.59, p = .03”

  • ConsistentF(2, 59) = 3.35reported p = .04 · recomputed p = .042

    “F(2,59) = 3.35, p = .04”

  • ConsistentF(2, 60) = 3.32reported p = .04 · recomputed p = .043

    “F(2,60) = 3.32, p = .04”

  • ConsistentF(6, 172) = 2.03reported p = .06 · recomputed p = .064

    “F(6, 172) = 2.03, p = .06”

  • ConsistentF(2, 88) = 3.48reported p = .04 · recomputed p = .035

    “F(2,88) = 3.48, p = .04”

  • ConsistentF(2, 90) = 1.22reported p = .3 · recomputed p = .300

    “F(2,90) = 1.22, p = .30”

  • ConsistentF(2, 90) = 4.08reported p = .02 · recomputed p = .020

    “F(2,90) = 4.08, p = .02”

  • ConsistentF(2, 89) = 4.63reported p = .01 · recomputed p = .012

    “F(2,89) = 4.63, p = .01”

  • Consistentt(78) = 2.74reported p < .01 · recomputed p = .008

    “t(78) = 2.74, p < .01”

  • Consistentt(69) = 2.1reported p = .04 · recomputed p = .039

    “t(69) = 2.10, p = .04”

  • ConsistentF(3, 85) = 0.23reported p = .87 · recomputed p = .875

    “F(3,85) = 0.23, p = .87”

  • Consistentt(91) = -0.63reported p = .53 · recomputed p = .530

    “t(91) = −0.63, p = .53”

  • Consistentt(91) = 1.09reported p = .28 · recomputed p = .279

    “t(91) = 1.09, p = .28”

  • Consistentt(91) = -0.03reported p = .98 · recomputed p = .976

    “t(91) = −0.03, p = .98”

19 of 19

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.

  • Confidence interval is not centered on the reported difference

    Minor

    The reported difference of 1.12 is not the midpoint of its 95% confidence interval, 0.07 to 3.10. That interval is centered on 1.585.

    Reported 1.12 · recomputed 1.585

  • A p = 0.06 result is called significant

    Minor

    The paper describes the second-task difference in pages visited as significant despite reporting p = 0.06. Its stated significance threshold is p < 0.05.

  • Claim of higher verification on all indicators conflicts with Table 4

    Minor

    The paper says the error-awareness group tended to score higher on every indicator at every level, but Table 4 reports substantially less time at level 2 in task 1 for that group than for the control group.

3 of 3

Cites

Flagged first, then by citations

Research works printed but not linked: 1

  1. Moor, 2006, The Dartmouth College artificial intelligence conference: the next fifty years, AI Mag., 27, 87, 10.1609/aimag.v27i4.1911 · DOI

Cited by

Sorted by citations

Related

Authors
Kupfer, C., Prassl, R., Fleiß, J., Malin, C., Thalmann, S., & Kubicek, B.
Affiliations
University of Graz (AT)
Corresponding author
Kupfer, Cordula
ORCID
Venue
Frontiers in Psychology
Publisher
Frontiers Media SA
Published
2023-04-05
Volume
14
Pages
1118723-1118723
ISSN
1664-1078 (ISSN-L)
DOI
10.3389/fpsyg.2023.1118723
PMC
PMC10113449
OpenAlex
W4362626331
Semantic Scholar
e68190b4b334e3550cf2da2ca3f2e531b7964bec
Cited by
59
References
68
Citations by year
2026: 22 · 2025: 17 · 2024: 16 · 2023: 4
FWCI
17.49
Type
journal article
Language
en
Open access
Licence
Locations
  • Frontiers in Psychology (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
Karl-Franzens-Universität Graz
Topics
  • Social Sciences › Psychology › Social Psychology › Human-Automation Interaction and Safety
  • Social Sciences › Business, Management and Accounting › Organizational Behavior and Human Resource Management › AI and HR Technologies
  • Social Sciences › Safety Research › Ethics and Social Impacts of AI
Keywords
automation bias, decision support system, human oversight, elaboration likelihood model, data aggregation level, decision quality, AI ethics

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 2026-10-30.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-30.

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