When Robots are Surprising: The Role of Cue Diagnosticity in Judging Robot Competence

When Robots are Surprising: The Role of Cue Diagnosticity in Judging Robot Competence | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article When Robots are Surprising: The Role of Cue Diagnosticity in Judging Robot Competence Nicholas Surdel, Melissa Ferguson This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-5852068/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 06 Feb, 2026 Read the published version in Scientific Reports → Version 1 posted 12 You are reading this latest preprint version Abstract Previous research showed people’s explicit (vs. implicit) competence impressions were more sensitive to a robot’s single inconsistent (“oddball”) behavior. We report nine pre-registered studies (N = 3,735 online participants) testing the scope and underlying causes of this dissociation. We found that the dissociation (a) generalized to industrial robots, surgical robots, and self-driving cars; (b) replicated with structurally aligned direct and indirect measures of competence; and (c) is at least partially explained by the inconsistent evidence’s diagnosticity. We discuss implications for social cognition and human-robot interaction. Biological sciences/Psychology Biological sciences/Psychology/Human behaviour Implicit Social Cognition Competence Human-Robot Interaction Automaticity Figures Figure 1 Figure 2 Introduction When Robots are Surprising: The Role of Cue Diagnosticity in Judging Robot Competence Competence is one of two core dimensions along which people evaluate others. 1,2 We spontaneously form competence impressions from the subtlest signals and leverage them to make consequential decisions. 3 But these signals always contain variance. An orator may follow an insightful point with word salad, just as a chef may perfectly caramelize the crust of an overcooked beef wellington. People are rarely (if ever) uniformly competent or incompetent. How do we navigate inconsistent evidence of ability? Two theoretical frameworks offer competing accounts. Dual-process theories predict people differently integrate inconsistent evidence in their explicit (i.e., directly measured, and thus relatively intentional) versus implicit (i.e., indirectly measured, and thus relatively unintentionally activated) impressions. An extensive literature appears to support this prediction. Many studies document circumstances in which explicit impressions update faster to new information than their implicit counterparts. 4–7 The dual-process framework postulates implicit impressions cannot rapidly update because they develop through a slow associative process, whereas explicit impressions can quickly update through a fast propositional process. 7–12 Thus, this perspective predicts that whereas explicit impressions will adjust according to each new piece of information learned – even if discrepant from the prior information – implicit impressions will be insensitive to infrequent, inconsistent evidence. In other words, explicit impressions should uniquely respond to oddball evidence. In contrast, an alternative framework suggests these apparent differences may instead reflect qualities of the evidence itself, particularly its diagnosticity —how indicative it is of someone's true characteristics. 13 Recent empirical work shows new counterattitudinal evidence can rapidly update both explicit and implicit impressions if the evidence is perceived as diagnostic and credible. 14–20 These researchers conclude that misalignments between explicit and implicit impressions stem from the nature of the inconsistent evidence (as well as social demands and expectations), rather than inherent differences between the two types of impressions. 21,22 This framework predicts both explicit and implicit impressions should integrate oddballs, provided they are diagnostic. 23 However, when new evidence is relatively weak or less diagnostic (which might have been the case in past work on this question), 21 this perspective predicts explicit impressions will uniquely update (due to demand considerations unique to directly measured evaluations). 24–27 Thus, this account argues that whether implicit impressions are influenced by oddball evidence depends on the evidence’s diagnosticity. Challenges with Generalizing Previous Results To date, nearly all research testing these theories focused on moral information and general (good-bad) evaluations. Although some work suggests convergence between general attitudes and traits attribution, 28 other work shows people seem to process trait competence distinctly from morality. 29 – 33 For instance, our learning systems seem to prioritize moral information: only after we understand an agent’s intentions can we contextualize their ability to act on them. 31 , 33 – 36 This prioritization may influence how diagnostic we consider different types of behavioral evidence – that is, moral behaviors may be seen as more revealing of a person's true nature than competence-related behaviors. Moreover, some accounts suggest that trait impressions (like competence) are grounded in semantic memory systems, 37 , 38 which can be more resistant to updating than the affective memory systems that underlie moral or valence-based evaluations. 39 Together, there is reason to suspect previous findings may not generalize to the competence domain. Surdel et al. 40 provides a notable exception. They tested how easily people formed and updated implicit versus explicit competence impressions of a robot. Robots can be used as sophisticated stimuli in order to test basic theoretical questions about social cognition. 41 , 42 People reason about robots similarly to humans, 43 – 45 ,cf. 46 with fewer priors about their capabilities. 47 – 49 Across six studies, Surdel et al.’s 40 participants played a series of tic-tac-toe games against a non-humanoid robot in a 2.5D virtual environment. The robot either played competently (i.e., such that it would never lose) or incompetently (i.e., such that it tried to lose). Throughout their studies, they manipulated the robot’s skill level and its performance consistency. After 3–10 games, participants completed a modified competence affect misattribution procedure (AMP) 50 , 51 to measure implicit impressions and a modified robot social attributes scale (RoSAS) 52 to measure explicit impressions. Their studies revealed two relevant findings. First, when the robot showed infrequent (20–25% of the total evidence), inconsistent evidence within the initial learning, explicit impressions uniquely incorporated it. Second, when the robot performed inconsistently between two learning sessions, explicit impressions updated to a greater extent. Notably, however, implicit impressions did update. This is likely because the inconsistent evidence was equally prevalent, and thus more diagnostic of the robot’s true capabilities. This work suggests that people’s explicit impressions of an agent’s competence routinely reflect oddball performances, whereas implicit impressions do not, producing a striking dissociation. Although these studies offer insight into how people form competence impressions (of a robot), they leave many unanswered questions. First, all their studies deployed the same robot, setting, and competence information. It is unclear whether this dissociation generalizes to different robots, contexts, and domains of competence. Second, their use of structurally different measures poses interpretational challenges. 20 , 53 – 55 Their measures differed on at least 17 dimensions (see Table 1), making it unclear which features are causally contributing to the dissociations. Structural-fit approaches improve clarity 56 and, occasionally, alter interpretations of previous findings. 55 Without a structurally-aligned design, we cannot determine which aspect(s) of the measures caused the dissociation, limiting the theoretical value of the data. Table 1 Differences between Surdel et al.’s 40 Indirect and Direct measures Feature Indirect Measure Direct Measure Instruction Complexity Careful reading and memory No instructions required Task Frequency Uncommon Extremely common Task Difficulty Hard Easy Response Format Binary Ordinal Number of Measures 15 per prime 4 per prime Unique Judgments 1 per prime 4 per prime Number of Novel Stimuli in Task 47 0 Task Sequence Shuffled trials Measures nested within blocks Basis of Evaluation Relative Absolute Stimuli per Response Four Five Ability to Revise Answers Absent Present Co-occurrent Stimuli Mask & response map Target and other scale items Stimuli Exposure Time < 100 ms ∞ ms Response times Fast (~ 1,500 ms) Slow (~ 10,000 ms) Method of Response Key press Mouse click Response Scale Unipolar Bipolar Intentionality Unintentional Intentional Type of Measurement Indirect Direct Most critically, Surdel et al. 40 did not address possible explanations for the dissociation. The new perspective on implicit and explicit updating is that both should update when the new evidence is relatively diagnostic. Perhaps the reason why implicit competence impressions were insensitive to the oddballs was because the oddballs (i.e., optimal vs. suboptimal tic-tac-toe performance) were seen as relatively nondiagnostic. Without manipulating diagnosticity, they could not directly test whether the dissociation stems from features of the evidence. The Present Research Across nine pre-registered studies, we demonstrate that this dissociation generalizes beyond the tic-tac-toe paradigm while identifying its psychological basis. We first generalize the effect to industrial robots, surgical robots, and self-driving cars (Experiments 1a-c). Using structurally-matched measures, we then demonstrate the explicit-implicit dissociation reflects differences in intentional versus unintentional responding (Experiments 2–4). Finally, we show the diagnosticity of inconsistent evidence drives these effects—when oddball behaviors are highly diagnostic of competence, both implicit and explicit impressions update (Experiments 5–7). Collectively, these data offer implications for social cognition, human-robot interaction, as well as marketing. First, these data are consistent with the theory that misalignments between implicit and explicit impressions stem from the nature of the evidence rather than inherent differences between the two types of impressions. Second, in settings in which evidence is perceived as relatively weak, one should expect competence evaluations to vary as a function of how they are measured (i.e., directly versus indirectly). As capability judgments are integral to users’ perceptions of a robot’s utility, 57 , 58 roboticists and marketeers should nest their downstream inferences within how they measure such impressions. We conducted all experiments under Yale University’s Institutional Review Board (IRB: 2000030817) in accordance with the Declaration of Helsinki. Each experiment began with participants’ informed consent and concluded with a study debriefing. Experiments 1a-c Experiments 1a–c examined whether the dissociation extends beyond simple game-playing contexts to domains where robot competence has real-world consequences. Methods We investigated three distinct settings where robots are increasingly deployed: industrial manufacturing (Experiment 1a), surgical assistance (Experiment 1b), and autonomous vehicles (Experiment 1c). These domains vary in stakes and complexity while allowing for clear markers of success and failure. We aimed to recruit 400 participants via Prolific for each experiment. Participants who either failed to engage the AMP (i.e., pressed one button), completed more than 5% of their AMP trials in either under 50 ms or over 7,500 ms, or who had incomplete data were excluded (as pre-registered). The final sample sizes were 338 participants for Experiment 1a (M age = 42 years, SD = 15; 150 males, 188 females), 362 participants for Experiment 1b (M age = 38 years, SD = 14; 178 males, 176 females), and 370 participants for Experiment 1c (M age = 39 years, SD = 13; 182 males, 183 females). Procedure In each experiment, participants were introduced to a novel target robot and a control robot, both pretested to be perceived as equally competent. The robots varied by context: Experiment 1a (Industrial Setting). Participants evaluated Boxy (target) and Shippy (control), two industrial robots. They observed Boxy performing five navigation trials through a simulated factory with randomly generated obstacles. Competent trials showed Boxy successfully navigating the path, whereas incompetent trials involved Boxy making a wrong turn and colliding with a shelf. Experiment 1b (Surgical Setting). Participants evaluated DocBot (target) and MedBot (control), two surgical robots. They observed DocBot performing five abstract navigation surgical protocols (Bezier curve, pi function, or sine wave). During competent trials, DocBot successfully followed the intended path, whereas during incompetent trials, the robot slightly deviated from the path. Experiment 1c (Autonomous Vehicle Setting). Participants evaluated Miles (target) and Rover (control), two autonomous vehicles. They observed Miles in five obstacle detection trials on a country road. Competent trials depicted Miles stopping to avoid a ball rolling across the road - preventing a collision. Incompetent trials showed Miles failing to stop, hitting the ball. Participants were randomly assigned to one of four conditions: consistently competent (five competent trials), inconsistently competent (four competent trials and one incompetent trial), inconsistently incompetent (four incompetent trials and one competent trial), or consistently incompetent (five incompetent trials). After the learning phase, participants completed a competence AMP, a 7-point bipolar self-report competence scale for both robots, exploratory items, and demographics. Like Surdel et al., 40 the AMP contained three primes (target robot, control robot, and a neural robot, i.e., Barret Technologies’ WAM Arm) and the self-report scale contained 4-items (competence, knowledgeability, reliability, capability). 52 The AMP contained 45 trials (15 per prime), and we removed any trials completed in less than 50 ms or more than 7,500 ms. We counterbalanced the robot image, name, color, and the order of explicit competence evaluations. For the exact materials used, please refer to the supplementary materials. (*A bug affected half of the first 351 participants in Experiment 1a. Those who were counterbalanced to explicitly evaluate the control robot before the target robot had all data subsequent to the AMP corrupted. This issue was resolved for the remaining 53 participants. Consequently, analyses involving explicit competence excluded data from an additional 138 participants. Importantly, all reported effects remained robust even after excluding these participants.) Results Explicit Competence Impressions To test the influence of the robot’s performance on explicit competence impressions, we ran an ANOVA with Majority Behavior (2: Competent, Incompetent) and Behavioral Consistency (2: Consistent, Inconsistent) as factors. In Experiment 1a, there was a main effect of Majority Behavior, F (1, 196) = 502, p < .001, η p 2 = .719, indicating that participants in the competent conditions rated the robot (M = 6.03, SE = .11) higher than those in the incompetent conditions (M = 2.42, SE = .11). The interaction term supported our hypothesis, F (1, 196) = 25.6, p < .001, η p 2 = .116, such that the competent oddball increased competence ratings, whereas incompetent oddballs decreased explicit competence ratings. In Experiment 1b, we saw the same pattern of results. A main effect of Majority Behavior, F (1, 358) = 446.3, p < .001, η p 2 = .555, revealing higher ratings in the competent (M = 5.66, SE = .09) than incompetent conditions (M = 3.08, SE = .09), and the same interaction, F (1, 358) = 44.6, p < .001, η p 2 = .111. We also saw a main effect of consistency, F (1, 358) = 18.5, p < .001, η p 2 = .049, suggesting the incompetent oddball was more impactful than the competent oddball. Experiment 1c was no different. The main effect of Majority Behavior, F (1, 366) = 881.3, p < .001, η p 2 = .707, showed folks in the competent condition (M = 5.65, SE = .08) rated the robot higher than those in the incompetent condition (M = 2.18, SE = .08). Although we did see the expected interaction, F (1, 366) = 15.6, p < .001, η p 2 = .040, it was driven by the incompetent oddball’s impact on impressions (as supported by an effect of consistency, F (1, 366) = 14.1, p < .001, η p 2 = .037). Implicit Competence Impressions We tested the influence of the robot’s performance on implicit competence impressions using ANCOVAs with the same factors, controlling for implicit impressions of the control robot. Each experiment revealed, as predicted, non-interactive main effects of Majority Behavior (see Table 2). Table 2 Implicit Competence Impressions for Experiments 1a–c Experiment Effect Inferential Stats Marginal Mean (SE) split by Majority Behavior 1a Majority Behavior Interaction Effect F (1, 333) = 17.7, p < .001, η p 2 = .050 F (1, 333) = 0, p = .912 .32 (.03), .11 (.03) 1b Majority Behavior Interaction Effect F (1, 357) = 11.4, p < .001, η p 2 = .031 F (1, 357) = 0, p = .911 .22 (.03), .07 (.03) 1c Majority Behavior Interaction Effect F (1, 365) = 6.8, p = .005, η p 2 = .018 F (1, 365) = 1.4, p = .237 .22 (.03), .05 (.03) Note. Green font: competent conditions, red font: incompetent conditions. Range: -1 to 1 Comparison of Explicit and Implicit Impressions To test for a dissociation, each study ran a mixed model with Majority Behavior, Behavioral Consistency, and Impression (2: Explicit, Implicit) predicting standardized competence impressions, controlling for implicit impressions of the control robot (see Fig. 1). In Experiments 1a and 1b, we saw the predicted three-way interaction, such that oddballs had a larger impact on explicit than implicit impressions. ( 1a : b = –.83, SE = .26, t (251) = − 3.22, p = .001; 1b : b = − 1.02, SE = 0.21, t (358) = − 4.84, p < .001). In Experiment 1c, the three-way interaction was not significant, b = − 0.20, SE = 0.21, t (366) = − 0.96, p = .339. We believe this is due to the competent oddball not influencing explicit impressions and random variance in implicit impressions. If we analyzed only participants who learned about a mostly competent vehicle, the Consistency × Impression Type interaction was marginally significant in the predicted direction, b = 0.27, SE = 0.14, t (350) = 1.96, p = .051. Discussion Experiments 1a-c generalized the dissociation between explicit and implicit competence impressions beyond simple game environments to contexts where competence has meaningful consequences. It is unclear, however, which of the (many) theoretically meaningful differences between the two measures is responsible (see Table 1). For instance, task fluency qualitatively affects response patterns. 