Beyond binary decisions: evaluating the effects of AI error type on trust and performance in AI-assisted tasks

May combine two claims.

Varying error patterns from AI impacted performance, reaction times, and trust.

“Varying error patterns from AI impacted performance, reaction times, and trust.”

May combine two claims.

The results did not support H3b: there was a nonsignificant difference between the three patterns when participants successfully recognized the false reassurance.

“The results did not support H3b. We found a nonsignificant difference between the three patterns when participants successfully recognized the false reassurance.”

May combine two claims.

The findings pertinent to H1 can be summarized as P_success|WC approximately equals P_success|WI, and both are greater than P_success|WIref.

“The findings pertinent to H1 can be summarized as P success |WC ≈ P success |WI > P success |WI ref .”

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

As automation reliability increases, trust and dependence on automation increase accordingly.

“The general findings indicate that as automation reliability increases, trust and dependence on automation increase accordingly.”

• cites Bhat et al., 2022, 2024Not checked yet

  Not checked in the free run.

• cites Lacson et al., 2005Not checked yet

  Not checked in the free run.

• cites Madhavan et al., 2006Not checked yet

  Not checked in the free run.

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

In Case E, the AI's error could potentially lead to catastrophic results because it mistakenly confirms the incorrectly filled medication as correct, misleading the human operator and providing false reassurance.

“Conversely, Case E also involves an error by the AI, but this mistake could potentially lead to catastrophic results, as it mistakenly confirms the incorrectly filled medication as correct, misleading the human operator and providing false reassurance.”

Classified as probably not a substantial claim.

If the AI's prediction were assumed to be only binary, Cases D and E would be combined into a single case.

“Interestingly, as shown in Figure 1(a), if we assumed the prediction of the AI is only binary, Cases D and E would be combined into a single case.”

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

When operators fail to respond to events due to automation misses, omission errors arise.

“When operators fail to respond to events due to automation misses, omission errors arise (Meyer, 2001, 2004; Skitka et al., 1999).”

• cites Meyer, 2001, 2004Not checked yet

  Not checked in the free run.

• cites Skitka et al., 1999Not checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

In Case D, the AI's erroneous identification of medication Y as Z can be considered beneficial from a utilitarian perspective because it differs from the prescribed medication X and rightly triggers a safety alert.

“In Case D, although the AI erroneously identifies medication Y as Z, this error can be considered beneficial from a utilitarian perspective. The incorrect identification differs from the prescribed medication X, which rightly triggers a safety alert.”

May combine two claims.

Some AI errors prompted safety checks and verifications, which, despite causing a moderate decrease in trust, ultimately enhanced overall performance.

“Paradoxically, some AI errors prompted safety checks and verifications, which, despite causing a moderate decrease in trust, ultimately enhanced overall performance.”

May combine two claims.

Some AI errors provided false reassurance, misleading operators into believing their incorrect decisions were correct, worsening performance and trust.

“Some AI errors provided false reassurance, misleading operators into believing their incorrect decisions were correct, worsening performance and trust.”

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

SDT categorizes the world into binary states-signal present or absent.

“However, SDT categorizes the world into binary states—signal present or absent—which oversimplifies the complex dynamics and variances observed in real-world scenarios.”

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

The hypothetical medication dispensing scenario reflects recent developments in medical dispensing.

“This hypothetical scenario, which reflects recent developments in medical dispensing (Q. Chen et al., 2024; J. Y. Kim et al., 2025; Lester et al., 2021; Tsai et al., 2025; Zheng et al., 2023), presents unique characteristics that are largely overlooked in existing research on trust in and dependence on automation/autonomy and AI systems.”

• cites Q. Chen et al., 2024Not checked yet

  Not checked in the free run.

• cites J. Y. Kim et al., 2025Not checked yet

  Not checked in the free run.

• cites Lester et al., 2021Not checked yet

  Not checked in the free run.

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

The stabilization property was first empirically found in Yang, Guo, and Schemanske (2023).

“The last property, stabilization, was first empirically found in Yang, Guo, and Schemanske (2023).”

• cites Yang, Guo, and Schemanske (2023)Not checked yet

  Not checked in the free run.

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

Operators may exhibit lower trust in systems with a high frequency of false alarms because the heightened prominence of false alarms reduces credibility as operators need to actively verify the presence of errors.

“Conversely, other studies indicate that operators may exhibit lower trust in systems with a high frequency of FAs, which may be attributed to the heightened prominence of FAs that reduces the credibility as operators need to actively verify the presence of errors, making the error more salient (Breznitz, 2013; Johnson et al., 2004).”

