RESEARCH DOSSIER
Trust Calibration in Artificial Partners
Abstract. Trust calibration in artificial partners concerns the alignment between a human decision maker's beliefs about an artificial advisor's competence and the conditions under which that advisor's outputs should guide action. This critical narrative synthesis examines classic automation research, formal distinctions between perceived trustworthiness and behavior, evidence from reliability manipulation and decision-aid experiments, algorithm aversion and appreciation, explanations and verifiability, uncertainty and confidence information, adaptive delegation, error recovery, individual differences, and the limits of human–AI complementarity. Empirical studies show that more trust or more advice acceptance can coexist with poorer discrimination of correct from erroneous recommendations; useful calibration depends on task, relative competence, consequences, opportunities to verify, and system change over time. The review develops an evidence-based research map and evaluates measurement, causal inference, ecological validity, and design implications without claiming a new empirical experiment, preregistered systematic review, meta-analysis, or original psychological theory.
Keywords: trust calibration; human–AI trust; appropriate reliance; automation bias; algorithm aversion; algorithm appreciation; decision support; confidence communication; human–artificial relations; explainable AI; trust repair; cognitive forcing; complementary performance; human factors; Artificial Era
The Scientific Object of Trust Calibration
Trust calibration is the process by which a person's expectations of an artificial partner's performance become suitably responsive to evidence about that partner's capabilities, limitations, and operating conditions. The psychological object is a developing relation among belief, observed behavior, task demands, and consequences. Calibration requires specifying which decision should be delegated, under what conditions, and relative to which available alternative. A person can trust an advisor in general yet correctly reject its answer to a particular question; conversely, a skeptical user can act on a useful recommendation after verification. Lee & See (2004) places appropriate reliance, rather than maximal trust, at the center of automation design.
Artificial partners include decision-support algorithms, interactive assistants, autonomous vehicles, robots, recommendation systems, and collaborative agents capable of suggesting or executing actions. The relevant unit of analysis is an identified human–system–task arrangement. A system's accuracy in a familiar laboratory benchmark cannot substitute for its reliability when tools, input distributions, interaction histories, and action permissions change. Researchers must describe the partner's configuration and the human's ability to inspect evidence, reject advice, or intervene. The breadth of artificial embodiments matters because users respond to visible competence, social form, and contextual expectations in different ways Glikson & Woolley (2020).
This monograph applies a structured critical narrative synthesis of original experiments, scholarly reviews, and selected recent research. Its questions are how trust and reliance develop, how they should be measured, which interventions improve discrimination between helpful and harmful advice, how failures alter subsequent cooperation, and where experimental findings generalize. It treats authorial Artificial Era terminology as context for studying contemporary partnerships, while empirical conclusions remain tied to operational behavior and available designs. The evidentiary emphasis follows the synthesis of contextual, dispositional, and learned determinants of automation trust Hoff & Bashir (2015).
Trust, Trustworthiness, Reliance, and Performance
Trust denotes an attitude or expectation that an entity will act suitably under conditions involving uncertainty and vulnerability. Trustworthiness denotes the warranted qualities of that entity as demonstrated by evidence in a specified context. Reliance is a behavior, such as accepting an answer, delegating a decision, or leaving an autonomous process in control. Performance is a consequence measured against a task criterion. These constructs influence one another but require separate observation. The early framework of Lee & See (2004) explained why reliance may be guided by trust when direct comprehension of an automation system is incomplete.
A calibration study requires an external criterion for the advice and a defensible standard for the user's choice. High acceptance can be rational when an advisor is consistently superior, and rejection can be rational when the human possesses evidence the advisor lacks. The crucial comparison is often conditional advantage: which party is expected to be more accurate or useful on this instance, given available information and verification costs? Fok & Weld (2024) separates outcome-graded appropriate reliance, judged after correctness is known, from strategy-graded reliance, judged using the decision maker's information and expectations.
Reports of trust, behavioral reliance, and decision accuracy should be presented in parallel rather than collapsed into a single score. A participant may report confidence in a technology while rarely encountering an opportunity to use it, or may follow an unwanted recommendation because institutional procedure requires doing so. Decision latency, monitoring frequency, overrides, requests for evidence, and subsequent error detection can reveal distinct aspects of interaction. The conceptual separation also protects against mistaking a charismatic interface for a well-calibrated partnership. The study of algorithmic trust therefore requires both psychological measurement and independent evaluation of the partner's actual behavior Jian et al. (2000).