59 It would be no shock to learn Prolific employees are more fluent in bipolar scales than AMPs. To make any claim about the nature of the dissociation, we must isolate the main theoretical difference between the two types of impressions: intentionality. Experiment 2 Experiments 2–4 revisited Surdel et al.’s 40 tic-tac-toe paradigm. We created a modified AMP to measure directly measured competence impressions while maintaining the AMP’s procedural features. Experiment 2 tested whether inconsistent evidence would influence directly measured competence impressions even with this disfluent instrument. We also manipulated response scales to see whether response format affected sensitivity to inconsistent evidence. Methods We recruited a total of 400 participants on Prolific. Participants with incomplete data ( n = 5) or who failed the attention check ( n = 63) were excluded (as pre-registered). Thus, we had a final sample of 332 (M age : 29 years, SD = 9; 165 males, 164 females). Most of the sample identified as White (242, 12 Asian, 51 Black, 16 Latino). Procedure This experiment followed the protocol established in Surdel et al. 40 . Participants were introduced to two robots: a novel non-humanoid robot named Taylor and a control robot named Jessie (i.e., Rethink Robotics’ Baxter). Participants were then redirected into a virtual, 2.5D office space and maneuvered (using the “w”, “a”, “s”, and “d” keys on their keyboard, as well as their mouse to rotate their field of vision) their first-person avatar to find Taylor. Once they found Taylor, they played four rounds of tic-tac-toe. The participant always went first in the first game to assure they were attending to the board before the first play occurred. Subsequent first moves were randomized. The robot either played optimally in all four games (competent condition) or optimally in three games and suboptimally in one (oddball condition). We did not include (mostly) incompetent conditions. Following the interaction, participants completed a direct AMP. Like the standard (i.e., indirect) AMP, the direct AMP presented: (a) a prime image, (b) a competence-neutral face, and (c) a noise mask until response. But unlike an indirect AMP, we explicitly instructed participants to evaluate the prime rather than the neutral face. Like the explicit competence scale, the direct AMP consisted of 4 items per prime. We manipulated the direct AMP response set between participants, such that half of participants completed a direct AMP with a binary response (e.g., “less competent” vs. “more competent) and the other half completed a direct AMP with an ordinal response (e.g., “much less competent”, “less competent”, “more competent”, and “much more competent”). Responses were submitted via key press (‘D’ and ‘K’ for the binary response condition, ‘S’, ‘D’, ‘K’, and ‘L’ for the ordinal response condition). After the direct AMP, we collected explicit competence impressions for both robots and demographics before debriefing. This experiment followed the protocol established in Surdel et al. 40 . Participants were introduced to two robots: a novel non-humanoid robot named Taylor and a control robot named Jessie (i.e., Rethink Robotics’ Baxter). Participants were then redirected into a virtual, 2.5D office space and maneuvered (using the “w”, “a”, “s”, and “d” keys on their keyboard, as well as their mouse to rotate their field of vision) their first-person avatar to find Taylor. Once they found Taylor, they played four rounds of tic-tac-toe. The participant always went first in the first game to assure they were attending to the board before the first play occurred. Subsequent first moves were randomized. The robot either played optimally in all four games (competent condition) or optimally in three games and suboptimally in one (oddball condition). We did not include (mostly) incompetent conditions. Following the interaction, participants completed a direct AMP. Like the standard (i.e., indirect) AMP, the direct AMP presented: (a) a prime image, (b) a competence-neutral face, and (c) a noise mask until response. But unlike an indirect AMP, we explicitly instructed participants to evaluate the prime rather than the neutral face. Like the explicit competence scale, the direct AMP consisted of 4 items per prime. We manipulated the direct AMP response set between participants, such that half of participants completed a direct AMP with a binary response (e.g., “less competent” vs. “more competent) and the other half completed a direct AMP with an ordinal response (e.g., “much less competent”, “less competent”, “more competent”, and “much more competent”). Responses were submitted via key press (‘D’ and ‘K’ for the binary response condition, ‘S’, ‘D’, ‘K’, and ‘L’ for the ordinal response condition). After the direct AMP, we collected explicit competence impressions for both robots and demographics before debriefing. Results Explicit Competence Impressions A Welch’s t -test, t (330) = 5.4, p < .001, d = .60, confirmed participants in the competent condition (M = 5.95, SD = .99) reported higher explicit competence impressions than those in the oddball condition (M = 5.32, SD = 1.14). Direct AMP Competence Impressions Participants in both conditions took a similar amount of time to answer each direct AMP trial (Oddball: x̄ G (RT) = 2178 ms; Competent: x̄ G (RT) = 2237 ms; p = .488). We tested the conditions’ impact on direct AMP evaluations with a Behavior (2: Competent, Oddball) by Response Method (2: Ordinal, Binary) ANOVA. Our pre-registered one-tailed test of behavior, F (1, 328) = 3.3, p = .035, η p 2 = .01, showed higher impressions in the competent (M = .68, SE = .07 versus M = .50, SE = .06). The interaction term was not significant, F (1, 328) = 1.0, p = .310, suggesting the response scale did not affect oddball sensitivity. Adding direct AMP evaluations of the control robot as a covariate did not alter the results. Lastly, we predicted procedural differences between the two measures would reduce the overall impact of the oddball on direct AMP evaluations. A Behavior by Measure (2: Direct, Explicit) mixed model predicting standardized competence impressions, controlling for Response Method, revealed a significant interaction, b = .371, SE = .11, t (330) = 3.4, p < .001. Specifically, explicit impressions were simultaneously more positive than direct AMP evaluations in the competent condition, x̄ d = .20, SE = .08, t (330) = 2.5, p = .012, and more negative than direct AMP evaluations in the oddball condition, x̄ d = − .17, SE = .07, t (330) = 2.3, p = .020, suggesting that explicit competence scores were more extreme in both conditions. Discussion Experiment 2 showed a single oddball performance could shift directly measured competence impressions even when measured under AMP-like conditions. The effect emerged regardless of response scale. However, the fact that the oddball’s influence was greater in participants’ explicit impressions than their direct AMP evaluations suggests we were justified in being concerned about the numerous differences between the two measures. This provides initial evidence the dissociation is associated with response intentionality. Experiment 3 Experiments 3 and 4 gave participants two different variants of structurally aligned direct and indirect AMPs (within-subjects). Experiment 3 aligned the instruments with Payne et al.’s 53 mouse-inputted ordinal responses, whereas Experiment 4 aligned both AMPs with key-inputted ordinal responses. Methods We recruited a total of 401 participants on Prolific. Participants who either had incomplete data ( n = 15) or who failed memory checks ( n = 21) were excluded (as pre-registered). Thus, we had a final sample of 365 (M age : 34 years, SD = 11; 183 males, 177 females). Most of the sample was White (257, 30 Asian, 35 Black, 20 Latino). Procedure This experiment followed a similar procedure to Experiment 2 with several notable exceptions. (1) Participants completed both a direct and indirect AMP (which both included the novel control robot prime, WAM Arm) in a counterbalanced order and, after each, a memory check to ensure they evaluated the correct target. (2) AMP responses were solicited as mouse inputs to a 4-point ordinal scale (-2 to 2; Less Competent to More Competent). Like Payne et al. 53 , the scale did not include a midpoint. (If we re-coded the ordinal responses to have equal steps between each response, all reported results across studies remain intact.) (3) We reintroduced the 2 (Majority Behavior: Competent, Incompetent) × 2 (Behavioral Consistency: Consistent, Inconsistent) between-subjects design. These changes resulted in the direct AMP differing from the indirect AMP in two structural ways. First, each trial’s prompt instructed participants to rate the [scale item] of the robot. Second, to avoid asking the same question about the same target multiple times, the direct AMP retained fewer trials (four per prime). We (correctly) assumed direct AMP evaluations would be less noisy than indirect AMP evaluations (see Table 3). Table 3 Experiment 3: Reliability of AMP Responses AMP All Trials Cronbach’s α First 4 Trials Cronbach's α Indirect .884 .660 Direct .878 .878 Note. These data represent the reliability scores of AMP responses separated by AMP (averaged within and then across primes). After completing the AMPs, participants completed explicit competence scales for the three robots, provided their demographics. and were debriefed. Results Unsurprisingly, explicit competence impressions were more similar to directly measured, r( 363) = .820, p < .001, than indirectly measured impressions, r( 363) = .177, p < .001. The correlation between direct and indirect measures was quite small, r( 363) = .218, p < .001. Explicit Competence Impressions A Majority Behavior by Behavioral Consistency ANOVA predicting explicit competence revealed the predicted interaction, F (1, 361) = 44.8, p < .001, η p 2 = .11, such that the oddballs influenced explicit competence in the expected ways. Directly vs. Indirectly Measured Competence Impressions To test for the dissociation, we submitted standardized direct and indirect AMP evaluations to a Measure by Majority Behavior by Behavioral Consistency mixed-effects model, controlling for impressions of the control robots. The 3-way interaction, b = .70, SE = .22, t (357) = -3.2, p = .001, η p 2 = .03, revealed directly measured impressions significantly accommodated the oddballs, but indirectly measured impressions did not, consistent with our predictions. Experiment 4 In Experiment 4, we modified the direct AMP to simplify the evaluation process by removing end-of-trial prompts and converting the four scale items to a repeatedly measured competence item. Methods We recruited a total of 476 participants on Prolific. Participants who failed either attention check ( n = 93), who failed the memory check ( n = 33), who failed to engage the indirect AMP (i.e., pressed one button; n = 20), who completed more than 5% of their AMP trials in either under 50 ms or over 7,500 ms ( n = 30), or who had incomplete data ( n = 4) were excluded (as pre-registered). Thus, we had a final sample of 294 (M age : 37 years, SD = 12; 148 males, 143 females). Most of the sample identified as White (190, 25 Asian, 43 Black, 11 Latino). Procedure This experiment’s procedure was like Experiment 3 with some changes. (1) Both AMPs now only solicited key-inputted (2) competence evaluations (3) without any end-of-trial prompts. (4) We also removed the incompetent conditions, once again. Following the AMPs, instead of eliciting explicit competence evaluations of the robots, we asked a series of behavior measures. We excluded these analyses because they did not replicate in unreported replication studies. The indirect and direct AMPs were reliable (see Table 4). Table 4 Experiment 4: Reliability of AMP Responses AMP Cronbach’s α Indirect .772 Direct .856 Note. These data represent the reliability scores of AMP responses separated by AMP (averaged within and then across primes). Results Directly and indirectly measured competence impressions were slightly correlated, r( 292) = .161, p = .006. Although participants rated the target robot more quickly in the indirect AMP (M = 1084, SE = 35) compared with the direct AMP (M = 1263, SE = 35), t (292) = 4.3, p < .001, this was likely a practicing effect. Direct vs. Indirect Competence Impressions To test for a dissociation, we ran a General Linear Model with Measure and Behavior predicting standardized impressions of the target robot while controlling for impressions of the control robots. The significant interaction, b = .22, SE = .10, t (291) = 2.2, p = .030, η p 2 = .02, revealed that while directly measured impressions were influenced by the oddball, t (573) = 2.2, p = .032, d = .09, indirectly measured impressions were not, t (573) = − .72, p = .475. Discussion In Experiments 2–4, participants’ directly measured impressions reliably and uniquely incorporated a single oddball. We believe this is strong evidence that the dissociation is related to the intentionality of the evaluation. But why didn’t implicit impressions shift? One explanation is that implicit competence impressions - by their very nature - are unable to incorporate infrequent, inconsistent performances. Provided the oddball remains odd (i.e., a statistical anomaly), it cannot be relevant. An alternative explanation is that the inconsistent evidence was not sufficiently diagnostic. If the perceiver only believed the oddball was indicative of the robot’s true capabilities, implicit impressions would incorporate it even if it remained statistically rare. To test these two explanations, Experiment 5 directly manipulated the diagnosticity of the inconsistent information. Experiment 5 Experiment 5 manipulated the inconsistent evidence’s diagnosticity with the industrial robot context in an impression updating setting. The first learning session consisted of 4 competent trials, whereas the second session consisted of 1–4 incompetent trials - the more voluminous the errors, the more diagnostic the inconsistent information. We also informed participants that previous samples found the initial learning insufficient to judge the robot, thereby increasing the diagnosticity of the second learning sessions in all conditions. Methods We recruited a total of 501 participants on Prolific. We excluded those who failed to engage the either AMP (i.e., pressed one button; n = 39) or who completed more than 5% of their AMP trials in either under 50 ms or over 7,500 ms ( n = 18; as pre-registered). In addition, a Google API data writing bug (Error: 429) caused some participants’ explicit data and demographics to not be recorded ( n = 68). Thus, we had a final sample of 444 (M age : 40 years, SD = 14; 195 males, 177 females). Most the sample was White (308, 36 Asian, 9 Black, 1 Latino). Procedure This experiment revisited Experiment 1a’s industrial robot paradigm with some changes. We added a blocked design with two learning and two measurement phases. The first learning phase always showed 4 competent trials, whereas the second learning phase showed either 1, 2, 3, or 4 incompetent trials. Participants completed an AMP and explicit responses after each learning phase. The second learning phase began by telling participants that “[p]revious participants found that 4 navigation tests were not enough to determine how reliable and competent Boxy is. Therefore, you will now view additional tests.” We also added an exploratory “How much did you learn about Boxy’s capabilities” (Not very much - A great deal) slider question after each explicit competence measurement. Lastly, we changed the post-trial feedback to “SUCCESS” and “FAILURE” rather than “No Errors” and “Errors.” Results Competence Impression Updating First, we tested for updating with paired samples t -tests contrasting initial and updated implicit and explicit impressions. Both explicit, t (375) = -29.6, p < .001, d = -1.53, and implicit impressions robustly updated, t (443) = -5.5, p < .001, d = − .26, even when controlling for the control robot (via mixed-effects model). Notably, even those in the 1 new piece of evidence condition updated their implicit impressions, t (112) = -1.9, p = .028, d = − .182, one-tailed . This is the first evidence showing a single oddball could shift implicit competence impressions. Explicit vs. Implicit Competence Impression Updating To test for a dissociation, we ran a mixed-effects model with impression and condition predicting standardized updating scores (Time 1- Time 2). As predicted, we saw a significant interaction, \(\:{\chi\:}^{2}\left(3\right)=22.2,\:p<\:.001,\:{{{\eta\:}}_{p}}^{2}=\:.06\) , such that explicit impressions updated based on the volume of new evidence, whereas implicit impressions updated non-linearly (see Fig. 2). Specifically, implicit impressions indifferently updated between 1, 2, and 3 new pieces of evidence, but spiked when the new evidence met parity. Discussion Experiment 5 showed that implicit competence impressions incorporate inconsistent information if it is diagnostic. As predicted, explicit updated in proportion to the volume of new evidence, whereas implicit did not. Notably, we found that a single piece of new evidence did result in implicit updating. Before we discuss this effect, we first sought to replicate it. Experiment 6 Methods We recruited a total of 201 participants on Prolific. We excluded those with incomplete data ( n = 2), who failed to engage the either AMP (i.e., pressed one button; n = 17) or who completed more than 5% of their AMP trials in either under 50 ms or over 7,500 ms ( n = 26; as pre-registered). Thus, we had a final sample of 175 (M age : 41 years, SD = 15; 90 males, 84 females). Most of the sample identified as White (148, 9 Asian, 7 Black, 2 Latino). Procedure This experiment was a direct replication of the 1 piece of new information condition from Experiment 5. Results Competence Impression Updating Our pre-registered one-tailed paired t -test revealed significant updating, t (174) = -1.7, p = .049, d = − .17. When we controlled for impressions of the control robot (via mixed-effects model), the effect became stronger, t (173) = -2, p = .024, d = − .18. Discussion Experiments 5 and 6 demonstrated a single oddball can influence implicit competence impressions. But why only in the updating (and not formation) studies? In the updating studies, we explicitly described the oddball as necessary to determine the robot’s capabilities. We expected this would ratchet up its diagnosticity. To confirm this explanation, we directly manipulated the diagnosticity of the oddball in an updating setting. Experiment 7 Experiment 7 tested whether the evidence’s diagnosticity caused implicit impressions to update. We predicted that implicit impressions would be more sensitive to a single failure described as recent versus outdated. Methods We recruited a total of 556 participants on Prolific. We excluded those with incomplete data ( n = 3), who failed to engage the either AMP (i.e., pressed one button; n = 35), or who completed more than 5% of their AMP trials in either under 50 ms or over 7,500 ms ( n = 30), or who failed the memory check ( n = 72; as pre-registered). Thus, we had a final sample of 416 (M age : 39 years, SD = 14; 206 males, 203 females). Most of the sample identified as White (345, 21 Asian, 23 Black, 3 Latino). Procedure This experiment revisited Experiment 6’s paradigm with an error diagnosticity factor (2: low, high) factor. Prior to the first learning session, participants read, “[t]he first set presents Boxy’s [most recent / preliminary] attempts at this task. These tests are [highly relevant and directly / outdated but] show Boxy’s [current capabilities / beginnings].” And then prior to the second learning session, participants read, “Now that you saw [Boxy’s most relevant, recent performances / the irrelevant, outdated evidence], you will see [an irrelevant earlier attempt / Boxy’s most relevant, recent test]. The final test occurred [before / after] Boxy was upgraded, and [is therefore outdated, irrelevant, and does not reflect / thus provides the best look at] Boxy’s true capabilities.” Following each learning session, participants completed an AMP and explicit competence scales for both robots. After the second explicit competence measure, they answered a binary memory check ("Which test session showed you the most recently measured performance?") and a manipulation check ("How much did the final test teach you about the robot?"; 1 = not very much, 7 = a great deal) before demographics and debriefing. Results Manipulation Check Participants in the high error diagnosticity condition (M = 5.3, SD = 1.4) reported more learning from the error than those in the low error diagnosticity condition (M = 3.5, SD = 2), t (283) = 10.2, p < .001, d = 1.08. Competence Impression Updating Participants in the high diagnosticity condition (M = .15, SD = .56) updated their implicit competence impressions more than those in the low diagnosticity condition (M = .02, SD = .45), t (409) = 2.6, p = .005, d = .25, one-tailed , even when we controlled for the control robot (via ANCOVA). Unsurprisingly, we find similar effects (of a larger magnitude) for explicit updating. Participants in the high diagnosticity condition (M = 2.6, SD = 1.6) updated their impressions more than those in the low diagnosticity condition (M = .6, SD = 1.2), t (412) = 14.9, p < .001, d = 1.45. Discussion Experiment 7 was the first experiment to demonstrate diagnosticity impacts implicit competence impressions. To date, the relationship between diagnosticity and implicit impressions was exclusively studied with general evaluations and moral information. 