• cites Breznitz, 2013Not checked yet

  Not checked in the free run.

• cites Johnson et al., 2004Not checked yet

  Not checked in the free run.

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

The difficulty of the stimuli was determined through a separate study using Qualtrics with participants recruited via Amazon Mechanical Turk.

“The difficulty of the stimuli was determined through a separate study using Qualtrics, and the participants were recruited via Amazon Mechanical Turk (MTurk).”

May combine two claims.

In the false-alarm condition, reliance increased as participants became more familiarized with automation, whereas reliance decreased in the miss condition.

“Results revealed a trend of increasing reliance in the FA condition as participants became more familiarized with automation, whereas reliance decreased in the miss condition (Sanchez et al., 2014).”

• cites Sanchez et al., 2014Not checked yet

  Not checked in the free run.

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

The research complied with the American Psychological Association code of ethics and was exempt from Institutional Review Board oversight at the University of Michigan.

“This research complied with the American Psychological Association code of ethics and was exempt from the Institutional Review Board oversight at the University of Michigan (HUM00230326).”

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

The Kenward-Roger method of estimating degrees of freedom mitigates Type 1 errors, generates precise p-values, and maintains reliability across varying sample sizes in mixed-effects models.

“The Kenward-Roger method of estimating degrees of freedom was used, which adjusts the variance-covariance matrix of the fixed effects using a Taylor series expansion, then calculates degrees of freedom (Kenward & Roger, 1997). This method is widely recognized in mixed-effects models for its ability to mitigate Type 1 errors, generate precise p-values, and maintain reliability across varying sample”

• cites Kenward & Roger, 1997Not checked yet

  Not checked in the free run.

• cites Kuznetsova et al., 2017Not checked yet

  Not checked in the free run.

• cites Luke, 2017Not checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

AI reliability above the baseline threshold level of around 70% is perceived as useful.

“This reliability is above the baseline threshold level of around 70% to be perceived as useful (Wickens & Dixon, 2007).”

• cites Wickens & Dixon, 2007Not checked yet

  Not checked in the free run.

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

False alarms had a stronger impact on trust than misses; trust mediated the relationship between false alarms and compliance but not between misses and reliance.

“Additionally, FAs had a stronger impact on trust than misses; trust mediated the relationship between FAs and compliance but not between misses and reliance (Chancey et al., 2017).”

• cites Chancey et al., 2017Not checked yet

  Not checked in the free run.

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

The presence of each pattern or outcome was participant-dependent, as the participant's initial and final responses could not be manipulated.

“It is important to note that the presence of each pattern or outcome was participant-dependent, as the participant’s initial and final responses could not be manipulated.”

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

Misses had a more harmful effect on trust than false alarms.

“Trust dynamics were evaluated after each interaction, revealing that misses had a more harmful effect on trust than FAs (Azevedo-Sa et al., 2020).”

• cites Azevedo-Sa et al., 2020Not checked yet

  Not checked in the free run.

May combine two claims.

Systems with more misses produced a greater trust decrement compared to false alarms; compliance was greater for the miss-prone aid, while reliance was greater for the false-alarm-prone aid.

“Trust was assessed through postsession questionnaires, revealing a greater trust decrement for systems with more misses compared to FAs. Compliance was greater for the miss-prone aid, while reliance was greater for the FA-prone aid (Davenport & Bustamante, 2010).”

• cites Davenport & Bustamante, 2010Not checked yet

  Not checked in the free run.

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

This model surpasses previous models in prediction accuracy while ensuring generalizability and explanatory power.

“This model surpasses previous models in prediction accuracy while ensuring generalizability and explanatory power, indicating the importance of the three properties of trust dynamics.”

• cites Guo & Yang, 2021Not checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

Mixed-effects models are appropriate for this analysis because they handle repeated measures and missing values.

“Mixed-effects models have advantages in dealing with repeated measures and missing values and are, therefore, particularly appropriate for our analysis.”

May combine two claims.

Excessive false alarms may affect trust more than misses, leading to disuse (neglect or under-dependence).

“However, excessive FAs may affect trust more than misses, leading to disuse (neglect or under-dependence), deviating from the original purpose (Chancey et al., 2017; Dixon & Wickens, 2006; Parasuraman & Riley, 1997).”

• cites Chancey et al., 2017Not checked yet

  Not checked in the free run.

• cites Dixon & Wickens, 2006Not checked yet

  Not checked in the free run.

• cites Parasuraman & Riley, 1997Not checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

The experiment employed a within-subjects design.

“The experiment employed a within-subjects design.”

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

A predictive computational model was developed using these properties, which claims that trust at any point follows a Beta distribution.