The Origins of Automation Reliance Research
The intellectual lineage of trust calibration extends through supervisory control, industrial process monitoring, aviation, and human factors engineering. Earlier automation systems operated in relatively bounded settings, but they already posed the general problem of determining when people should retain control and when they should depend on technical aids. Muir & Moray (1996) reported process-control simulations examining trust and human intervention. Their findings placed trust inside a changing control relationship: participants responded to histories of machine performance and made interventions that affected the combined system rather than merely expressing detached opinions.
The vocabulary of automation use was refined by Parasuraman & Riley (1997), who distinguished use, misuse, disuse, and abuse. Misuse includes excessive reliance that undermines monitoring and independent judgment; disuse includes neglect of useful assistance; abuse concerns organizational or design choices that impose inappropriate automation. These distinctions remain important for AI because a user's apparent compliance may follow a workplace's mandatory process, time pressure, or a design that obscures the possibility of reversal. An individual-centered account without institutional constraints misses a substantial part of the phenomenon.
Historical studies also showed that failures vary in their informational consequences. An alarm's false-positive rate may encourage disabling an otherwise useful warning system, whereas an unseen failure to detect a hazard creates a different pattern of risk. Base rates, signal costs, operator workload, and detection opportunities mediate both reliance and perceptions of reliability. Consequently, an idealized percentage of correct answers has limited explanatory power in the absence of an error matrix. The older automation literature provides methodological resources for current AI studies, especially the discipline of manipulating actual reliability rather than inferring it from enthusiasm Parasuraman & Riley (1997).
Controlled Experiments on Learned Trust
Research shifted from general attitudes toward experimentally induced changes in trust by exposing people to system performance. In a simulated supervisory-control task, Muir & Moray (1996) examined how confidence and intervention changed during interactions with automation. Such paradigms offer causal leverage because investigators can vary reliability, show errors, and observe whether a human resumes manual action. They also expose a challenge: behavior can be shaped by strategy, perceived risk, or workload even when explicit trust ratings remain stable. Measurement therefore benefits from simultaneous behavioral and verbal outcomes.
The three studies reported by Dzindolet et al. (2003) used an automated decision aid for detecting camouflaged soldiers in terrain images. The authors investigated relationships among automation reliability, trust, and decisions to use the aid, showing why trust helps explain reliance behavior. Their task contained visible judgments that permitted evaluation of the human and machine contribution. At the same time, the experimental setting provided constrained stimuli and recruited students; a transfer to medicine, emergency response, or longitudinal conversational systems must be supported by additional domain-specific evidence.
Learned trust develops across sequences. The location of an error within a sequence, whether an explanation accompanies it, and whether performance recovers can produce different later behavior even when aggregate reliability is identical. Research design should retain trial-level histories rather than reporting only initial and final averages. A well-calibrated participant updates in response to meaningful evidence but avoids overgeneralizing from one extraordinary failure. This framework is compatible with the dispositional and history-based distinctions tested by Merritt & Ilgen (2008), while requiring independent measurement of how accurately participants estimate the machine's capabilities.
Measuring Trust as a Psychological Construct
Psychometric studies attempt to characterize attitudes using multiple indicators instead of a single willingness-to-use item. Jian et al. (2000) developed an empirically grounded scale through elicitation, questionnaires, and comparative judgments concerning trust in automation. A multi-item instrument can identify dimensions that respondents associate with dependability, confidence, or suspicion. Yet the existence of a reliable attitude scale does not establish that its values are aligned with actual system competence. Researchers should separately validate content, factor structure, internal consistency, and relations to observed reliance in the target population.
Measurement can be contaminated by task-specific incentives. Participants asked whether they trust a classifier may interpret the question as an evaluation of its vendor, the experimenter, the interface, or the person's own confidence. Human-centered teams often operate under organizational instructions that alter the meaning of voluntary acceptance. Behavioral measures improve ecological relevance but can still confound reliance with ignorance of alternative evidence. A strong protocol obtains trust ratings at several stages while recording information access, advice acceptance, overrides, and outcome feedback. Early trust research emphasized differences between dispositional and learned contributions Merritt & Ilgen (2008).