15 , 60 These results challenge purely process-based accounts of the dissociation, suggesting instead that the nature of the evidence itself shapes how much it influences implicit impressions. General Discussion Recent work revealed a novel dissociation in how people form and update competence impressions (of a robot; Surdel et al. 40 ). Specifically, people’s explicit impressions were more sensitive to inconsistency than their implicit impressions. In the present paper, we conducted nine experiments identifying the psychological mechanisms of this dissociation. Our results show that the dissociation (a) generalized to the ecologically relevant settings of industrial robots, surgical robots, and autonomous vehicles, (b) stems from differences in intentional versus unintentional processing, and (c) reflects the diagnosticity of inconsistent evidence rather than constraints inherent to implicit impressions. Together, these findings simultaneously teach us about how humans process competence information while also offering practical insights for robot deployment. Theoretical Contributions These data offer four key insights to the social cognitive literature. First, we demonstrate robust generalization of an explicit-implicit dissociation across diverse contexts. Psychologists often underemphasize generalizability, causing an overextrapolation of narrow findings. 61 While Surdel et al. 40 documented this pattern in a game-based tic-tac-toe setting, we show it persists across industrial navigation tasks, precise surgical operations, and autonomous driving scenarios - domains where competence judgments drive real-world consequences. Second, we provided strong evidence the dissociation is associated with the judgment’s intentionality rather than structural misalignment between the two types of measures. 20 , 53 – 56 By clearly connecting the dissociation to the measure’s intentionality, we identify specific circumstances in which competence impressions should be sensitive to relatively undiagnostic information (i.e., intentionally measured judgments) and when they should not (i.e., unintentionally measured judgments). This offers practical insights for psychologists, marketeers, and engineers. Third, we bridge theoretical frameworks between moral and competence domains. Previous research demonstrated that implicit global impressions rapidly updated in the face of highly diagnostic evidence. 15 , 60 We show parallel effects in competence impressions—when a robot's error was perceived as diagnostic of its true capabilities, implicit impressions readily incorporated it. This suggests that common mechanisms may govern how people update competence and morality impressions. Finally, these findings help reconcile competing theoretical perspectives. Traditional dual-process theories argue implicit impressions develop through slow associative learning, resistant to rapid updating. 4 , 11 Our findings support a newer account 15 , 60 where both explicit and implicit impressions are sensitive to evidence quality but differ in their evidentiary thresholds. When new evidence is perceived as sufficiently diagnostic, both implicit and explicit impressions rapidly update. But when new evidence is relatively weak, implicit impressions show slower updating. Why, then, did explicit impressions respond to weakly diagnostic evidence? The exchange of information via intentional evaluation is rife with conversational implicatum and demand. 24 – 27 , 62 First, conversational norms necessitate the responder to treat the various inputs as though they are relevant to the output - almost an assumed diagnosticity. Suppose you ask your neighbor if his nephew is reliable and she replies, “He owns a cow, what do you think?” Although you likely lack an association between cow ownership and reliability, you – being a well-socialized communicator – will assume relevance and harvest diagnosticity (e.g., cows require regular care, thus he’s reliable). Second, intentional measures may trigger unique expectations about which evidence participants should incorporate. 21 , 22 Had your neighbor merely replied, “He owns a cow,” perhaps the hunt for diagnosticity may never have commenced. But the mere presence of the question about competence (of the target) informs the conversational partner that the evidence should be sufficient to answer the question. Practical Implications These findings offer direct guidance for robot deployment in real-world settings. First, they highlight that measurement approach matters—direct and indirect measures may yield divergent conclusions about a robot’s capabilities, particularly when performance varies. It is an open question whether explicit or implicit impressions better predict user behavior, but previous work suggests they both might in some circumstances, though more research is needed on this question. 63 – 67 ,cf. 68 Second, our findings lend specific strategies for managing users’ impressions of their robot (as tools, coworkers, or products). Rather than attempting to eliminate all inconsistencies—an impossible goal—manufacturers should focus on preventing their robots from committing highly diagnostic errors that impact both explicit and implicit impressions. When errors occur, actively reducing their perceived diagnosticity through appropriate framing (e.g., highlighting situational factors or system updates) may help preserve trust. Notably, however, it is likely challenging to anticipate exactly which actions (for which perceivers) will be perceived as (non)diagnostic. Limitations and Future Directions Several limitations suggest important directions for future research. First, although we demonstrated generalization across robot types and contexts, all our studies used relatively simple, discrete tasks with clear success criteria and feedback. Future work should examine whether similar patterns emerge in tasks where competence assessment is less straightforward, such as dynamic surgical settings. It would be fascinating to see how an expert surgeon learns about a surgical robot who performs well, but not perfectly (e.g., 20% bigger incision than necessary). Second, our studies focused on impressions formed over a single session. Longitudinal research is needed to understand how explicit-implicit dissociations evolve over repeated interactions and whether they predict important outcomes like trust, reliance, and acceptance. This is particularly relevant for understanding how people adapt to learning robots whose performance may improve over time. Finally, although we identified diagnosticity as a key mechanism, other factors likely influence the integration of inconsistent evidence. These could include the perceived intentionality of robot actions, 69 the presence of explanations, 70 and users' construal of the robot's role. 46 Understanding these moderators could help organizations better manage human-robot interactions. Conclusion A growing wave of psychologists are looking to robots as tools for studying basic social cognition. 41 , 42 Separately, computer science and marketing departments are becoming increasingly interested in how people judge robots. Our research shows that when inconsistent evidence about a robot is relatively nondiagnostic, explicit and implicit impressions will dissociate. The present work contributes to at least four theoretical debates in social cognition, while also lending interdisciplinary insights. As automation continues to expand, understanding competence impression formation becomes increasingly relevant to daily life. Declarations Data Availability All data, materials, and code can be found at: https://osf.io/zwued/?view_only=9bc9367ab70e4935b278972efb662a1c Acknowledgements We thank John Bargh, Margaret Clark, the Implicit Social Cognition Lab at Yale, OpenAI’s ChatGPT, and Anthropic’s Claude for helpful comments on an earlier version of the manuscript. We also thank Malte Jung and Wen-Ying Lee for their software support, and the Office of Naval Research for funding this research (Award Number: N00014-19-1-2299). Author Contributions N.S. and M.F. equally contributed to project conceptualization and methodology. N.S. led data collection and analyses, and drafted the initial manuscript text. M.F. supervised, acquired funding, and revised the manuscript text. Both authors reviewed the manuscript. Competing Interests The authors declare no competing interests. References Fiske, S. T., Cuddy, A. J. & Glick, P. Universal dimensions of social cognition: Warmth and competence. Trends Cogn. Sci. 11 , 77–83 (2007). Todorov, A. & Oh, D. The structure and perceptual basis of social judgments from faces. Adv. Exp. Soc. Psychol. 63 , 189–245 (2021). Todorov, A., Mandisodza, A. N., Goren, A. & Hall, C. C. Inferences of Competence from Faces Predict Election Outcomes. Science 308 , 1623–1626 (2005). Gawronski, B. & Sritharan, R. Formation, change, and contextualization of mental associations: Determinants and principles of variations in implicit measures. in Handbook of implicit social cognition: Measurement, theory, and applications 216–240 (The Guilford Press, New York, NY, US, 2010). Gregg, A. P., Seibt, B. & Banaji, M. R. Easier done than undone: Asymmetry in the malleability of implicit preferences. J. Pers. Soc. Psychol. 90 , 1–20 (2006). Peters, K. R. & Gawronski, B. Are We Puppets on a String? Comparing the Impact of Contingency and Validity on Implicit and Explicit Evaluations. Pers. Soc. Psychol. Bull. 37 , 557–569 (2011). Rydell, R. J. & McConnell, A. R. Understanding implicit and explicit attitude change: A systems of reasoning analysis. J. Pers. Soc. Psychol. 91 , 995–1008 (2006). Cao, J. & Banaji, M. R. The base rate principle and the fairness principle in social judgment. Proc. Natl. Acad. Sci. 113 , 7475–7480 (2016). Gawronski, B. & Bodenhausen, G. V. Associative and propositional processes in evaluation: An integrative review of implicit and explicit attitude change. Psychol. Bull. 132 , 692–731 (2006). Gawronski, B. & Bodenhausen, G. V. The associative–propositional evaluation model: Theory, evidence, and open questions. in Advances in experimental social psychology vol. 44 59–127 (Elsevier, 2011). Gawronski, B., Luke, D. M. & Creighton, L. A. Dual-Process Theories. in The Oxford Handbook of Social Cognition, Second Edition (eds. Carlston, D. E., Hugenberg, K. & Johnson, K. L.) 319–353 (Oxford University Press, 2024). doi:10.1093/oxfordhb/9780197763414.013.12. McConnell, A. R. & Rydell, R. J. The systems of evaluation model. in Dual-process theories of the social mind (eds. Sherman, J. W., Gawronski, B. & Trope, Y.) 204–217 (New York: Guilford Press, 2014). Skowronski, J. J. & Carlston, D. E. Negativity and extremity biases in impression formation: A review of explanations. Psychol. Bull. 105 , 131–142 (1989). Brannon, S. M. & Gawronski, B. A Second Chance for First Impressions? Exploring the Context-(In)Dependent Updating of Implicit Evaluations. Soc. Psychol. Personal. Sci. 8 , 275–283 (2017). Cone, J. & Ferguson, M. J. He did what? The role of diagnosticity in revising implicit evaluations. J. Pers. Soc. Psychol. 108 , 37–57 (2015). Kurdi, B., Mann, T. C. & Ferguson, M. J. Persuading the Implicit Mind: Changing Negative Implicit Evaluations With an 8-Minute Podcast. Soc. Psychol. Personal. Sci. 13 , 688–697 (2022). Mann, T. C. & Ferguson, M. J. Can we undo our first impressions? The role of reinterpretation in reversing implicit evaluations. J. Pers. Soc. Psychol. 108 , 823–849 (2015). Mann, T. C., Cone, J., Heggeseth, B. & Ferguson, M. J. Updating implicit impressions: New evidence on intentionality and the affect misattribution procedure. J. Pers. Soc. Psychol. 116 , 349–374 (2019). Shen, X. & Ferguson, M. J. How resistant are implicit impressions of facial trustworthiness? When new evidence leads to durable updating. J. Exp. Soc. Psychol. 97 , 104219 (2021). Van Dessel, P., Cone, J., Gast, A. & De Houwer, J. The impact of valenced verbal information on implicit and explicit evaluation: the role of information diagnosticity, primacy, and memory cueing. Cogn. Emot. 34 , 74–85 (2020). Ferguson, M. J., Mann, T. C., Cone, J. & Shen, X. When and How Implicit First Impressions Can Be Updated. Curr. Dir. Psychol. Sci. 28 , 331–336 (2019). Ferguson, M. J., Mann, T. C., Cone, J. & Shen, X. The role of context in implicit and explicit evaluation. (2025). Shen, X., Mann, T. C. & Ferguson, M. J. Beware a dishonest face: Updating face-based implicit impressions using diagnostic behavioral information. J. Exp. Soc. Psychol. 86 , 103888 (2020). Corneille, O. & Lush, P. Sixty Years After Orne’s American Psychologist Article: A Conceptual Framework for Subjective Experiences Elicited by Demand Characteristics. Personal. Soc. Psychol. Rev. 27 , 83–101 (2023). Orne, M. T. Communication by the total experimental situation: Why it is important, how it is evaluated, and its significance for the ecological validity of findings. in Communication and affect (eds. Pliner, P., Krames, L. & Alloway, T.) 157–191 (Elsevier, 1973). Orne, M. T. On the social psychology of the psychological experiment: With particular reference to demand characteristics and their implications. Am. Psychol. 17 , 776–783 (1962). Schwarz, N. Measurement as cooperative communication: What research participants learn from questionnaires. in The SAGE handbook of measurement 43–59 (Sage Thousand Oaks, CA, 2010). Kurdi, B., Mann, T. C., Charlesworth, T. E. S. & Banaji, M. R. The relationship between implicit intergroup attitudes and beliefs. Proc. Natl. Acad. Sci. 116 , 5862–5871 (2019). Skowronski, J. J. & Carlston, D. E. Social judgment and social memory: The role of cue diagnosticity in negativity, positivity, and extremity biases. J. Pers. Soc. Psychol. 52 , 689–699 (1987). Wojciszke, B. Morality and competence in person- and self-perception. Eur. Rev. Soc. Psychol. 155–188 (2005). Landy, J. F., Piazza, J. & Goodwin, G. P. When It’s Bad to Be Friendly and Smart: The Desirability of Sociability and Competence Depends on Morality. Pers. Soc. Psychol. Bull. 42 , 1272–1290 (2016). Mende-Siedlecki, P., Baron, S. G. & Todorov, A. Diagnostic Value Underlies Asymmetric Updating of Impressions in the Morality and Ability Domains. J. Neurosci. 33 , 19406–19415 (2013). Luttrell, A., Sacchi, S. & Brambilla, M. Changing impressions in competence-oriented domains: The primacy of morality endures. J. Exp. Soc. Psychol. 98 , 104246 (2022). Asch, S. E. Forming impressions of personality. J. Abnorm. Soc. Psychol. 41 , 258–290 (1946). Brambilla, M., Carraro, L., Castelli, L. & Sacchi, S. Changing impressions: Moral character dominates impression updating. J. Exp. Soc. Psychol. 82 , 64–73 (2019). Brambilla, M., Sacchi, S., Rusconi, P. & Goodwin, G. P. The primacy of morality in impression development: Theory, research, and future directions. in Advances in Experimental Social Psychology vol. 64 187–262 (Elsevier, 2021). Amodio, D. M. & Devine, P. G. Stereotyping and evaluation in implicit race bias: Evidence for independent constructs and unique effects on behavior. J. Pers. Soc. Psychol. 652–661 (2006). Amodio, D. M. & Ratner, K. G. A Memory Systems Model of Implicit Social Cognition. Curr. Dir. Psychol. Sci. 20 , 143–148 (2011). Amodio, D. M. & Hamilton, H. K. Intergroup anxiety effects on implicit racial evaluation and stereotyping. Emotion 12 , (2012). Surdel, N. et al. Judging Robot Ability: How People Form Implicit and Explicit Impressions of Robot Competence. J. Exp. Psychol. Gen. 153 , 1309–1335 (2024). Wykowska, A. Robots as mirrors of the human mind. Curr. Dir. Psychol. Sci. 30 , 34–40 (2021). Wykowska, A., Chaminade, T. & Cheng, G. Embodied artificial agents for understanding human social cognition. Philos. Trans. R. Soc. B Biol. Sci. 371 , 20150375 (2016). Bigman, Y., Surdel, N. & Ferguson, M. J. Trait attribution explains human-robot interactions. Behav. Brain Sci. 46 , E23 (2023). Cross, E. S. & Ramsey, R. Mind Meets Machine: Towards a Cognitive Science of Human–Machine Interactions. Trends Cogn. Sci. 25 , 200–212 (2021). De Graaf, M. M. & Malle, B. F. How people explain action (and autonomous intelligent systems should too). in 2017 AAAI Fall Symposium Series 19–26 (2017). Clark, H. H. & Fischer, K. Social robots as depictions of social agents. Behav. Brain Sci. 46 , 1–65 (2023). Gray, K. & Wegner, D. M. Feeling robots and human zombies: Mind perception and the uncanny valley. Cognition 125 , 125–130 (2012). Bonezzi, A., Ostinelli, M. & Melzner, J. The human black-box: The illusion of understanding human better than algorithmic decision-making. J. Exp. Psychol. Gen. 151 , 2250–2258 (2022). Tully, S., Longoni, C. & Appel, G. Knowledge of Artificial Intelligence Predicts Lower AI Receptivity . https://osf.io/t9u8g (2023) doi:10.31234/osf.io/t9u8g. Payne, B. K., Cheng, C. M., Govorun, O. & Stewart, B. D. An inkblot for attitudes: Affect misattribution as implicit measurement. J. Pers. Soc. Psychol. 89 , 277–293 (2005). Cone, J., Flaharty, K. & Ferguson, M. J. Believability of evidence matters for correcting social impressions. Proc. Natl. Acad. Sci. 116 , 9802–9807 (2019). Carpinella, C. M., Wyman, A. B., Perez, M. A. & Stroessner, S. J. The Robotic Social Attributes Scale (RoSAS): Development and Validation. in Proceedings of the 2017 ACM/IEEE International Conference on Human-Robot Interaction 254–262 (ACM, Vienna Austria, 2017). doi:10.1145/2909824.3020208. Payne, B. K., Burkley, M. A. & Stokes, M. B. Why do implicit and explicit attitude tests diverge? The role of structural fit. J. Pers. Soc. Psychol. 94 , 16–31 (2008). Payne, B. K. et al. Intention Invention and the Affect Misattribution Procedure: Reply to Bar-Anan and Nosek (2012). Pers. Soc. Psychol. Bull. 39 , 375–386 (2013). Béna, J., Melnikoff, D. E., Mierop, A. & Corneille, O. Revisiting dissociation hypotheses with a structural fit approach: The case of the prepared reflex framework. J. Exp. Soc. Psychol. 100 , 104297 (2022). Corneille, O. & Gawronski, B. Self-reports are better measurement instruments than implicit measures. Nat. Rev. Psychol. 3 , 835–846 (2024). Hancock, P. A. et al. A Meta-Analysis of Factors Affecting Trust in Human-Robot Interaction. Hum. Factors 53 , 517–527 (2011). Baker, A. L., Phillips, E. K., Ullman, D. & Keebler, J. R. Toward an Understanding of Trust Repair in Human-Robot Interaction: Current Research and Future Directions. ACM Trans. Interact. Intell. Syst. 8 , 1–30 (2018). Logan, G. D. Toward an instance theory of automatization. Psychol. Rev. 95 , 492–527 (1988). Cone, J., Mann, T. C. & Ferguson, M. J. Changing Our Implicit Minds: How, When, and Why Implicit Evaluations Can Be Rapidly Revised. in Advances in Experimental Social Psychology vol. 56 131–199 (Elsevier, 2017). Yarkoni, T. The generalizability crisis. Behav. Brain Sci. 45 , e1 (2022). Grice, H. P. Logic and conversation. in Speech acts 41–58 (Brill, 1975). Cameron, C. D., Brown-Iannuzzi, J. L. & Payne, B. K. Sequential Priming Measures of Implicit Social Cognition: A Meta-Analysis of Associations With Behavior and Explicit Attitudes. Personal. Soc. Psychol. Rev. 16 , 330–350 (2012). Kurdi, B. et al. Relationship between the Implicit Association Test and intergroup behavior: A meta-analysis. Am. Psychol. 74 , 569–586 (2019). Perugini, M., Richetin, J. & Zogmaister, C. Prediction of behavior. in Handbook of Implicit Social Cognition: Measurement, Theory, and Applications (eds. Gawronski, B. & Payne, B. K.) 255–277 (New York: Guilford Press, 2010). Friese, M., Hofmann, W. & Schmitt, M. When and why do implicit measures predict behaviour? Empirical evidence for the moderating role of opportunity, motivation, and process reliance. Eur. Rev. Soc. Psychol. 19 , 285–338 (2008). Greenwald, A. G., Poehlman, T. A., Uhlmann, E. L. & Banaji, M. R. Understanding and Using the Implicit Association Test: III. Meta-Analysis of Predictive Validity. J. Pers. Soc. Psychol. 97 , 17–41 (2009). Oswald, F. L., Mitchell, G., Blanton, H., Jaccard, J. & Tetlock, P. E. Predicting Ethnic and Racial Discrimination: A Meta-Analysis of IAT Criterion Studies. J. Pers. Soc. Psychol. 