“Using these properties, a predictive computational model was developed, which claims that trust at any point follows a Beta distribution (Guo & Yang, 2021).”

• cites Guo & Yang, 2021Not checked yet

  Not checked in the free run.

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

Informed consent was obtained from each participant.

“Informed consent was obtained from each participant.”

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

Error type did not affect participants' trust.

“Trust was measured after each block, and their findings indicated that error type did not affect the participants’ trust (Guzman-Bonilla & Patton, 2024).”

• cites Guzman-Bonilla & Patton, 2024Not checked yet

  Not checked in the free run.

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

Data from participants who failed to match the validation pairs correctly were discarded.

“Data from participants who failed to match the validation pairs (reference-reference image pairs) correctly were discarded.”

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

The 15-second duration for initial choice was determined based on pilot testing.

“The 15 second duration was determined based on pilot testing.”

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

The 10-second duration for final choice was determined based on a pilot study.

“Similarly, the duration was determined based on a pilot study.”

May combine two claims.

False alarms primarily influence compliance, though some studies indicate they may also affect reliance, whereas misses predominantly impact reliance.

“Research suggests that FAs primarily influence compliance, though some studies indicate they may also affect reliance, whereas misses predominantly impact reliance (Chancey et al., 2015; Chancey et al., 2017; Dixon & Wickens, 2006; Dixon et al., 2007; Meyer, 2001).”

• cites Chancey et al., 2015Not checked yet

  Not checked in the free run.

• cites Chancey et al., 2017Not checked yet

  Not checked in the free run.

• cites Dixon & Wickens, 2006Not checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

The stabilization property of trust dynamics was recently reported in the studies of Yang, Guo, and Schemanske (2023) and Guo and Yang (2021).

“The stabilization property was recently reported in the studies of Yang, Guo, and Schemanske (2023) and Guo and Yang (2021).”

• cites Yang, Guo, and Schemanske (2023)Not checked yet

  Not checked in the free run.

• cites Guo and Yang (2021)Not checked yet

  Not checked in the free run.

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

Under the assumption that C_success = 1 and C_failure = −1, the calculated average consequences of the three cases are 0.24, 0.14, and −0.5.

“Furthermore, if we are willing to assume C success = 1 and C failure = −1, we could calculate the average consequences of the three cases: Consequence WC = 0.24, Consequence WC = 0.14, and Consequence WC = −0.5.”

• AssumptionThe authors state this as an assumption: The cost of success is 1 and the cost of failure is −1.

May combine two claims.

The results supported H3a: when human operators were tricked by incorrect reassurance from the AI, they reacted very fast, with WIref resulting in significantly quicker reaction time than WC and WI.

“The results supported H3a—when the human operators were tricked by the incorrect reassurance from the AI, they reacted very fast, WI ref resulting in significantly quicker reaction time than WC and WI.”

May combine two claims.

When an automated decision aid erroneously confirms a participant's wrong initial answer as correct (WIref), there is a high likelihood of making errors and participants falsely believe their initial mistakes were correct.

“This highlights the high likelihood of making errors when automated decision aid erroneously confirms participants’ wrong initial answer as correct, making participants falsely believe their initial mistakes were correct.”

May combine two claims.

Certain AI errors are beneficial, and this finding only surfaces when going beyond the binary categorization of signal detection theory.

“This counter-intuitive finding—that certain AI errors are, in fact, beneficial—only surfaces when we go beyond the binary categorization of signal detection theory (SDT).”

The quote only partly states this claim.

As a person uses an automated technology over and over again, this person's trust will stabilize over time.

“Our study provides further validation of this property—as a person uses an automated technology over and over again, this person’s trust will stabilize over time.”

May combine two claims.

Prior research has consistently shown that automation performance is a key determinant of the direction of trust change: automation success leads to trust increment and failure leads to trust decrement.

“Prior research has consistently shown that automation performance is a key determinant of the direction of trust change—automation success leads to trust increment and failure leads to trust decrement (Lee & Moray, 1992; Manzey et al., 2012; Yang et al., 2017; Yang, Schemanske, & Searle, 2023).”

• cites Lee & Moray, 1992Not checked yet

  Not checked in the free run.

• cites Manzey et al., 2012Not checked yet

  Not checked in the free run.

• cites Yang et al., 2017Not checked yet

  Not checked in the free run.

May combine two claims.

AI errors can provide false reassurance, misleading operators into believing their incorrect decisions were correct, which leads to worsened performance and a significant decline in trust.

“AI errors can provide false reassurance, misleading operators into believing their incorrect decisions were correct. This leads to worsened performance and a significant decline in trust.”