Calibration itself requires a target function. A correlation between trust and objective overall reliability, computed over a few systems, is vulnerable to aggregation bias when conditions vary. An individual who correctly understands that an advisor is highly accurate on ordinary questions and poor on exceptional ones can be more calibrated than someone assigning a single average confidence to all outputs. The unit of analysis should therefore include item difficulty, system confidence, observed failures, and potential costs. The scale should be interpreted as psychological evidence alongside objective discrimination, not as the definitive criterion of appropriateness Lee & See (2004).
From Automation to Artificial Collaboration
Contemporary AI partners differ from classic automation in the diversity of tasks, continuous conversation, generative output, and potential changes to model configuration. These features expand the space of judgments about competence. A highly fluent system can produce plausible language across domains while performing unevenly on specific factual, predictive, or practical tasks. The user encounters competence through interface behavior rather than directly inspecting the full training and evaluation history. Glikson & Woolley (2020) identifies representation and perceived machine intelligence among influential dimensions in the empirical literature on human trust in AI.
The transition from passive aid to collaborating agent also introduces long sequences of delegation: searching for data, selecting a tool, drafting a recommendation, executing a transaction, or making a revision after feedback. Calibrated trust becomes distributed across stages, each with different evidence and opportunities for correction. The principal empirical question is whether a person knows when to request verification and when to permit the system to proceed. Trust repair scholarship anticipates this challenge by examining how increasingly autonomous systems alter expectations after errors and how recovery efforts may affect human cooperation de Visser et al. (2018).
Human–artificial partnership should be studied through the relationship among human judgment, artificial capability, and shared outcomes. An artificial system's fluent response is an observable behavior; conclusions about inner experiences or interpersonal reciprocity require distinct evidence. Psychological studies can rigorously investigate expectations, attachment, perceived agency, and delegation without allowing those attributions to replace task measures. An integrated research architecture links cognitive psychology, social attribution, human factors, and human–AI interaction, retaining a common empirical demand: the purported relationship must be measured in sequences of decisions and consequences Bansal et al. (2019).
Dispositional, Situational, and Learned Trust
The most influential integrative taxonomy distinguishes stable personal tendencies, contextual influences, and learned expectations. Hoff & Bashir (2015) reviewed 101 papers containing 127 eligible studies conducted from 2002 to June 2013 and organized findings into dispositional, situational, and learned trust. Dispositional trust may shape how a user approaches a newly encountered advisor. Situational factors include workload, urgency, apparent stakes, and the presence of alternatives. Learned trust reflects experience with actual system behavior. These layers can interact: the same error may have very different consequences for novices, domain experts, and users whose initial confidence in technology differs.
Studies that treat user characteristics only as nuisance covariates risk confusing the effect of reliability with the participant's baseline expectations. Merritt & Ilgen (2008) used an X-ray luggage-screening task and investigated dispositional and history-based trust. Their study supports attention to individual differences while illustrating the methodological difficulty of generalizing from a student sample and a specific visual-search setting. Experimental comparisons should record prior experience with the relevant technology, domain expertise, and familiarity with error types, then test whether these variables actually moderate behavior rather than assume an invariant psychological profile.
Trust can also depend on who is accountable for an outcome. An employee making an individually consequential decision may use an advisor differently from a participant who receives a small laboratory reward. Where institutional authority demands use of a particular tool, observed acceptance partly reflects compliance. The psychology of calibration should therefore measure discretion, reversibility, task ownership, and anticipated responsibility. Parasuraman & Riley (1997) shows why automation practice can result from system design and organizational choice alongside individual attitudes. An appropriate experiment makes these constraints visible and treats their causal roles as substantive variables.
System Reliability, Error Types, and Base Rates
A reliability figure summarizes correct outcomes over a distribution of cases. It does not tell a person which cases are unusually difficult, which error types are costly, or how performance changes under distribution shift. A classifier can achieve attractive overall accuracy while missing rare consequential events. Trust calibration requires aligning reliance with the error profile relevant to the decision, including class prevalence and false-positive and false-negative costs. The legacy of signal detection and alarm research in Parasuraman & Riley (1997) demonstrates that system use must be understood against base rates and consequences rather than a single headline measure.