105 , 171–192 (2013). Malle, B. F. Attribution theories: How people make sense of behavior. in Theories in social psychology vol. 23 93–119 (Wiley-Blackwell, 2022). Arrieta, A. B. et al. Explainable Artificial Intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible AI. Inf. Fusion 58 , 82–115 (2020). Additional Declarations No competing interests reported. Supplementary Files SurdelFergusonSupplementaryMaterials.docx Cite Share Download PDF Status: Published Journal Publication published 06 Feb, 2026 Read the published version in Scientific Reports → Version 1 posted Editorial decision: Revision requested 23 May, 2025 Reviews received at journal 02 May, 2025 Reviews received at journal 01 May, 2025 Reviews received at journal 26 Apr, 2025 Reviewers agreed at journal 21 Apr, 2025 Reviewers agreed at journal 17 Apr, 2025 Reviewers agreed at journal 16 Apr, 2025 Reviewers agreed at journal 15 Apr, 2025 Reviewers invited by journal 03 Feb, 2025 Editor assigned by journal 29 Jan, 2025 Submission checks completed at journal 28 Jan, 2025 First submitted to journal 28 Jan, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-5852068","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":415405110,"identity":"3957ca32-6755-4dca-9056-45546e877351","order_by":0,"name":"Nicholas Surdel","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABAklEQVRIiWNgGAWjYDACZgY2BoYCKIeHwUIOKNQAFsevxQCuRcKYgYGRgBYGNC2JDYS0mLMzP3vwwcDOnoH/jNmHNxUS6RuOH2x8wFBhndiAQ4tlM5u54QyD5MQGiRzjmXPOSORuOJPYbMBwJh2nFoPDPGzSPAbMCQwSPMbMvG1ALQcS2yQY2w4T0lIPchhQyz+JdIPzD9t/MP4jqOUw0Ms5QC0NEgkGNxLbgCGATwubmeQMg+NAx6QVM845JmE488bDZomEY+nGOLWcP/xM4kNFtT0//+HNDG9qbOT5zicf/PChxloWlxY4YIMxFA4AiQRCylGAPEHTR8EoGAWjYKQBACTdUbnUy7AHAAAAAElFTkSuQmCC","orcid":"","institution":"Yale University","correspondingAuthor":true,"prefix":"","firstName":"Nicholas","middleName":"","lastName":"Surdel","suffix":""},{"id":415405111,"identity":"08a35d2e-eea3-4b32-baa2-acb152d1d0a2","order_by":1,"name":"Melissa Ferguson","email":"","orcid":"","institution":"Yale University","correspondingAuthor":false,"prefix":"","firstName":"Melissa","middleName":"","lastName":"Ferguson","suffix":""}],"badges":[],"createdAt":"2025-01-18 00:08:04","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-5852068/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-5852068/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1038/s41598-026-35375-y","type":"published","date":"2026-02-06T15:59:35+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":78248293,"identity":"4621cefd-2f2c-44b9-85b5-7a1080293b93","added_by":"auto","created_at":"2025-03-11 09:38:43","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":1194520,"visible":true,"origin":"","legend":"\u003cp\u003eSee image above for figure legend\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-5852068/v1/2f26cb8e5c0d9bf116308b8b.png"},{"id":78249207,"identity":"6b98166a-b9a0-4957-b812-a4cd1faa5285","added_by":"auto","created_at":"2025-03-11 09:46:43","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":159592,"visible":true,"origin":"","legend":"\u003cp\u003eExperiment 5: Impression Updating by Amount of New Evidence\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eNote. \u003c/em\u003eExperiment 5 (\u003cem\u003eN\u003c/em\u003e = 444) standardized competence updating of the target robot predicted by the amount of new evidence (between-subjects) while controlling for competence updating of the control robots. Error bars denote 95% confidence intervals of the marginal means.\u003c/p\u003e","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-5852068/v1/a3e1333728cc969dc5516b85.png"},{"id":102234300,"identity":"aa927fb3-9360-4642-932e-4570530f540f","added_by":"auto","created_at":"2026-02-09 16:09:30","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2984072,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-5852068/v1/5842e9a2-42a2-41a6-949e-429983cb68c1.pdf"},{"id":78246656,"identity":"bdbe9392-ee34-4078-b6a4-9d27424f23bd","added_by":"auto","created_at":"2025-03-11 09:30:44","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":8505404,"visible":true,"origin":"","legend":"","description":"","filename":"SurdelFergusonSupplementaryMaterials.docx","url":"https://assets-eu.researchsquare.com/files/rs-5852068/v1/829032da1eff3d396623e413.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"When Robots are Surprising: The Role of Cue Diagnosticity in Judging Robot Competence","fulltext":[{"header":"Introduction","content":"\u003cp\u003eWhen Robots are Surprising: The Role of Cue Diagnosticity in Judging Robot Competence\u003c/p\u003e\n\u003cp\u003eCompetence is one of two core dimensions along which people evaluate others.\u003csup\u003e1,2\u003c/sup\u003e We spontaneously form competence impressions from the subtlest signals and leverage them to make consequential decisions.\u003csup\u003e3\u003c/sup\u003e But these\u003cem\u003e\u0026nbsp;\u003c/em\u003esignals always contain variance. An orator may follow an insightful point with word salad, just as a chef may perfectly caramelize the crust of an overcooked beef wellington. People are rarely (if ever) uniformly competent or incompetent. How do we navigate inconsistent evidence of ability?\u003c/p\u003e\n\u003cp\u003eTwo theoretical frameworks offer competing accounts. Dual-process theories predict people differently integrate inconsistent evidence in their explicit (i.e., directly measured, and thus relatively intentional) versus implicit (i.e., indirectly measured, and thus relatively unintentionally activated) impressions. An extensive literature appears to support this prediction. Many studies document circumstances in which explicit impressions update faster to new information than their implicit counterparts.\u003csup\u003e4\u0026ndash;7\u003c/sup\u003e The dual-process framework postulates implicit impressions cannot rapidly update because they develop through a slow associative process, whereas explicit impressions can quickly update through a fast propositional process.\u003csup\u003e7\u0026ndash;12\u003c/sup\u003e Thus, this perspective predicts that whereas explicit impressions will adjust according to each new piece of information learned \u0026ndash; even if discrepant from the prior information \u0026ndash; implicit impressions will be insensitive to infrequent, inconsistent evidence. In other words, explicit impressions should uniquely respond to oddball evidence.\u003c/p\u003e\n\u003cp\u003eIn contrast, an alternative framework suggests these\u003cem\u003e\u0026nbsp;\u003c/em\u003eapparent differences may instead reflect qualities of the evidence itself, particularly its \u003cem\u003ediagnosticity\u003c/em\u003e\u0026mdash;how indicative it is of someone\u0026apos;s true characteristics.\u003csup\u003e13\u003c/sup\u003e Recent empirical work shows new counterattitudinal evidence can rapidly update both explicit and implicit impressions \u003cem\u003eif\u003c/em\u003e the evidence is perceived as diagnostic and credible.\u003csup\u003e14\u0026ndash;20\u003c/sup\u003e These\u003cem\u003e\u0026nbsp;\u003c/em\u003eresearchers conclude that misalignments between explicit and implicit impressions stem from the nature of the inconsistent evidence (as well as social demands and expectations), rather than inherent differences between the two types of impressions.\u003csup\u003e21,22\u003c/sup\u003e This framework predicts both explicit and implicit impressions should integrate oddballs, provided they are diagnostic.\u003csup\u003e23\u003c/sup\u003e However, when new evidence is relatively weak or less diagnostic (which might have been the case in past work on this question),\u003csup\u003e21\u003c/sup\u003e this perspective predicts explicit impressions will uniquely update (due to demand considerations unique to directly measured evaluations).\u003csup\u003e24\u0026ndash;27\u003c/sup\u003e Thus, this account argues that whether implicit impressions are influenced by oddball evidence depends on the evidence\u0026rsquo;s diagnosticity.\u0026nbsp;\u003c/p\u003e\n\u003ch3\u003eChallenges with Generalizing Previous Results\u003c/h3\u003e\n\u003cp\u003eTo date, nearly all research testing these theories focused on moral information and general (good-bad) evaluations. Although some work suggests convergence between general attitudes and traits attribution,\u003csup\u003e\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u003c/sup\u003e other work shows people seem to process trait competence distinctly from morality.\u003csup\u003e\u003cspan additionalcitationids=\"CR30 CR31 CR32\" citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u003c/sup\u003e For instance, our learning systems seem to prioritize moral information: only after we understand an agent\u0026rsquo;s intentions can we contextualize their ability to act on them.\u003csup\u003e\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e,\u003cspan additionalcitationids=\"CR34 CR35\" citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e\u003c/sup\u003e This prioritization may influence how diagnostic we consider different types of behavioral evidence \u0026ndash; that is, moral behaviors may be seen as more revealing of a person's true nature than competence-related behaviors. Moreover, some accounts suggest that trait impressions (like competence) are grounded in semantic memory systems,\u003csup\u003e\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e,\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e\u003c/sup\u003e which can be more resistant to updating than the affective memory systems that underlie moral or valence-based evaluations.\u003csup\u003e\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e\u003c/sup\u003e Together, there is reason to suspect previous findings may not generalize to the competence domain.\u003c/p\u003e \u003cp\u003eSurdel et al.\u003csup\u003e\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e\u003c/sup\u003e provides a notable exception. They tested how easily people formed and updated implicit versus explicit competence impressions of a robot. Robots can be used as sophisticated stimuli in order to test basic theoretical questions about social cognition.\u003csup\u003e\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e,\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e\u003c/sup\u003e People reason about robots similarly to humans,\u003csup\u003e\u003cspan additionalcitationids=\"CR44\" citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e,cf. \u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e\u003c/sup\u003e with fewer priors about their capabilities.\u003csup\u003e\u003cspan additionalcitationids=\"CR48\" citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e\u003c/sup\u003e Across six studies, Surdel et al.\u0026rsquo;s\u003csup\u003e\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e\u003c/sup\u003e participants played a series of tic-tac-toe games against a non-humanoid robot in a 2.5D virtual environment. The robot either played competently (i.e., such that it would never lose) or incompetently (i.e., such that it tried to lose). Throughout their studies, they manipulated the robot\u0026rsquo;s skill level and its performance consistency. After 3\u0026ndash;10 games, participants completed a modified competence affect misattribution procedure (AMP)\u003csup\u003e\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e,\u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e\u003c/sup\u003e to measure implicit impressions and a modified robot social attributes scale (RoSAS)\u003csup\u003e\u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e52\u003c/span\u003e\u003c/sup\u003e to measure explicit impressions. Their studies revealed two relevant findings. First, when the robot showed infrequent (20\u0026ndash;25% of the total evidence), inconsistent evidence \u003cem\u003ewithin\u003c/em\u003e the initial learning, explicit impressions uniquely incorporated it. Second, when the robot performed inconsistently \u003cem\u003ebetween\u003c/em\u003e two learning sessions, explicit impressions updated to a greater extent. Notably, however, implicit impressions \u003cem\u003edid\u003c/em\u003e update. This is likely because the inconsistent evidence was equally prevalent, and thus more diagnostic of the robot\u0026rsquo;s true capabilities. This work suggests that people\u0026rsquo;s explicit impressions of an agent\u0026rsquo;s competence routinely reflect oddball performances, whereas implicit impressions do not, producing a striking dissociation.\u003c/p\u003e \u003cp\u003eAlthough these studies offer insight into how people form competence impressions (of a robot), they leave many unanswered questions. First, all their studies deployed the same robot, setting, and competence information. It is unclear whether this dissociation generalizes to different robots, contexts, and domains of competence. Second, their use of structurally different measures poses interpretational challenges.\u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e,\u003cspan additionalcitationids=\"CR54\" citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e55\u003c/span\u003e\u003c/sup\u003e Their measures differed on at least 17 dimensions (see Table\u0026nbsp;1), making it unclear which features are causally contributing to the dissociations. Structural-fit approaches improve clarity\u003csup\u003e\u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e56\u003c/span\u003e\u003c/sup\u003e and, occasionally, alter interpretations of previous findings.\u003csup\u003e\u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e55\u003c/span\u003e\u003c/sup\u003e Without a structurally-aligned design, we cannot determine which aspect(s) of the measures caused the dissociation, limiting the theoretical value of the data.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eDifferences between Surdel et al.\u0026rsquo;s\u003csup\u003e\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e\u003c/sup\u003e Indirect and Direct measures\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFeature\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIndirect Measure\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDirect Measure\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eInstruction Complexity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCareful reading and memory\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNo instructions required\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTask Frequency\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eUncommon\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eExtremely common\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTask Difficulty\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHard\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eEasy\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eResponse Format\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBinary\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOrdinal\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNumber of Measures\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e15 per prime\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4 per prime\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUnique Judgments\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1 per prime\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4 per prime\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNumber of Novel \u003c/p\u003e \u003cp\u003eStimuli in Task\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e47\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTask Sequence\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eShuffled trials\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMeasures nested within blocks\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBasis of Evaluation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRelative\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAbsolute\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStimuli per Response\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFour\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eFive\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAbility to Revise Answers\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAbsent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePresent\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCo-occurrent Stimuli\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMask \u0026amp; response map\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTarget and other scale items\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStimuli Exposure Time\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;100 ms\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026infin; ms\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eResponse times\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFast (~\u0026thinsp;1,500 ms)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSlow (~\u0026thinsp;10,000 ms)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMethod of Response\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eKey press\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMouse click\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eResponse Scale\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eUnipolar\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eBipolar\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eIntentionality\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eUnintentional\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eIntentional\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eType of Measurement\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIndirect\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDirect\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eMost critically, Surdel et al.\u003csup\u003e\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e\u003c/sup\u003e did not address possible explanations for the dissociation. The new perspective on implicit and explicit updating is that both should update when the new evidence is relatively diagnostic. Perhaps the reason why implicit competence impressions were insensitive to the oddballs was because the oddballs (i.e., optimal vs. suboptimal tic-tac-toe performance) were seen as relatively nondiagnostic. Without manipulating diagnosticity, they could not directly test whether the dissociation stems from features of the evidence.\u003c/p\u003e\n\u003ch3\u003eThe Present Research\u003c/h3\u003e\n\u003cp\u003eAcross nine pre-registered studies, we demonstrate that this dissociation generalizes beyond the tic-tac-toe paradigm while identifying its psychological basis. We first generalize the effect to industrial robots, surgical robots, and self-driving cars (Experiments 1a-c). Using structurally-matched measures, we then demonstrate the explicit-implicit dissociation reflects differences in intentional versus unintentional responding (Experiments 2–4). Finally, we show the diagnosticity of inconsistent evidence drives these effects—when oddball behaviors are highly diagnostic of competence, both implicit and explicit impressions update (Experiments 5–7).\u003c/p\u003e \u003cp\u003eCollectively, these data offer implications for social cognition, human-robot interaction, as well as marketing. First, these data are consistent with the theory that misalignments between implicit and explicit impressions stem from the nature of the evidence rather than inherent differences between the two types of impressions. Second, in settings in which evidence is perceived as relatively weak, one should expect competence evaluations to vary as a function of how they are measured (i.e., directly versus indirectly). As capability judgments are integral to users’ perceptions of a robot’s utility,\u003csup\u003e\u003cspan citationid=\"CR57\" class=\"CitationRef\"\u003e57\u003c/span\u003e,\u003cspan citationid=\"CR58\" class=\"CitationRef\"\u003e58\u003c/span\u003e\u003c/sup\u003e roboticists and marketeers should nest their downstream inferences within \u003cem\u003ehow\u003c/em\u003e they measure such impressions.\u003c/p\u003e \u003cp\u003e We conducted all experiments under Yale University’s Institutional Review Board (IRB: 2000030817) in accordance with the Declaration of Helsinki. Each experiment began with participants’ informed consent and concluded with a study debriefing.\u003c/p\u003e "},{"header":"Experiments 1a-c","content":"\u003cp\u003eExperiments 1a–c examined whether the dissociation extends beyond simple game-playing contexts to domains where robot competence has real-world consequences.\u003c/p\u003e\n\u003ch3\u003eMethods\u003c/h3\u003e\n\u003cp\u003eWe investigated three distinct settings where robots are increasingly deployed: industrial manufacturing (Experiment 1a), surgical assistance (Experiment 1b), and autonomous vehicles (Experiment 1c). These domains vary in stakes and complexity while allowing for clear markers of success and failure.\u003c/p\u003e \u003cp\u003e We aimed to recruit 400 participants via Prolific for each experiment. Participants who either failed to engage the AMP (i.e., pressed one button), completed more than 5% of their AMP trials in either under 50 ms or over 7,500 ms, or who had incomplete data were excluded (as pre-registered). The final sample sizes were 338 participants for Experiment 1a \u003cem\u003e(M\u003c/em\u003e\u003csub\u003e\u003cem\u003eage\u003c/em\u003e\u003c/sub\u003e = 42 years, \u003cem\u003eSD\u003c/em\u003e\u0026thinsp;=\u0026thinsp;15; 150 males, 188 females), 362 participants for Experiment 1b \u003cem\u003e(M\u003c/em\u003e\u003csub\u003e\u003cem\u003eage\u003c/em\u003e\u003c/sub\u003e = 38 years, \u003cem\u003eSD\u003c/em\u003e\u0026thinsp;=\u0026thinsp;14; 178 males, 176 females), and 370 participants for Experiment 1c \u003cem\u003e(M\u003c/em\u003e\u003csub\u003e\u003cem\u003eage\u003c/em\u003e\u003c/sub\u003e = 39 years, \u003cem\u003eSD\u003c/em\u003e\u0026thinsp;=\u0026thinsp;13; 182 males, 183 females).\u003c/p\u003e\n\u003ch3\u003eProcedure\u003c/h3\u003e\n\u003cp\u003eIn each experiment, participants were introduced to a novel target robot and a control robot, both pretested to be perceived as equally competent. The robots varied by context:\u003c/p\u003e \u003cp\u003e \u003cb\u003eExperiment 1a (Industrial Setting).\u003c/b\u003e Participants evaluated Boxy (target) and Shippy (control), two industrial robots. They observed Boxy performing five navigation trials through a simulated factory with randomly generated obstacles. Competent trials showed Boxy successfully navigating the path, whereas incompetent trials involved Boxy making a wrong turn and colliding with a shelf.\u003c/p\u003e \u003cp\u003e \u003cb\u003eExperiment 1b (Surgical Setting).\u003c/b\u003e Participants evaluated DocBot (target) and MedBot (control), two surgical robots. They observed DocBot performing five abstract navigation surgical protocols (Bezier curve, pi function, or sine wave). During competent trials, DocBot successfully followed the intended path, whereas during incompetent trials, the robot slightly deviated from the path.\u003c/p\u003e \u003cp\u003e \u003cb\u003eExperiment 1c (Autonomous Vehicle Setting).\u003c/b\u003e Participants evaluated Miles (target) and Rover (control), two autonomous vehicles. They observed Miles in five obstacle detection trials on a country road. Competent trials depicted Miles stopping to avoid a ball rolling across the road - preventing a collision. Incompetent trials showed Miles failing to stop, hitting the ball.\u003c/p\u003e \u003cp\u003eParticipants were randomly assigned to one of four conditions: consistently competent (five competent trials), inconsistently competent (four competent trials and one incompetent trial), inconsistently incompetent (four incompetent trials and one competent trial), or consistently incompetent (five incompetent trials). After the learning phase, participants completed a competence AMP, a 7-point bipolar self-report competence scale for both robots, exploratory items, and demographics. Like Surdel et al.,\u003csup\u003e\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e\u003c/sup\u003e the AMP contained three primes (target robot, control robot, and a neural robot, i.e., Barret Technologies\u0026rsquo; WAM Arm) and the self-report scale contained 4-items (competence, knowledgeability, reliability, capability).\u003csup\u003e\u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e52\u003c/span\u003e\u003c/sup\u003e The AMP contained 45 trials (15 per prime), and we removed any trials completed in less than 50 ms or more than 7,500 ms.\u003c/p\u003e \u003cp\u003eWe counterbalanced the robot image, name, color, and the order of explicit competence evaluations. For the exact materials used, please refer to the supplementary materials.\u003c/p\u003e \u003cp\u003e(*A bug affected half of the first 351 participants in Experiment 1a. Those who were counterbalanced to explicitly evaluate the control robot before the target robot had all data subsequent to the AMP corrupted. This issue was resolved for the remaining 53 participants. Consequently, analyses involving explicit competence excluded data from an additional 138 participants. Importantly, all reported effects remained robust even after excluding these participants.)\u003c/p\u003e\n\u003ch3\u003eResults\u003c/h3\u003e\n\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003eExplicit Competence Impressions\u003c/h2\u003e \u003cp\u003eTo test the influence of the robot\u0026rsquo;s performance on explicit competence impressions, we ran an ANOVA with Majority Behavior (2: Competent, Incompetent) and Behavioral Consistency (2: Consistent, Inconsistent) as factors.\u003c/p\u003e \u003cp\u003eIn Experiment 1a, there was a main effect of Majority Behavior, \u003cem\u003eF\u003c/em\u003e(1, 196)\u0026thinsp;=\u0026thinsp;502, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;.001, η\u003csub\u003ep\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.719, indicating that participants in the competent conditions rated the robot \u003cem\u003e(M\u003c/em\u003e\u0026thinsp;=\u0026thinsp;6.03, \u003cem\u003eSE\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.11) higher than those in the incompetent conditions \u003cem\u003e(M\u003c/em\u003e\u0026thinsp;=\u0026thinsp;2.42, \u003cem\u003eSE\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.11). The interaction term supported our hypothesis, \u003cem\u003eF\u003c/em\u003e(1, 196)\u0026thinsp;=\u0026thinsp;25.6, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;.001, η\u003csub\u003ep\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.116, such that the competent oddball increased competence ratings, whereas incompetent oddballs decreased explicit competence ratings.\u003c/p\u003e \u003cp\u003eIn Experiment 1b, we saw the same pattern of results. A main effect of Majority Behavior, \u003cem\u003eF\u003c/em\u003e(1, 358)\u0026thinsp;=\u0026thinsp;446.3, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;.001, η\u003csub\u003ep\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.555, revealing higher ratings in the competent \u003cem\u003e(M\u003c/em\u003e\u0026thinsp;=\u0026thinsp;5.66, \u003cem\u003eSE\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.09) than incompetent conditions \u003cem\u003e(M\u003c/em\u003e\u0026thinsp;=\u0026thinsp;3.08, \u003cem\u003eSE\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.09), and the same interaction, \u003cem\u003eF\u003c/em\u003e(1, 358)\u0026thinsp;=\u0026thinsp;44.6, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;.001, η\u003csub\u003ep\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.111. We also saw a main effect of consistency, \u003cem\u003eF\u003c/em\u003e(1, 358)\u0026thinsp;=\u0026thinsp;18.5, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;.001, η\u003csub\u003ep\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.049, suggesting the incompetent oddball was more impactful than the competent oddball.\u003c/p\u003e \u003cp\u003eExperiment 1c was no different. The main effect of Majority Behavior, \u003cem\u003eF\u003c/em\u003e(1, 366)\u0026thinsp;=\u0026thinsp;881.3, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;.001, η\u003csub\u003ep\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.707, showed folks in the competent condition \u003cem\u003e(M\u003c/em\u003e\u0026thinsp;=\u0026thinsp;5.65, \u003cem\u003eSE\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.08) rated the robot higher than those in the incompetent condition \u003cem\u003e(M\u003c/em\u003e\u0026thinsp;=\u0026thinsp;2.18, \u003cem\u003eSE\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.08). Although we did see the expected interaction, \u003cem\u003eF\u003c/em\u003e(1, 366)\u0026thinsp;=\u0026thinsp;15.6, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;.001, η\u003csub\u003ep\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.040, it was driven by the incompetent oddball\u0026rsquo;s impact on impressions (as supported by an effect of consistency, \u003cem\u003eF\u003c/em\u003e(1, 366)\u0026thinsp;=\u0026thinsp;14.1, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;.001, η\u003csub\u003ep\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.037).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eImplicit Competence Impressions\u003c/h2\u003e \u003cp\u003eWe tested the influence of the robot\u0026rsquo;s performance on implicit competence impressions using ANCOVAs with the same factors, controlling for implicit impressions of the control robot. Each experiment revealed, as predicted, non-interactive main effects of Majority Behavior (see Table\u0026nbsp;2).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eImplicit Competence Impressions for Experiments 1a\u0026ndash;c\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eExperiment\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEffect\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eInferential Stats\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eMarginal Mean (SE) split by Majority Behavior\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1a\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMajority Behavior\u003c/p\u003e \u003cp\u003eInteraction Effect\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cem\u003eF\u003c/em\u003e(1, 333)\u0026thinsp;=\u0026thinsp;17.7, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;.001, \u003cem\u003eη\u003c/em\u003e\u003csub\u003e\u003cem\u003ep\u003c/em\u003e\u003c/sub\u003e\u003csup\u003e\u003cem\u003e2\u003c/em\u003e\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.050\u003c/p\u003e \u003cp\u003e\u003cem\u003eF\u003c/em\u003e(1, 333)\u0026thinsp;=\u0026thinsp;0, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.912\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e.32 (.03), .11 (.03)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1b\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMajority Behavior\u003c/p\u003e \u003cp\u003eInteraction Effect\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cem\u003eF\u003c/em\u003e(1, 357)\u0026thinsp;=\u0026thinsp;11.4, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;.001, \u003cem\u003eη\u003c/em\u003e\u003csub\u003e\u003cem\u003ep\u003c/em\u003e\u003c/sub\u003e\u003csup\u003e\u003cem\u003e2\u003c/em\u003e\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.031\u003c/p\u003e \u003cp\u003e\u003cem\u003eF\u003c/em\u003e(1, 357)\u0026thinsp;=\u0026thinsp;0, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.911\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e.22 (.03), .07 (.03)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1c\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMajority Behavior\u003c/p\u003e \u003cp\u003eInteraction Effect\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cem\u003eF\u003c/em\u003e(1, 365)\u0026thinsp;=\u0026thinsp;6.8, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.005, \u003cem\u003eη\u003c/em\u003e\u003csub\u003e\u003cem\u003ep\u003c/em\u003e\u003c/sub\u003e\u003csup\u003e\u003cem\u003e2\u003c/em\u003e\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.018\u003c/p\u003e \u003cp\u003e\u003cem\u003eF\u003c/em\u003e(1, 365)\u0026thinsp;=\u0026thinsp;1.4, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.237\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e.22 (.03), .05 (.03)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cem\u003eNote.\u003c/em\u003e Green font: competent conditions, red font: incompetent conditions. Range: -1 to 1\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eComparison of Explicit and Implicit Impressions\u003c/h3\u003e\n\u003cp\u003eTo test for a dissociation, each study ran a mixed model with Majority Behavior, Behavioral Consistency, and Impression (2: Explicit, Implicit) predicting standardized competence impressions, controlling for implicit impressions of the control robot (see Fig.\u0026nbsp;1).\u003c/p\u003e \u003cp\u003eIn Experiments 1a and 1b, we saw the predicted three-way interaction, such that oddballs had a larger impact on explicit than implicit impressions. (\u003cb\u003e1a\u003c/b\u003e: \u003cem\u003eb\u003c/em\u003e = \u0026ndash;.83, \u003cem\u003eSE\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.26, \u003cem\u003et\u003c/em\u003e(251) = \u0026minus;\u0026thinsp;3.22, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.001; \u003cb\u003e1b\u003c/b\u003e: \u003cem\u003eb\u003c/em\u003e = \u0026minus;\u0026thinsp;1.02, \u003cem\u003eSE\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.21, \u003cem\u003et\u003c/em\u003e(358) = \u0026minus;\u0026thinsp;4.84, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;.001). In Experiment 1c, the three-way interaction was not significant, \u003cem\u003eb\u003c/em\u003e = \u0026minus;\u0026thinsp;0.20, \u003cem\u003eSE\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.21, \u003cem\u003et\u003c/em\u003e(366) = \u0026minus;\u0026thinsp;0.96, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.339. We believe this is due to the competent oddball not influencing explicit impressions and random variance in implicit impressions. If we analyzed only participants who learned about a mostly competent vehicle, the Consistency \u0026times; Impression Type interaction was marginally significant in the predicted direction, \u003cem\u003eb\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.27, \u003cem\u003eSE\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.14, \u003cem\u003et\u003c/em\u003e(350)\u0026thinsp;=\u0026thinsp;1.96, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.051.\u003c/p\u003e\n\u003ch3\u003eDiscussion\u003c/h3\u003e\n\u003cp\u003eExperiments 1a-c generalized the dissociation between explicit and implicit competence impressions beyond simple game environments to contexts where competence has meaningful consequences.\u003c/p\u003e \u003cp\u003eIt is unclear, however, which of the (many) theoretically meaningful differences between the two measures is responsible (see Table\u0026nbsp;1). For instance, task fluency qualitatively affects response patterns.\u003csup\u003e\u003cspan citationid=\"CR59\" class=\"CitationRef\"\u003e59\u003c/span\u003e\u003c/sup\u003e It would be no shock to learn Prolific employees are more fluent in bipolar scales than AMPs. To make any claim about the nature of the dissociation, we must isolate the main theoretical difference between the two types of impressions: intentionality.\u003c/p\u003e "},{"header":"Experiment 2","content":"\u003cp\u003eExperiments 2–4 revisited Surdel et al.’s\u003csup\u003e\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e\u003c/sup\u003e tic-tac-toe paradigm. We created a modified AMP to measure directly measured competence impressions while maintaining the AMP’s procedural features. Experiment 2 tested whether inconsistent evidence would influence directly measured competence impressions even with this disfluent instrument. We also manipulated response scales to see whether response format affected sensitivity to inconsistent evidence.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e\u003cp\u003e We recruited a total of 400 participants on Prolific. Participants with incomplete data (\u003cem\u003en\u003c/em\u003e = 5) or who failed the attention check (\u003cem\u003en\u003c/em\u003e = 63) were excluded (as pre-registered). Thus, we had a final sample of 332 \u003cem\u003e(M\u003c/em\u003e\u003csub\u003eage\u003c/sub\u003e: 29 years, \u003cem\u003eSD\u003c/em\u003e = 9; 165 males, 164 females). Most of the sample identified as White (242, 12 Asian, 51 Black, 16 Latino).\u003c/p\u003e\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eProcedure\u003c/h2\u003e \u003cp\u003eThis experiment followed the protocol established in Surdel et al.\u003csup\u003e\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e\u003c/sup\u003e. Participants were introduced to two robots: a novel non-humanoid robot named Taylor and a control robot named Jessie (i.e., Rethink Robotics’ Baxter). Participants were then redirected into a virtual, 2.5D office space and maneuvered (using the “w”, “a”, “s”, and “d” keys on their keyboard, as well as their mouse to rotate their field of vision) their first-person avatar to find Taylor. Once they found Taylor, they played four rounds of tic-tac-toe. The participant always went first in the first game to assure they were attending to the board before the first play occurred. Subsequent first moves were randomized. The robot either played optimally in all four games (competent condition) or optimally in three games and suboptimally in one (oddball condition). We did not include (mostly) incompetent conditions.\u003c/p\u003e \u003cp\u003e Following the interaction, participants completed a direct AMP. Like the standard (i.e., indirect) AMP, the direct AMP presented: (a) a prime image, (b) a competence-neutral face, and (c) a noise mask until response. But unlike an indirect AMP, we explicitly instructed participants to evaluate the prime rather than the neutral face. Like the explicit competence scale, the direct AMP consisted of 4 items per prime. We manipulated the direct AMP response set between participants, such that half of participants completed a direct AMP with a binary response (e.g., “less competent” vs. “more competent) and the other half completed a direct AMP with an ordinal response (e.g., “much less competent”, “less competent”, “more competent”, and “much more competent”). Responses were submitted via key press (‘D’ and ‘K’ for the binary response condition, ‘S’, ‘D’, ‘K’, and ‘L’ for the ordinal response condition).\u003c/p\u003e \u003cp\u003eAfter the direct AMP, we collected explicit competence impressions for both robots and demographics before debriefing.\u003c/p\u003e \u003c/div\u003e\u003cp\u003eThis experiment followed the protocol established in Surdel et al.\u003csup\u003e\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e\u003c/sup\u003e. Participants were introduced to two robots: a novel non-humanoid robot named Taylor and a control robot named Jessie (i.e., Rethink Robotics’ Baxter). Participants were then redirected into a virtual, 2.5D office space and maneuvered (using the “w”, “a”, “s”, and “d” keys on their keyboard, as well as their mouse to rotate their field of vision) their first-person avatar to find Taylor. Once they found Taylor, they played four rounds of tic-tac-toe. The participant always went first in the first game to assure they were attending to the board before the first play occurred. Subsequent first moves were randomized. The robot either played optimally in all four games (competent condition) or optimally in three games and suboptimally in one (oddball condition). We did not include (mostly) incompetent conditions.\u003c/p\u003e\u003cp\u003e Following the interaction, participants completed a direct AMP. Like the standard (i.e., indirect) AMP, the direct AMP presented: (a) a prime image, (b) a competence-neutral face, and (c) a noise mask until response. But unlike an indirect AMP, we explicitly instructed participants to evaluate the prime rather than the neutral face. Like the explicit competence scale, the direct AMP consisted of 4 items per prime. We manipulated the direct AMP response set between participants, such that half of participants completed a direct AMP with a binary response (e.g., “less competent” vs. “more competent) and the other half completed a direct AMP with an ordinal response (e.g., “much less competent”, “less competent”, “more competent”, and “much more competent”). Responses were submitted via key press (‘D’ and ‘K’ for the binary response condition, ‘S’, ‘D’, ‘K’, and ‘L’ for the ordinal response condition).\u003c/p\u003e\u003cp\u003eAfter the direct AMP, we collected explicit competence impressions for both robots and demographics before debriefing.\u003c/p\u003e\n\u003ch3\u003eResults\u003c/h3\u003e\n\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eExplicit Competence Impressions\u003c/h2\u003e \u003cp\u003eA Welch\u0026rsquo;s \u003cem\u003et\u003c/em\u003e-test, \u003cem\u003et\u003c/em\u003e(330)\u0026thinsp;=\u0026thinsp;5.4, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;.001, \u003cem\u003ed\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.60, confirmed participants in the competent condition \u003cem\u003e(M\u003c/em\u003e\u0026thinsp;=\u0026thinsp;5.95, \u003cem\u003eSD\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.99) reported higher explicit competence impressions than those in the oddball condition \u003cem\u003e(M\u003c/em\u003e\u0026thinsp;=\u0026thinsp;5.32, \u003cem\u003eSD\u003c/em\u003e\u0026thinsp;=\u0026thinsp;1.14).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eDirect AMP Competence Impressions\u003c/h2\u003e \u003cp\u003eParticipants in both conditions took a similar amount of time to answer each direct AMP trial (Oddball: \u003cem\u003ex̄\u003c/em\u003e\u003csub\u003e\u003cem\u003eG (RT)\u003c/em\u003e\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;2178 ms; Competent: \u003cem\u003ex̄\u003c/em\u003e\u003csub\u003e\u003cem\u003eG (RT)\u003c/em\u003e\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;2237 ms; \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.488).\u003c/p\u003e \u003cp\u003eWe tested the conditions\u0026rsquo; impact on direct AMP evaluations with a Behavior (2: Competent, Oddball) by Response Method (2: Ordinal, Binary) ANOVA. Our pre-registered one-tailed test of behavior, \u003cem\u003eF\u003c/em\u003e(1, 328)\u0026thinsp;=\u0026thinsp;3.3, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.035, η\u003csub\u003ep\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.01, showed higher impressions in the competent \u003cem\u003e(M\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.68, \u003cem\u003eSE\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.07 versus \u003cem\u003eM\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.50, \u003cem\u003eSE\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.06). The interaction term was not significant, \u003cem\u003eF\u003c/em\u003e(1, 328)\u0026thinsp;=\u0026thinsp;1.0, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.310, suggesting the response scale did not affect oddball sensitivity. Adding direct AMP evaluations of the control robot as a covariate did not alter the results.\u003c/p\u003e \u003cp\u003eLastly, we predicted procedural differences between the two measures would reduce the overall impact of the oddball on direct AMP evaluations. A Behavior by Measure (2: Direct, Explicit) mixed model predicting standardized competence impressions, controlling for Response Method, revealed a significant interaction, \u003cem\u003eb\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.371, \u003cem\u003eSE\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.11, \u003cem\u003et\u003c/em\u003e(330)\u0026thinsp;=\u0026thinsp;3.4, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;.001. Specifically, explicit impressions were simultaneously more positive than direct AMP evaluations in the competent condition, x̄\u003csub\u003ed\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;.20, \u003cem\u003eSE\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.08, \u003cem\u003et\u003c/em\u003e(330)\u0026thinsp;=\u0026thinsp;2.5, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.012, and more negative than direct AMP evaluations in the oddball condition, x̄\u003csub\u003ed\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;.17, \u003cem\u003eSE\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.07, \u003cem\u003et\u003c/em\u003e(330)\u0026thinsp;=\u0026thinsp;2.3, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.020, suggesting that explicit competence scores were more extreme in both conditions.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eDiscussion\u003c/h3\u003e\n\u003cp\u003eExperiment 2 showed a single oddball performance could shift directly measured competence impressions even when measured under AMP-like conditions. The effect emerged regardless of response scale. However, the fact that the oddball’s influence was greater in participants’ explicit impressions than their direct AMP evaluations suggests we were justified in being concerned about the numerous differences between the two measures. This provides initial evidence the dissociation is associated with response intentionality.\u003c/p\u003e "},{"header":"Experiment 3","content":"\u003cp\u003eExperiments 3 and 4 gave participants two different variants of structurally aligned direct and indirect AMPs (within-subjects). Experiment 3 aligned the instruments with Payne et al.’s\u003csup\u003e\u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e\u003c/sup\u003e mouse-inputted ordinal responses, whereas Experiment 4 aligned both AMPs with key-inputted ordinal responses.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e\u003cp\u003e We recruited a total of 401 participants on Prolific. Participants who either had incomplete data (\u003cem\u003en\u003c/em\u003e = 15) or who failed memory checks (\u003cem\u003en\u003c/em\u003e = 21) were excluded (as pre-registered). Thus, we had a final sample of 365 \u003cem\u003e(M\u003c/em\u003e\u003csub\u003eage\u003c/sub\u003e: 34 years, \u003cem\u003eSD\u003c/em\u003e = 11; 183 males, 177 females). Most of the sample was White (257, 30 Asian, 35 Black, 20 Latino).\u003c/p\u003e\u003ch2\u003eProcedure\u003c/h2\u003e\u003cp\u003eThis experiment followed a similar procedure to Experiment 2 with several notable exceptions. (1) Participants completed both a direct and indirect AMP (which both included the novel control robot prime, WAM Arm) in a counterbalanced order and, after each, a memory check to ensure they evaluated the correct target. (2) AMP responses were solicited as mouse inputs to a 4-point ordinal scale (-2 to 2; Less Competent to More Competent). Like Payne et al.\u003csup\u003e\u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e\u003c/sup\u003e, the scale did not include a midpoint. (If we re-coded the ordinal responses to have equal steps between each response, all reported results across studies remain intact.) (3) We reintroduced the 2 (Majority Behavior: Competent, Incompetent) × 2 (Behavioral Consistency: Consistent, Inconsistent) between-subjects design.\u003c/p\u003e\u003cp\u003eThese changes resulted in the direct AMP differing from the indirect AMP in two structural ways. First, each trial’s prompt instructed participants to rate the [scale item] of the robot. Second, to avoid asking the same question about the same target multiple times, the direct AMP retained fewer trials (four per prime). We (correctly) assumed direct AMP evaluations would be less noisy than indirect AMP evaluations (see Table\u0026nbsp;3).\u003c/p\u003e\u003cdiv class=\"gridtable\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eExperiment 3: Reliability of AMP Responses\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e\u003ccolgroup cols=\"3\"\u003e\u003c/colgroup\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAMP\u003c/p\u003e \u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAll Trials\u003c/p\u003e \u003cp\u003eCronbach’s α\u003c/p\u003e \u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eFirst 4 Trials\u003c/p\u003e \u003cp\u003eCronbach's α\u003c/p\u003e \u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIndirect\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e.884\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e.660\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDirect\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e.878\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e.878\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/table\u003e\u003c/div\u003e\u003cp\u003e \u003cem\u003eNote.\u003c/em\u003e These data represent the reliability scores of AMP responses separated by AMP (averaged within and then across primes).\u003c/p\u003e\u003cp\u003eAfter completing the AMPs, participants completed explicit competence scales for the three robots, provided their demographics. and were debriefed.\u003c/p\u003e\n\u003ch3\u003eResults\u003c/h3\u003e\n\u003cp\u003eUnsurprisingly, explicit competence impressions were more similar to directly measured, \u003cem\u003er(\u003c/em\u003e363) = .820, \u003cem\u003ep\u003c/em\u003e \u0026lt; .001, than indirectly measured impressions, \u003cem\u003er(\u003c/em\u003e363) = .177, \u003cem\u003ep\u003c/em\u003e \u0026lt; .001. The correlation between direct and indirect measures was quite small, \u003cem\u003er(\u003c/em\u003e363) = .218, \u003cem\u003ep\u003c/em\u003e \u0026lt; .001.\u003c/p\u003e \u003cdiv id=\"Sec22\" class=\"Section2\"\u003e \u003ch2\u003eExplicit Competence Impressions\u003c/h2\u003e \u003cp\u003eA Majority Behavior by Behavioral Consistency ANOVA predicting explicit competence revealed the predicted interaction, \u003cem\u003eF\u003c/em\u003e(1, 361) = 44.8, \u003cem\u003ep\u003c/em\u003e \u0026lt; .001, η\u003csub\u003ep\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e = .11, such that the oddballs influenced explicit competence in the expected ways.\u003c/p\u003e \u003cdiv id=\"Sec23\" class=\"Section3\"\u003e \u003ch2\u003eDirectly vs. Indirectly Measured Competence Impressions\u003c/h2\u003e \u003cp\u003eTo test for the dissociation, we submitted standardized direct and indirect AMP evaluations to a Measure by Majority Behavior by Behavioral Consistency mixed-effects model, controlling for impressions of the control robots. The 3-way interaction, \u003cem\u003eb\u003c/em\u003e = .70, \u003cem\u003eSE\u003c/em\u003e = .22, \u003cem\u003et\u003c/em\u003e(357) = -3.2, \u003cem\u003ep\u003c/em\u003e = .001, η\u003csub\u003ep\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e = .03, revealed directly measured impressions significantly accommodated the oddballs, but indirectly measured impressions did not, consistent with our predictions.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec24\" class=\"Section2\"\u003e \u003cdiv id=\"Sec25\" class=\"Section3\"\u003e \u003c/div\u003e \u003cdiv id=\"Sec26\" class=\"Section3\"\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"Experiment 4","content":"\u003cp\u003eIn Experiment 4, we modified the direct AMP to simplify the evaluation process by removing end-of-trial prompts and converting the four scale items to a repeatedly measured competence item.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e\u003cp\u003e We recruited a total of 476 participants on Prolific. Participants who failed either attention check (\u003cem\u003en\u003c/em\u003e = 93), who failed the memory check (\u003cem\u003en\u003c/em\u003e = 33), who failed to engage the indirect AMP (i.e., pressed one button; \u003cem\u003en\u003c/em\u003e = 20), who completed more than 5% of their AMP trials in either under 50 ms or over 7,500 ms (\u003cem\u003en\u003c/em\u003e = 30), or who had incomplete data (\u003cem\u003en\u003c/em\u003e = 4) were excluded (as pre-registered). Thus, we had a final sample of 294 \u003cem\u003e(M\u003c/em\u003e\u003csub\u003eage\u003c/sub\u003e: 37 years, \u003cem\u003eSD\u003c/em\u003e = 12; 148 males, 143 females). Most of the sample identified as White (190, 25 Asian, 43 Black, 11 Latino).\u003c/p\u003e\u003ch2\u003eProcedure\u003c/h2\u003e\u003cp\u003eThis experiment’s procedure was like Experiment 3 with some changes. (1) Both AMPs now only solicited key-inputted (2) competence evaluations (3) without any end-of-trial prompts. (4) We also removed the incompetent conditions, once again. Following the AMPs, instead of eliciting explicit competence evaluations of the robots, we asked a series of behavior measures. We excluded these analyses because they did not replicate in unreported replication studies.\u003c/p\u003e\u003cp\u003eThe indirect and direct AMPs were reliable (see Table\u0026nbsp;4).\u003c/p\u003e\u003cdiv class=\"gridtable\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eExperiment 4: Reliability of AMP Responses\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e\u003ccolgroup cols=\"2\"\u003e\u003c/colgroup\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAMP\u003c/p\u003e \u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCronbach’s α\u003c/p\u003e \u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIndirect\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e.772\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDirect\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e.856\u003c/p\u003e \u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/table\u003e\u003c/div\u003e\u003cp\u003e \u003cem\u003eNote.\u003c/em\u003e These data represent the reliability scores of AMP responses separated by AMP (averaged within and then across primes).\u003c/p\u003e\n\u003ch3\u003eResults\u003c/h3\u003e\n\u003cp\u003eDirectly and indirectly measured competence impressions were slightly correlated, \u003cem\u003er(\u003c/em\u003e292)\u0026thinsp;=\u0026thinsp;.161, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.006. Although participants rated the target robot more quickly in the indirect AMP \u003cem\u003e(M\u003c/em\u003e\u0026thinsp;=\u0026thinsp;1084, \u003cem\u003eSE\u003c/em\u003e\u0026thinsp;=\u0026thinsp;35) compared with the direct AMP \u003cem\u003e(M\u003c/em\u003e\u0026thinsp;=\u0026thinsp;1263, \u003cem\u003eSE\u003c/em\u003e\u0026thinsp;=\u0026thinsp;35), \u003cem\u003et\u003c/em\u003e(292)\u0026thinsp;=\u0026thinsp;4.3, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;.001, this was likely a practicing effect.\u003c/p\u003e \u003cdiv id=\"Sec28\" class=\"Section2\"\u003e \u003ch2\u003eDirect vs. Indirect Competence Impressions\u003c/h2\u003e \u003cp\u003eTo test for a dissociation, we ran a General Linear Model with Measure and Behavior predicting standardized impressions of the target robot while controlling for impressions of the control robots. The significant interaction, \u003cem\u003eb\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.22, \u003cem\u003eSE\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.10, \u003cem\u003et\u003c/em\u003e(291)\u0026thinsp;=\u0026thinsp;2.2, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.030, η\u003csub\u003ep\u003c/sub\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;.02, revealed that while directly measured impressions were influenced by the oddball, \u003cem\u003et\u003c/em\u003e(573)\u0026thinsp;=\u0026thinsp;2.2, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.032, \u003cem\u003ed\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.09, indirectly measured impressions were not, \u003cem\u003et\u003c/em\u003e(573)\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;.72, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.475.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eDiscussion\u003c/h3\u003e\n\u003cp\u003e In Experiments 2\u0026ndash;4, participants\u0026rsquo; directly measured impressions reliably and uniquely incorporated a single oddball. We believe this is strong evidence that the dissociation is related to the intentionality of the evaluation. But why didn\u0026rsquo;t implicit impressions shift?\u003c/p\u003e \u003cp\u003eOne explanation is that implicit competence impressions - by their very nature - are unable to incorporate infrequent, inconsistent performances. Provided the oddball remains odd (i.e., a statistical anomaly), it cannot be relevant. An alternative explanation is that the inconsistent evidence was not sufficiently diagnostic. If the perceiver only believed the oddball was indicative of the robot\u0026rsquo;s \u003cem\u003etrue\u003c/em\u003e capabilities, implicit impressions would incorporate it even if it remained statistically rare. To test these two explanations, Experiment 5 directly manipulated the diagnosticity of the inconsistent information.\u003c/p\u003e"},{"header":"Experiment 5","content":"\u003cp\u003eExperiment 5 manipulated the inconsistent evidence\u0026rsquo;s diagnosticity with the industrial robot context in an impression updating setting. The first learning session consisted of 4 competent trials, whereas the second session consisted of 1\u0026ndash;4 incompetent trials - the more voluminous the errors, the more diagnostic the inconsistent information.\u003c/p\u003e \u003cp\u003eWe also informed participants that previous samples found the initial learning insufficient to judge the robot, thereby increasing the diagnosticity of the second learning sessions in all conditions.\u003c/p\u003e \u003cdiv id=\"Sec31\" class=\"Section2\"\u003e \u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003e We recruited a total of 501 participants on Prolific. We excluded those who failed to engage the either AMP (i.e., pressed one button; \u003cem\u003en\u003c/em\u003e\u0026thinsp;=\u0026thinsp;39) or who completed more than 5% of their AMP trials in either under 50 ms or over 7,500 ms (\u003cem\u003en\u003c/em\u003e\u0026thinsp;=\u0026thinsp;18; as pre-registered). In addition, a Google API data writing bug (Error: 429) caused some participants\u0026rsquo; explicit data and demographics to not be recorded (\u003cem\u003en\u003c/em\u003e\u0026thinsp;=\u0026thinsp;68). Thus, we had a final sample of 444 \u003cem\u003e(M\u003c/em\u003e\u003csub\u003eage\u003c/sub\u003e: 40 years, \u003cem\u003eSD\u003c/em\u003e\u0026thinsp;=\u0026thinsp;14; 195 males, 177 females). Most the sample was White (308, 36 Asian, 9 Black, 1 Latino).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec32\" class=\"Section2\"\u003e \u003ch2\u003eProcedure\u003c/h2\u003e \u003cp\u003eThis experiment revisited Experiment 1a\u0026rsquo;s industrial robot paradigm with some changes. We added a blocked design with two learning and two measurement phases. The first learning phase always showed 4 competent trials, whereas the second learning phase showed either 1, 2, 3, or 4 incompetent trials. Participants completed an AMP and explicit responses after each learning phase. The second learning phase began by telling participants that \u0026ldquo;[p]revious participants found that 4 navigation tests were not enough to determine how reliable and competent Boxy is. Therefore, you will now view additional tests.\u0026rdquo; We also added an exploratory \u0026ldquo;How much did you learn about Boxy\u0026rsquo;s capabilities\u0026rdquo; (Not very much - A great deal) slider question after each explicit competence measurement. Lastly, we changed the post-trial feedback to \u0026ldquo;SUCCESS\u0026rdquo; and \u0026ldquo;FAILURE\u0026rdquo; rather than \u0026ldquo;No Errors\u0026rdquo; and \u0026ldquo;Errors.\u0026rdquo;\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eResults\u003c/h3\u003e\n\u003cdiv id=\"Sec34\" class=\"Section2\"\u003e \u003ch2\u003eCompetence Impression Updating\u003c/h2\u003e \u003cp\u003eFirst, we tested for updating with paired samples \u003cem\u003et\u003c/em\u003e-tests contrasting initial and updated implicit and explicit impressions. Both explicit, \u003cem\u003et\u003c/em\u003e(375) = -29.6, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;.001, \u003cem\u003ed\u003c/em\u003e = -1.53, and implicit impressions robustly updated, \u003cem\u003et\u003c/em\u003e(443) = -5.5, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;.001, \u003cem\u003ed\u003c/em\u003e\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;.26, even when controlling for the control robot (via mixed-effects model).\u003c/p\u003e \u003cp\u003eNotably, even those in the 1 new piece of evidence condition updated their implicit impressions, \u003cem\u003et\u003c/em\u003e(112) = -1.9, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.028, \u003cem\u003ed\u003c/em\u003e\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;.182, \u003cem\u003eone-tailed\u003c/em\u003e. This is the first evidence showing a single oddball could shift implicit competence impressions.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eExplicit vs. Implicit Competence Impression Updating\u003c/h3\u003e\n\u003cp\u003eTo test for a dissociation, we ran a mixed-effects model with impression and condition predicting standardized updating scores (Time 1- Time 2). As predicted, we saw a significant interaction, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\chi\\:}^{2}\\left(3\\right)=22.2,\\:p\u0026lt;\\:.001,\\:{{{\\eta\\:}}_{p}}^{2}=\\:.06\\)\u003c/span\u003e\u003c/span\u003e, such that explicit impressions updated based on the volume of new evidence, whereas implicit impressions updated non-linearly (see Fig.\u0026nbsp;2). Specifically, implicit impressions indifferently updated between 1, 2, and 3 new pieces of evidence, but spiked when the new evidence met parity.\u003c/p\u003e \n\u003ch3\u003eDiscussion\u003c/h3\u003e\n\u003cp\u003eExperiment 5 showed that implicit competence impressions incorporate inconsistent information if it is diagnostic. As predicted, explicit updated in proportion to the volume of new evidence, whereas implicit did not. Notably, we found that a single piece of new evidence \u003cem\u003edid\u003c/em\u003e result in implicit updating. Before we discuss this effect, we first sought to replicate it.\u003c/p\u003e"},{"header":"Experiment 6","content":"\u003ch2\u003eMethods\u003c/h2\u003e\u003cp\u003e We recruited a total of 201 participants on Prolific. We excluded those with incomplete data (\u003cem\u003en\u003c/em\u003e = 2), who failed to engage the either AMP (i.e., pressed one button; \u003cem\u003en\u003c/em\u003e = 17) or who completed more than 5% of their AMP trials in either under 50 ms or over 7,500 ms (\u003cem\u003en\u003c/em\u003e = 26; as pre-registered). Thus, we had a final sample of 175 \u003cem\u003e(M\u003c/em\u003e\u003csub\u003eage\u003c/sub\u003e: 41 years, \u003cem\u003eSD\u003c/em\u003e = 15; 90 males, 84 females). Most of the sample identified as White (148, 9 Asian, 7 Black, 2 Latino).\u003c/p\u003e\u003ch2\u003eProcedure\u003c/h2\u003e\u003cp\u003eThis experiment was a direct replication of the 1 piece of new information condition from Experiment 5.\u003c/p\u003e\n\u003ch3\u003eResults\u003c/h3\u003e\n\u003cp\u003e \u003cb\u003eCompetence Impression Updating\u003c/b\u003e \u003c/p\u003e \u003cp\u003eOur pre-registered one-tailed paired \u003cem\u003et\u003c/em\u003e-test revealed significant updating, \u003cem\u003et\u003c/em\u003e(174) = -1.7, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.049, \u003cem\u003ed\u003c/em\u003e\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;.17. When we controlled for impressions of the control robot (via mixed-effects model), the effect became stronger, \u003cem\u003et\u003c/em\u003e(173) = -2, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.024, \u003cem\u003ed\u003c/em\u003e\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;.18.\u003c/p\u003e\n\u003ch3\u003eDiscussion\u003c/h3\u003e\n\u003cp\u003eExperiments 5 and 6 demonstrated a single oddball can influence implicit competence impressions. But why only in the updating (and not formation) studies?\u003c/p\u003e \u003cp\u003eIn the updating studies, we explicitly described the oddball as \u003cem\u003enecessary\u003c/em\u003e to determine the robot\u0026rsquo;s capabilities. We expected this would ratchet up its diagnosticity. To confirm this explanation, we directly manipulated the diagnosticity of the oddball in an updating setting.\u003c/p\u003e"},{"header":"Experiment 7","content":"\u003cp\u003eExperiment 7 tested whether the evidence\u0026rsquo;s diagnosticity caused implicit impressions to update. We predicted that implicit impressions would be more sensitive to a single failure described as recent versus outdated.\u003c/p\u003e\n\u003ch3\u003eMethods\u003c/h3\u003e\n\u003cp\u003e We recruited a total of 556 participants on Prolific. We excluded those with incomplete data (\u003cem\u003en\u003c/em\u003e\u0026thinsp;=\u0026thinsp;3), who failed to engage the either AMP (i.e., pressed one button; \u003cem\u003en\u003c/em\u003e\u0026thinsp;=\u0026thinsp;35), or who completed more than 5% of their AMP trials in either under 50 ms or over 7,500 ms (\u003cem\u003en\u003c/em\u003e\u0026thinsp;=\u0026thinsp;30), or who failed the memory check (\u003cem\u003en\u003c/em\u003e\u0026thinsp;=\u0026thinsp;72; as pre-registered). Thus, we had a final sample of 416 \u003cem\u003e(M\u003c/em\u003e\u003csub\u003eage\u003c/sub\u003e: 39 years, \u003cem\u003eSD\u003c/em\u003e\u0026thinsp;=\u0026thinsp;14; 206 males, 203 females). Most of the sample identified as White (345, 21 Asian, 23 Black, 3 Latino).\u003c/p\u003e\n\u003ch3\u003eProcedure\u003c/h3\u003e\n\u003cp\u003eThis experiment revisited Experiment 6\u0026rsquo;s paradigm with an error diagnosticity factor (2: low, high) factor. Prior to the first learning session, participants read, \u0026ldquo;[t]he first set presents Boxy\u0026rsquo;s [most recent / preliminary] attempts at this task. These tests are [highly relevant and directly / outdated but] show Boxy\u0026rsquo;s [current capabilities / beginnings].\u0026rdquo; And then prior to the second learning session, participants read, \u0026ldquo;Now that you saw [Boxy\u0026rsquo;s most relevant, recent performances / the irrelevant, outdated evidence], you will see [an irrelevant earlier attempt / Boxy\u0026rsquo;s most relevant, recent test]. The final test occurred [before / after] Boxy was upgraded, and [is therefore outdated, irrelevant, and does not reflect / thus provides the best look at] Boxy\u0026rsquo;s true capabilities.\u0026rdquo;\u003c/p\u003e \u003cp\u003eFollowing each learning session, participants completed an AMP and explicit competence scales for both robots. After the second explicit competence measure, they answered a binary memory check (\"Which test session showed you the most recently measured performance?\") and a manipulation check (\"How much did the final test teach you about the robot?\"; 1\u0026thinsp;=\u0026thinsp;not very much, 7\u0026thinsp;=\u0026thinsp;a great deal) before demographics and debriefing.\u003c/p\u003e\n\u003ch3\u003eResults\u003c/h3\u003e\n\u003cp\u003e \u003cb\u003eManipulation Check\u003c/b\u003e \u003c/p\u003e \u003cp\u003eParticipants in the high error diagnosticity condition \u003cem\u003e(M\u003c/em\u003e\u0026thinsp;=\u0026thinsp;5.3, \u003cem\u003eSD\u003c/em\u003e\u0026thinsp;=\u0026thinsp;1.4) reported more learning from the error than those in the low error diagnosticity condition \u003cem\u003e(M\u003c/em\u003e\u0026thinsp;=\u0026thinsp;3.5, \u003cem\u003eSD\u003c/em\u003e\u0026thinsp;=\u0026thinsp;2), \u003cem\u003et\u003c/em\u003e(283)\u0026thinsp;=\u0026thinsp;10.2, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;.001, \u003cem\u003ed\u003c/em\u003e\u0026thinsp;=\u0026thinsp;1.08.\u003c/p\u003e \u003cp\u003e \u003cb\u003eCompetence Impression Updating\u003c/b\u003e \u003c/p\u003e \u003cp\u003eParticipants in the high diagnosticity condition \u003cem\u003e(M\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.15, \u003cem\u003eSD\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.56) updated their implicit competence impressions more than those in the low diagnosticity condition \u003cem\u003e(M\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.02, \u003cem\u003eSD\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.45), \u003cem\u003et\u003c/em\u003e(409)\u0026thinsp;=\u0026thinsp;2.6, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.005, \u003cem\u003ed\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.25, \u003cem\u003eone-tailed\u003c/em\u003e, even when we controlled for the control robot (via ANCOVA).\u003c/p\u003e \u003cp\u003eUnsurprisingly, we find similar effects (of a larger magnitude) for explicit updating. Participants in the high diagnosticity condition \u003cem\u003e(M\u003c/em\u003e\u0026thinsp;=\u0026thinsp;2.6, \u003cem\u003eSD\u003c/em\u003e\u0026thinsp;=\u0026thinsp;1.6) updated their impressions more than those in the low diagnosticity condition \u003cem\u003e(M\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.6, \u003cem\u003eSD\u003c/em\u003e\u0026thinsp;=\u0026thinsp;1.2), \u003cem\u003et\u003c/em\u003e(412)\u0026thinsp;=\u0026thinsp;14.9, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;.001, \u003cem\u003ed\u003c/em\u003e\u0026thinsp;=\u0026thinsp;1.45.\u003c/p\u003e\n\u003ch3\u003eDiscussion\u003c/h3\u003e\n\u003cp\u003eExperiment 7 was the first experiment to demonstrate diagnosticity impacts implicit competence impressions. To date, the relationship between diagnosticity and implicit impressions was exclusively studied with general evaluations and moral information.\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e,\u003cspan citationid=\"CR60\" class=\"CitationRef\"\u003e60\u003c/span\u003e\u003c/sup\u003e These results challenge purely process-based accounts of the dissociation, suggesting instead that the nature of the evidence itself shapes how much it influences implicit impressions.\u003c/p\u003e"},{"header":"General Discussion","content":"\u003cp\u003eRecent work revealed a novel dissociation in how people form and update competence impressions (of a robot; Surdel et al.\u003csup\u003e\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e\u003c/sup\u003e). Specifically, people\u0026rsquo;s explicit impressions were more sensitive to inconsistency than their implicit impressions. In the present paper, we conducted nine experiments identifying the psychological mechanisms of this dissociation. Our results show that the dissociation (a) generalized to the ecologically relevant settings of industrial robots, surgical robots, and autonomous vehicles, (b) stems from differences in intentional versus unintentional processing, and (c) reflects the diagnosticity of inconsistent evidence rather than constraints inherent to implicit impressions. Together, these findings simultaneously teach us about how humans process competence information while also offering practical insights for robot deployment.\u003c/p\u003e"},{"header":"Theoretical Contributions","content":"\u003cp\u003eThese data offer four key insights to the social cognitive literature. First, we demonstrate robust generalization of an explicit-implicit dissociation across diverse contexts. Psychologists often underemphasize generalizability, causing an overextrapolation of narrow findings.\u003csup\u003e\u003cspan citationid=\"CR61\" class=\"CitationRef\"\u003e61\u003c/span\u003e\u003c/sup\u003e While Surdel et al.\u003csup\u003e\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e\u003c/sup\u003e documented this pattern in a game-based tic-tac-toe setting, we show it persists across industrial navigation tasks, precise surgical operations, and autonomous driving scenarios - domains where competence judgments drive real-world consequences.\u003c/p\u003e \u003cp\u003eSecond, we provided strong evidence the dissociation is associated with the judgment\u0026rsquo;s intentionality rather than structural misalignment between the two types of measures.\u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e,\u003cspan additionalcitationids=\"CR54 CR55\" citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e56\u003c/span\u003e\u003c/sup\u003e By clearly connecting the dissociation to the measure\u0026rsquo;s intentionality, we identify specific circumstances in which competence impressions should be sensitive to relatively undiagnostic information (i.e., intentionally measured judgments) and when they should not (i.e., unintentionally measured judgments). This offers practical insights for psychologists, marketeers, and engineers.\u003c/p\u003e \u003cp\u003eThird, we bridge theoretical frameworks between moral and competence domains. Previous research demonstrated that implicit global impressions rapidly updated in the face of highly diagnostic evidence.\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e,\u003cspan citationid=\"CR60\" class=\"CitationRef\"\u003e60\u003c/span\u003e\u003c/sup\u003e We show parallel effects in competence impressions\u0026mdash;when a robot's error was perceived as diagnostic of its true capabilities, implicit impressions readily incorporated it. This suggests that common mechanisms may govern how people update competence and morality impressions.\u003c/p\u003e \u003cp\u003eFinally, these findings help reconcile competing theoretical perspectives. Traditional dual-process theories argue implicit impressions develop through slow associative learning, resistant to rapid updating.\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e,\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e Our findings support a newer account\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e,\u003cspan citationid=\"CR60\" class=\"CitationRef\"\u003e60\u003c/span\u003e\u003c/sup\u003e where both explicit and implicit impressions are sensitive to evidence quality but differ in their evidentiary thresholds. When new evidence is perceived as sufficiently diagnostic, both implicit and explicit impressions rapidly update. But when new evidence is relatively weak, implicit impressions show slower updating.\u003c/p\u003e \u003cp\u003eWhy, then, did explicit impressions respond to weakly diagnostic evidence? The exchange of information via intentional evaluation is rife with conversational implicatum and demand.\u003csup\u003e\u003cspan additionalcitationids=\"CR25 CR26\" citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e,\u003cspan citationid=\"CR62\" class=\"CitationRef\"\u003e62\u003c/span\u003e\u003c/sup\u003e First, conversational norms necessitate the responder to treat the various inputs \u003cem\u003eas though\u003c/em\u003e they are relevant to the output - almost an assumed diagnosticity. Suppose you ask your neighbor if his nephew is reliable and she replies, \u0026ldquo;He owns a cow, what do you think?\u0026rdquo; Although you likely lack an association between cow ownership and reliability, you \u0026ndash; being a well-socialized communicator \u0026ndash; will assume relevance and harvest diagnosticity (e.g., cows require regular care, thus he\u0026rsquo;s reliable). Second, intentional measures may trigger unique expectations about which evidence participants should incorporate.\u003csup\u003e\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e,\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e Had your neighbor merely replied, \u0026ldquo;He owns a cow,\u0026rdquo; perhaps the hunt for diagnosticity may never have commenced. But the mere presence of the question about competence (of the target) informs the conversational partner that the evidence \u003cem\u003eshould be sufficient\u003c/em\u003e to answer the question.\u003c/p\u003e"},{"header":"Practical Implications","content":"\u003cp\u003eThese findings offer direct guidance for robot deployment in real-world settings. First, they highlight that measurement approach matters\u0026mdash;direct and indirect measures may yield divergent conclusions about a robot\u0026rsquo;s capabilities, particularly when performance varies. It is an open question whether explicit or implicit impressions better predict user behavior, but previous work suggests they both might in some circumstances, though more research is needed on this question.\u003csup\u003e\u003cspan additionalcitationids=\"CR64 CR65 CR66\" citationid=\"CR63\" class=\"CitationRef\"\u003e63\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR67\" class=\"CitationRef\"\u003e67\u003c/span\u003e,cf. \u003cspan citationid=\"CR68\" class=\"CitationRef\"\u003e68\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eSecond, our findings lend specific strategies for managing users\u0026rsquo; impressions of their robot (as tools, coworkers, or products). Rather than attempting to eliminate all inconsistencies\u0026mdash;an impossible goal\u0026mdash;manufacturers should focus on preventing their robots from committing highly diagnostic errors that impact both explicit and implicit impressions. When errors occur, actively reducing their perceived diagnosticity through appropriate framing (e.g., highlighting situational factors or system updates) may help preserve trust. Notably, however, it is likely challenging to anticipate exactly which actions (for which perceivers) will be perceived as (non)diagnostic.\u003c/p\u003e"},{"header":"Limitations and Future Directions","content":"\u003cp\u003eSeveral limitations suggest important directions for future research. First, although we demonstrated generalization across robot types and contexts, all our studies used relatively simple, discrete tasks with clear success criteria and feedback. Future work should examine whether similar patterns emerge in tasks where competence assessment is less straightforward, such as dynamic surgical settings. It would be fascinating to see how an expert surgeon learns about a surgical robot who performs well, but not \u003cem\u003eperfectly\u003c/em\u003e (e.g., 20% bigger incision than necessary).\u003c/p\u003e \u003cp\u003eSecond, our studies focused on impressions formed over a single session. Longitudinal research is needed to understand how explicit-implicit dissociations evolve over repeated interactions and whether they predict important outcomes like trust, reliance, and acceptance. This is particularly relevant for understanding how people adapt to learning robots whose performance may improve over time.\u003c/p\u003e \u003cp\u003eFinally, although we identified diagnosticity as a key mechanism, other factors likely influence the integration of inconsistent evidence. These could include the perceived intentionality of robot actions,\u003csup\u003e\u003cspan citationid=\"CR69\" class=\"CitationRef\"\u003e69\u003c/span\u003e\u003c/sup\u003e the presence of explanations,\u003csup\u003e\u003cspan citationid=\"CR70\" class=\"CitationRef\"\u003e70\u003c/span\u003e\u003c/sup\u003e and users' construal of the robot's role.\u003csup\u003e\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e\u003c/sup\u003e Understanding these moderators could help organizations better manage human-robot interactions.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eA growing wave of psychologists are looking to robots as tools for studying basic social cognition.\u003csup\u003e\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e,\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e\u003c/sup\u003e Separately, computer science and marketing departments are becoming increasingly interested in how people judge robots. Our research shows that when inconsistent evidence about a robot is relatively nondiagnostic, explicit and implicit impressions will dissociate. The present work contributes to at least four theoretical debates in social cognition, while also lending interdisciplinary insights. As automation continues to expand, understanding competence impression formation becomes increasingly relevant to daily life.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eData Availability\u003c/p\u003e\n\u003cp\u003eAll data, materials, and code can be found at: https://osf.io/zwued/?view_only=9bc9367ab70e4935b278972efb662a1c\u003c/p\u003e\u003cp\u003eAcknowledgements\u003c/p\u003e\n\u003cp\u003eWe thank John Bargh, Margaret Clark, the Implicit Social Cognition Lab at Yale, OpenAI\u0026rsquo;s ChatGPT, and Anthropic\u0026rsquo;s Claude for helpful comments on an earlier version of the manuscript. We also thank Malte Jung and Wen-Ying Lee for their software support, and the Office of Naval Research for funding this research (Award Number: N00014-19-1-2299).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAuthor Contributions\u003c/p\u003e\n\u003cp\u003eN.S. and M.F. equally contributed to project conceptualization and methodology. N.S. led data collection and analyses, and drafted the initial manuscript text. M.F. supervised, acquired funding, and revised the manuscript text. Both authors reviewed the manuscript.\u003c/p\u003e\n\u003cp\u003eCompeting Interests\u003c/p\u003e\n\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eFiske, S. T., Cuddy, A. J. \u0026amp; Glick, P. Universal dimensions of social cognition: Warmth and competence. \u003cem\u003eTrends Cogn. Sci.\u003c/em\u003e \u003cstrong\u003e11\u003c/strong\u003e, 77\u0026ndash;83 (2007).\u003c/li\u003e\n\u003cli\u003eTodorov, A. \u0026amp; Oh, D. The structure and perceptual basis of social judgments from faces. \u003cem\u003eAdv. Exp. Soc. Psychol.\u003c/em\u003e \u003cstrong\u003e63\u003c/strong\u003e, 189\u0026ndash;245 (2021).\u003c/li\u003e\n\u003cli\u003eTodorov, A., Mandisodza, A. N., Goren, A. \u0026amp; Hall, C. C. Inferences of Competence from Faces Predict Election Outcomes. \u003cem\u003eScience\u003c/em\u003e \u003cstrong\u003e308\u003c/strong\u003e, 1623\u0026ndash;1626 (2005).\u003c/li\u003e\n\u003cli\u003eGawronski, B. \u0026amp; Sritharan, R. Formation, change, and contextualization of mental associations: Determinants and principles of variations in implicit measures. in \u003cem\u003eHandbook of implicit social cognition: Measurement, theory, and applications\u003c/em\u003e 216\u0026ndash;240 (The Guilford Press, New York, NY, US, 2010).\u003c/li\u003e\n\u003cli\u003eGregg, A. P., Seibt, B. \u0026amp; Banaji, M. R. Easier done than undone: Asymmetry in the malleability of implicit preferences. \u003cem\u003eJ. Pers. Soc. Psychol.\u003c/em\u003e \u003cstrong\u003e90\u003c/strong\u003e, 1\u0026ndash;20 (2006).\u003c/li\u003e\n\u003cli\u003ePeters, K. R. \u0026amp; Gawronski, B. Are We Puppets on a String? Comparing the Impact of Contingency and Validity on Implicit and Explicit Evaluations. \u003cem\u003ePers. Soc. Psychol. Bull.\u003c/em\u003e \u003cstrong\u003e37\u003c/strong\u003e, 557\u0026ndash;569 (2011).\u003c/li\u003e\n\u003cli\u003eRydell, R. J. \u0026amp; McConnell, A. R. Understanding implicit and explicit attitude change: A systems of reasoning analysis. \u003cem\u003eJ. Pers. Soc. Psychol.\u003c/em\u003e \u003cstrong\u003e91\u003c/strong\u003e, 995\u0026ndash;1008 (2006).\u003c/li\u003e\n\u003cli\u003eCao, J. \u0026amp; Banaji, M. R. The base rate principle and the fairness principle in social judgment. \u003cem\u003eProc. Natl. Acad. Sci.\u003c/em\u003e \u003cstrong\u003e113\u003c/strong\u003e, 7475\u0026ndash;7480 (2016).\u003c/li\u003e\n\u003cli\u003eGawronski, B. \u0026amp; Bodenhausen, G. V. Associative and propositional processes in evaluation: An integrative review of implicit and explicit attitude change. \u003cem\u003ePsychol. Bull.\u003c/em\u003e \u003cstrong\u003e132\u003c/strong\u003e, 692\u0026ndash;731 (2006).\u003c/li\u003e\n\u003cli\u003eGawronski, B. \u0026amp; Bodenhausen, G. V. The associative\u0026ndash;propositional evaluation model: Theory, evidence, and open questions. in \u003cem\u003eAdvances in experimental social psychology\u003c/em\u003e vol. 44 59\u0026ndash;127 (Elsevier, 2011).\u003c/li\u003e\n\u003cli\u003eGawronski, B., Luke, D. M. \u0026amp; Creighton, L. A. Dual-Process Theories. in \u003cem\u003eThe Oxford Handbook of Social Cognition, Second Edition\u003c/em\u003e (eds. Carlston, D. E., Hugenberg, K. \u0026amp; Johnson, K. L.) 319\u0026ndash;353 (Oxford University Press, 2024). doi:10.1093/oxfordhb/9780197763414.013.12.\u003c/li\u003e\n\u003cli\u003eMcConnell, A. R. \u0026amp; Rydell, R. J. The systems of evaluation model. in \u003cem\u003eDual-process theories of the social mind\u003c/em\u003e (eds. Sherman, J. W., Gawronski, B. \u0026amp; Trope, Y.) 204\u0026ndash;217 (New York: Guilford Press, 2014).\u003c/li\u003e\n\u003cli\u003eSkowronski, J. J. \u0026amp; Carlston, D. E. Negativity and extremity biases in impression formation: A review of explanations. \u003cem\u003ePsychol. Bull.\u003c/em\u003e \u003cstrong\u003e105\u003c/strong\u003e, 131\u0026ndash;142 (1989).\u003c/li\u003e\n\u003cli\u003eBrannon, S. M. \u0026amp; Gawronski, B. A Second Chance for First Impressions? Exploring the Context-(In)Dependent Updating of Implicit Evaluations. \u003cem\u003eSoc. Psychol. Personal. Sci.\u003c/em\u003e \u003cstrong\u003e8\u003c/strong\u003e, 275\u0026ndash;283 (2017).\u003c/li\u003e\n\u003cli\u003eCone, J. \u0026amp; Ferguson, M. J. He did what? The role of diagnosticity in revising implicit evaluations. \u003cem\u003eJ. Pers. Soc. Psychol.\u003c/em\u003e \u003cstrong\u003e108\u003c/strong\u003e, 37\u0026ndash;57 (2015).\u003c/li\u003e\n\u003cli\u003eKurdi, B., Mann, T. C. \u0026amp; Ferguson, M. J. Persuading the Implicit Mind: Changing Negative Implicit Evaluations With an 8-Minute Podcast. \u003cem\u003eSoc. Psychol. Personal. Sci.\u003c/em\u003e \u003cstrong\u003e13\u003c/strong\u003e, 688\u0026ndash;697 (2022).\u003c/li\u003e\n\u003cli\u003eMann, T. C. \u0026amp; Ferguson, M. J. Can we undo our first impressions? The role of reinterpretation in reversing implicit evaluations. \u003cem\u003eJ. Pers. Soc. Psychol.\u003c/em\u003e \u003cstrong\u003e108\u003c/strong\u003e, 823\u0026ndash;849 (2015).\u003c/li\u003e\n\u003cli\u003eMann, T. C., Cone, J., Heggeseth, B. \u0026amp; Ferguson, M. J. Updating implicit impressions: New evidence on intentionality and the affect misattribution procedure. \u003cem\u003eJ. Pers. Soc. Psychol.\u003c/em\u003e \u003cstrong\u003e116\u003c/strong\u003e, 349\u0026ndash;374 (2019).\u003c/li\u003e\n\u003cli\u003eShen, X. \u0026amp; Ferguson, M. J. How resistant are implicit impressions of facial trustworthiness? When new evidence leads to durable updating. \u003cem\u003eJ. Exp. Soc. Psychol.\u003c/em\u003e \u003cstrong\u003e97\u003c/strong\u003e, 104219 (2021).\u003c/li\u003e\n\u003cli\u003eVan Dessel, P., Cone, J., Gast, A. \u0026amp; De Houwer, J. The impact of valenced verbal information on implicit and explicit evaluation: the role of information diagnosticity, primacy, and memory cueing. \u003cem\u003eCogn. Emot.\u003c/em\u003e \u003cstrong\u003e34\u003c/strong\u003e, 74\u0026ndash;85 (2020).\u003c/li\u003e\n\u003cli\u003eFerguson, M. J., Mann, T. C., Cone, J. \u0026amp; Shen, X. When and How Implicit First Impressions Can Be Updated. \u003cem\u003eCurr. Dir. Psychol. Sci.\u003c/em\u003e \u003cstrong\u003e28\u003c/strong\u003e, 331\u0026ndash;336 (2019).\u003c/li\u003e\n\u003cli\u003eFerguson, M. J., Mann, T. C., Cone, J. \u0026amp; Shen, X. The role of context in implicit and explicit evaluation. (2025).\u003c/li\u003e\n\u003cli\u003eShen, X., Mann, T. C. \u0026amp; Ferguson, M. J. Beware a dishonest face: Updating face-based implicit impressions using diagnostic behavioral information. \u003cem\u003eJ. Exp. Soc. Psychol.\u003c/em\u003e \u003cstrong\u003e86\u003c/strong\u003e, 103888 (2020).\u003c/li\u003e\n\u003cli\u003eCorneille, O. \u0026amp; Lush, P. Sixty Years After Orne\u0026rsquo;s American Psychologist Article: A Conceptual Framework for Subjective Experiences Elicited by Demand Characteristics. \u003cem\u003ePersonal. Soc. Psychol. Rev.\u003c/em\u003e \u003cstrong\u003e27\u003c/strong\u003e, 83\u0026ndash;101 (2023).\u003c/li\u003e\n\u003cli\u003eOrne, M. T. Communication by the total experimental situation: Why it is important, how it is evaluated, and its significance for the ecological validity of findings. in \u003cem\u003eCommunication and affect\u003c/em\u003e (eds. Pliner, P., Krames, L. \u0026amp; Alloway, T.) 157\u0026ndash;191 (Elsevier, 1973).\u003c/li\u003e\n\u003cli\u003eOrne, M. T. On the social psychology of the psychological experiment: With particular reference to demand characteristics and their implications. \u003cem\u003eAm. Psychol.\u003c/em\u003e \u003cstrong\u003e17\u003c/strong\u003e, 776\u0026ndash;783 (1962).\u003c/li\u003e\n\u003cli\u003eSchwarz, N. Measurement as cooperative communication: What research participants learn from questionnaires. in \u003cem\u003eThe SAGE handbook of measurement\u003c/em\u003e 43\u0026ndash;59 (Sage Thousand Oaks, CA, 2010).\u003c/li\u003e\n\u003cli\u003eKurdi, B., Mann, T. C., Charlesworth, T. E. S. \u0026amp; Banaji, M. R. The relationship between implicit intergroup attitudes and beliefs. \u003cem\u003eProc. Natl. Acad. Sci.\u003c/em\u003e \u003cstrong\u003e116\u003c/strong\u003e, 5862\u0026ndash;5871 (2019).\u003c/li\u003e\n\u003cli\u003eSkowronski, J. J. \u0026amp; Carlston, D. E. Social judgment and social memory: The role of cue diagnosticity in negativity, positivity, and extremity biases. \u003cem\u003eJ. Pers. Soc. Psychol.\u003c/em\u003e \u003cstrong\u003e52\u003c/strong\u003e, 689\u0026ndash;699 (1987).\u003c/li\u003e\n\u003cli\u003eWojciszke, B. Morality and competence in person- and self-perception. \u003cem\u003eEur. Rev. Soc. Psychol.\u003c/em\u003e 155\u0026ndash;188 (2005).\u003c/li\u003e\n\u003cli\u003eLandy, J. F., Piazza, J. \u0026amp; Goodwin, G. P. When It\u0026rsquo;s Bad to Be Friendly and Smart: The Desirability of Sociability and Competence Depends on Morality. \u003cem\u003ePers. Soc. Psychol. Bull.\u003c/em\u003e \u003cstrong\u003e42\u003c/strong\u003e, 1272\u0026ndash;1290 (2016).\u003c/li\u003e\n\u003cli\u003eMende-Siedlecki, P., Baron, S. G. \u0026amp; Todorov, A. Diagnostic Value Underlies Asymmetric Updating of Impressions in the Morality and Ability Domains. \u003cem\u003eJ. Neurosci.\u003c/em\u003e \u003cstrong\u003e33\u003c/strong\u003e, 19406\u0026ndash;19415 (2013).\u003c/li\u003e\n\u003cli\u003eLuttrell, A., Sacchi, S. \u0026amp; Brambilla, M. Changing impressions in competence-oriented domains: The primacy of morality endures. \u003cem\u003eJ. Exp. Soc. Psychol.\u003c/em\u003e \u003cstrong\u003e98\u003c/strong\u003e, 104246 (2022).\u003c/li\u003e\n\u003cli\u003eAsch, S. E. Forming impressions of personality. \u003cem\u003eJ. Abnorm. Soc. Psychol.\u003c/em\u003e \u003cstrong\u003e41\u003c/strong\u003e, 258\u0026ndash;290 (1946).\u003c/li\u003e\n\u003cli\u003eBrambilla, M., Carraro, L., Castelli, L. \u0026amp; Sacchi, S. Changing impressions: Moral character dominates impression updating. \u003cem\u003eJ. Exp. Soc. Psychol.\u003c/em\u003e \u003cstrong\u003e82\u003c/strong\u003e, 64\u0026ndash;73 (2019).\u003c/li\u003e\n\u003cli\u003eBrambilla, M., Sacchi, S., Rusconi, P. \u0026amp; Goodwin, G. P. The primacy of morality in impression development: Theory, research, and future directions. in \u003cem\u003eAdvances in Experimental Social Psychology\u003c/em\u003e vol. 64 187\u0026ndash;262 (Elsevier, 2021).\u003c/li\u003e\n\u003cli\u003eAmodio, D. M. \u0026amp; Devine, P. G. Stereotyping and evaluation in implicit race bias: Evidence for independent constructs and unique effects on behavior. \u003cem\u003eJ. Pers. Soc. Psychol.\u003c/em\u003e 652\u0026ndash;661 (2006).\u003c/li\u003e\n\u003cli\u003eAmodio, D. M. \u0026amp; Ratner, K. G. A Memory Systems Model of Implicit Social Cognition. \u003cem\u003eCurr. Dir. Psychol. Sci.\u003c/em\u003e \u003cstrong\u003e20\u003c/strong\u003e, 143\u0026ndash;148 (2011).\u003c/li\u003e\n\u003cli\u003eAmodio, D. M. \u0026amp; Hamilton, H. K. Intergroup anxiety effects on implicit racial evaluation and stereotyping. \u003cem\u003eEmotion\u003c/em\u003e \u003cstrong\u003e12\u003c/strong\u003e, (2012).\u003c/li\u003e\n\u003cli\u003eSurdel, N. \u003cem\u003eet al.\u003c/em\u003e Judging Robot Ability: How People Form Implicit and Explicit Impressions of Robot Competence. \u003cem\u003eJ. Exp. Psychol. Gen.\u003c/em\u003e \u003cstrong\u003e153\u003c/strong\u003e, 1309\u0026ndash;1335 (2024).\u003c/li\u003e\n\u003cli\u003eWykowska, A. Robots as mirrors of the human mind. \u003cem\u003eCurr. Dir. Psychol. Sci.\u003c/em\u003e \u003cstrong\u003e30\u003c/strong\u003e, 34\u0026ndash;40 (2021).\u003c/li\u003e\n\u003cli\u003eWykowska, A., Chaminade, T. \u0026amp; Cheng, G. Embodied artificial agents for understanding human social cognition. \u003cem\u003ePhilos. Trans. R. Soc. B Biol. Sci.\u003c/em\u003e \u003cstrong\u003e371\u003c/strong\u003e, 20150375 (2016).\u003c/li\u003e\n\u003cli\u003eBigman, Y., Surdel, N. \u0026amp; Ferguson, M. J. Trait attribution explains human-robot interactions. \u003cem\u003eBehav. Brain Sci.\u003c/em\u003e \u003cstrong\u003e46\u003c/strong\u003e, E23 (2023).\u003c/li\u003e\n\u003cli\u003eCross, E. S. \u0026amp; Ramsey, R. Mind Meets Machine: Towards a Cognitive Science of Human\u0026ndash;Machine Interactions. \u003cem\u003eTrends Cogn. Sci.\u003c/em\u003e \u003cstrong\u003e25\u003c/strong\u003e, 200\u0026ndash;212 (2021).\u003c/li\u003e\n\u003cli\u003eDe Graaf, M. M. \u0026amp; Malle, B. F. How people explain action (and autonomous intelligent systems should too). in \u003cem\u003e2017 AAAI Fall Symposium Series\u003c/em\u003e 19\u0026ndash;26 (2017).\u003c/li\u003e\n\u003cli\u003eClark, H. H. \u0026amp; Fischer, K. Social robots as depictions of social agents. \u003cem\u003eBehav. Brain Sci.\u003c/em\u003e \u003cstrong\u003e46\u003c/strong\u003e, 1\u0026ndash;65 (2023).\u003c/li\u003e\n\u003cli\u003eGray, K. \u0026amp; Wegner, D. M. Feeling robots and human zombies: Mind perception and the uncanny valley. \u003cem\u003eCognition\u003c/em\u003e \u003cstrong\u003e125\u003c/strong\u003e, 125\u0026ndash;130 (2012).\u003c/li\u003e\n\u003cli\u003eBonezzi, A., Ostinelli, M. \u0026amp; Melzner, J. The human black-box: The illusion of understanding human better than algorithmic decision-making. \u003cem\u003eJ. Exp. Psychol. Gen.\u003c/em\u003e \u003cstrong\u003e151\u003c/strong\u003e, 2250\u0026ndash;2258 (2022).\u003c/li\u003e\n\u003cli\u003eTully, S., Longoni, C. \u0026amp; Appel, G. \u003cem\u003eKnowledge of Artificial Intelligence Predicts Lower AI Receptivity\u003c/em\u003e. https://osf.io/t9u8g (2023) doi:10.31234/osf.io/t9u8g.\u003c/li\u003e\n\u003cli\u003ePayne, B. K., Cheng, C. M., Govorun, O. \u0026amp; Stewart, B. D. An inkblot for attitudes: Affect misattribution as implicit measurement. \u003cem\u003eJ. Pers. Soc. Psychol.\u003c/em\u003e \u003cstrong\u003e89\u003c/strong\u003e, 277\u0026ndash;293 (2005).\u003c/li\u003e\n\u003cli\u003eCone, J., Flaharty, K. \u0026amp; Ferguson, M. J. Believability of evidence matters for correcting social impressions. \u003cem\u003eProc. Natl. Acad. Sci.\u003c/em\u003e \u003cstrong\u003e116\u003c/strong\u003e, 9802\u0026ndash;9807 (2019).\u003c/li\u003e\n\u003cli\u003eCarpinella, C. M., Wyman, A. B., Perez, M. A. \u0026amp; Stroessner, S. J. The Robotic Social Attributes Scale (RoSAS): Development and Validation. in \u003cem\u003eProceedings of the 2017 ACM/IEEE International Conference on Human-Robot Interaction\u003c/em\u003e 254\u0026ndash;262 (ACM, Vienna Austria, 2017). doi:10.1145/2909824.3020208.\u003c/li\u003e\n\u003cli\u003ePayne, B. K., Burkley, M. A. \u0026amp; Stokes, M. B. Why do implicit and explicit attitude tests diverge? The role of structural fit. \u003cem\u003eJ. Pers. Soc. Psychol.\u003c/em\u003e \u003cstrong\u003e94\u003c/strong\u003e, 16\u0026ndash;31 (2008).\u003c/li\u003e\n\u003cli\u003ePayne, B. K. \u003cem\u003eet al.\u003c/em\u003e Intention Invention and the Affect Misattribution Procedure: Reply to Bar-Anan and Nosek (2012). \u003cem\u003ePers. Soc. Psychol. Bull.\u003c/em\u003e \u003cstrong\u003e39\u003c/strong\u003e, 375\u0026ndash;386 (2013).\u003c/li\u003e\n\u003cli\u003eB\u0026eacute;na, J., Melnikoff, D. E., Mierop, A. \u0026amp; Corneille, O. Revisiting dissociation hypotheses with a structural fit approach: The case of the prepared reflex framework. \u003cem\u003eJ. Exp. Soc. Psychol.\u003c/em\u003e \u003cstrong\u003e100\u003c/strong\u003e, 104297 (2022).\u003c/li\u003e\n\u003cli\u003eCorneille, O. \u0026amp; Gawronski, B. Self-reports are better measurement instruments than implicit measures. \u003cem\u003eNat. Rev. Psychol.\u003c/em\u003e \u003cstrong\u003e3\u003c/strong\u003e, 835\u0026ndash;846 (2024).\u003c/li\u003e\n\u003cli\u003eHancock, P. A. \u003cem\u003eet al.\u003c/em\u003e A Meta-Analysis of Factors Affecting Trust in Human-Robot Interaction. \u003cem\u003eHum. Factors\u003c/em\u003e \u003cstrong\u003e53\u003c/strong\u003e, 517\u0026ndash;527 (2011).\u003c/li\u003e\n\u003cli\u003eBaker, A. L., Phillips, E. K., Ullman, D. \u0026amp; Keebler, J. R. Toward an Understanding of Trust Repair in Human-Robot Interaction: Current Research and Future Directions. \u003cem\u003eACM Trans. Interact. Intell. Syst.\u003c/em\u003e \u003cstrong\u003e8\u003c/strong\u003e, 1\u0026ndash;30 (2018).\u003c/li\u003e\n\u003cli\u003eLogan, G. D. Toward an instance theory of automatization. \u003cem\u003ePsychol. Rev.\u003c/em\u003e \u003cstrong\u003e95\u003c/strong\u003e, 492\u0026ndash;527 (1988).\u003c/li\u003e\n\u003cli\u003eCone, J., Mann, T. C. \u0026amp; Ferguson, M. J. Changing Our Implicit Minds: How, When, and Why Implicit Evaluations Can Be Rapidly Revised. in \u003cem\u003eAdvances in Experimental Social Psychology\u003c/em\u003e vol. 56 131\u0026ndash;199 (Elsevier, 2017).\u003c/li\u003e\n\u003cli\u003eYarkoni, T. The generalizability crisis. \u003cem\u003eBehav. Brain Sci.\u003c/em\u003e \u003cstrong\u003e45\u003c/strong\u003e, e1 (2022).\u003c/li\u003e\n\u003cli\u003eGrice, H. P. Logic and conversation. in \u003cem\u003eSpeech acts\u003c/em\u003e 41\u0026ndash;58 (Brill, 1975).\u003c/li\u003e\n\u003cli\u003eCameron, C. D., Brown-Iannuzzi, J. L. \u0026amp; Payne, B. K. Sequential Priming Measures of Implicit Social Cognition: A Meta-Analysis of Associations With Behavior and Explicit Attitudes. \u003cem\u003ePersonal. Soc. Psychol. Rev.\u003c/em\u003e \u003cstrong\u003e16\u003c/strong\u003e, 330\u0026ndash;350 (2012).\u003c/li\u003e\n\u003cli\u003eKurdi, B. \u003cem\u003eet al.\u003c/em\u003e Relationship between the Implicit Association Test and intergroup behavior: A meta-analysis. \u003cem\u003eAm. Psychol.\u003c/em\u003e \u003cstrong\u003e74\u003c/strong\u003e, 569\u0026ndash;586 (2019).\u003c/li\u003e\n\u003cli\u003ePerugini, M., Richetin, J. \u0026amp; Zogmaister, C. Prediction of behavior. in \u003cem\u003eHandbook of Implicit Social Cognition: Measurement, Theory, and Applications\u003c/em\u003e (eds. Gawronski, B. \u0026amp; Payne, B. K.) 255\u0026ndash;277 (New York: Guilford Press, 2010).\u003c/li\u003e\n\u003cli\u003eFriese, M., Hofmann, W. \u0026amp; Schmitt, M. When and why do implicit measures predict behaviour? Empirical evidence for the moderating role of opportunity, motivation, and process reliance. \u003cem\u003eEur. Rev. Soc. Psychol.\u003c/em\u003e \u003cstrong\u003e19\u003c/strong\u003e, 285\u0026ndash;338 (2008).\u003c/li\u003e\n\u003cli\u003eGreenwald, A. G., Poehlman, T. A., Uhlmann, E. L. \u0026amp; Banaji, M. R. Understanding and Using the Implicit Association Test: III. Meta-Analysis of Predictive Validity. \u003cem\u003eJ. Pers. Soc. Psychol.\u003c/em\u003e \u003cstrong\u003e97\u003c/strong\u003e, 17\u0026ndash;41 (2009).\u003c/li\u003e\n\u003cli\u003eOswald, F. L., Mitchell, G., Blanton, H., Jaccard, J. \u0026amp; Tetlock, P. E. Predicting Ethnic and Racial Discrimination: A Meta-Analysis of IAT Criterion Studies. \u003cem\u003eJ. Pers. Soc. Psychol.\u003c/em\u003e \u003cstrong\u003e105\u003c/strong\u003e, 171\u0026ndash;192 (2013).\u003c/li\u003e\n\u003cli\u003eMalle, B. F. Attribution theories: How people make sense of behavior. in \u003cem\u003eTheories in social psychology\u003c/em\u003e vol. 23 93\u0026ndash;119 (Wiley-Blackwell, 2022).\u003c/li\u003e\n\u003cli\u003eArrieta, A. B. \u003cem\u003eet al.\u003c/em\u003e Explainable Artificial Intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible AI. \u003cem\u003eInf. Fusion\u003c/em\u003e\u003cstrong\u003e58\u003c/strong\u003e, 82\u0026ndash;115 (2020).\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Implicit Social Cognition, Competence, Human-Robot Interaction, Automaticity","lastPublishedDoi":"10.21203/rs.3.rs-5852068/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-5852068/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003ePrevious research showed people\u0026rsquo;s explicit (vs. implicit) competence impressions were more sensitive to a robot\u0026rsquo;s single inconsistent (\u0026ldquo;oddball\u0026rdquo;) behavior. We report nine pre-registered studies (N\u0026thinsp;=\u0026thinsp;3,735 online participants) testing the scope and underlying causes of this dissociation. We found that the dissociation (a) generalized to industrial robots, surgical robots, and self-driving cars; (b) replicated with structurally aligned direct and indirect measures of competence; and (c) is at least partially explained by the inconsistent evidence\u0026rsquo;s diagnosticity. We discuss implications for social cognition and human-robot interaction.\u003c/p\u003e","manuscriptTitle":"When Robots are Surprising: The Role of Cue Diagnosticity in Judging Robot Competence","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-03-11 09:30:38","doi":"10.21203/rs.3.rs-5852068/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-05-23T11:39:12+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-05-02T04:53:48+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-05-01T22:13:52+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-04-26T09:44:05+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"75835016470672657822852962461354872237","date":"2025-04-21T11:46:26+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"31695877927800393361093875620420219801","date":"2025-04-17T09:18:09+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"217767313065075211974556766829605399826","date":"2025-04-16T14:03:37+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"256846545267579400050804081377729069958","date":"2025-04-15T05:06:01+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-02-03T15:10:40+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-01-29T14:21:14+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-01-28T15:08:34+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2025-01-28T14:49:24+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"a863336b-2dff-47f8-837d-e9e490bc7079","owner":[],"postedDate":"March 11th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[{"id":44297227,"name":"Biological sciences/Psychology"},{"id":44297228,"name":"Biological sciences/Psychology/Human behaviour"}],"tags":[],"updatedAt":"2026-02-09T16:05:08+00:00","versionOfRecord":{"articleIdentity":"rs-5852068","link":"https://doi.org/10.1038/s41598-026-35375-y","journal":{"identity":"scientific-reports","isVorOnly":false,"title":"Scientific Reports"},"publishedOn":"2026-02-06 15:59:35","publishedOnDateReadable":"February 6th, 2026"},"versionCreatedAt":"2025-03-11 09:30:38","video":"","vorDoi":"10.1038/s41598-026-35375-y","vorDoiUrl":"https://doi.org/10.1038/s41598-026-35375-y","workflowStages":[]},"version":"v1","identity":"rs-5852068","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-5852068","identity":"rs-5852068","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}