In a controlled design, investigators can vary the reliability of an advisor while fixing the quality of the human's unaided performance. They can then ask whether participants become more likely to accept advice precisely in those conditions where assistance offers an advantage. Trial-by-trial manipulation should distinguish randomly scattered errors from systematic failure modes. Participants may learn that an algorithm is dependable for frequent categories but brittle for rare contexts. Trust that adapts to this conditional structure is more scientifically informative than a rise in a global attitude score after a sequence of accurate predictions Dzindolet et al. (2003).
Risk changes the criterion for acceptable reliance. False reassurance about a medication interaction, missed detection of a hazard, or erroneous financial advice may justify more verification than a low-stakes content suggestion. If verification costs time or specialized knowledge, the rational policy changes again. Studies should state the task's loss function, how errors are scored, and whose interests bear the consequences. Lee & See (2004) grounded reliance in the user's goals and situational context; this framework remains indispensable when assessing systems whose output can be accepted immediately or passed through more demanding review.
Algorithm Aversion after Observable Mistakes
Algorithm aversion describes decisions to avoid a statistical advisor after witnessing an error, even when the available evidence indicates superior overall performance. Across five studies, Dietvorst et al. (2015) examined forecasts produced by people and algorithms and found that witnessing mistakes could reduce participants' willingness to choose an algorithm. The phenomenon demonstrates that experience with errors does not guarantee optimal learning: the evidentiary weight assigned to one failure may be disproportionate. The important psychological mechanism concerns evaluation of expected competence rather than the mere recognition that the system is imperfect.
An error's meaning depends on the comparison the observer makes. People expect some mistakes from human advisers, whereas algorithms may be imagined as mechanical sources of consistency. A visible algorithmic mistake can then appear to violate a stronger prior expectation. Such a process may produce underuse of an advisor that remains superior to alternatives. However, aversion should not be assumed across every task: expert users, outcome stakes, prior successes, and ability to modify the model can change reliance. Experimental findings must be interpreted in relation to the studied forecast tasks and the alternatives participants actually observed Dietvorst et al. (2015).
Calibrated correction after error requires more than encouraging trust restoration. The system might be unreliable on a newly discovered class of cases, making reduced reliance appropriate; alternatively, the error might be a rare event with little bearing on future decisions. Distinguishing these scenarios requires conditional performance evidence. Researchers should expose comparable histories for human and artificial advisers, measure perceived performance distributions, and test whether a participant's later behavior approaches the best available choice. An emphasis on repair without validating actual competence risks reinstating misuse under the guise of resilience de Visser et al. (2018).
Algorithm Appreciation and Perceived Expertise
The opposite tendency, algorithm appreciation, has also been demonstrated experimentally. Logg et al. (2019) reported six experiments in which lay participants often weighted advice more when it was believed to originate from an algorithm rather than a human. The observed pattern depended on the comparison being made and could weaken when participants compared an algorithm's judgment with their own expertise. This evidence counters the assumption that users universally resist automated recommendations. It also shows how source labels can affect decisions independently of the objective contents of advice.
Source attribution shapes perceived expertise and responsibility. A user may infer that a model aggregates large datasets or employs rigorous calculation, even when such assumptions do not accurately characterize the system's actual performance. In a financially incentivized experimental setting, Klingbeil et al. (2024) found detrimental overreliance on AI advice, including cases where it conflicted with available contextual information. These results illustrate the possibility that confidence in technology can lead users to discount their own task-relevant evidence. They should be interpreted within the experiment's incentive structure and decision environment.
Aversion and appreciation can coexist in a population or emerge sequentially in the same user. An initial AI label can generate high expectations, while a later salient failure can reduce use disproportionately. Expertise, information about competing sources, and previous feedback determine how this trajectory develops. The empirical task is to explain the switch by measuring expectations before an error, behavioral reactions to it, and subsequent evidence accumulation. Glikson & Woolley (2020) reviews the broader literature on representations and trust antecedents, reinforcing the need to distinguish interface impressions from demonstrated reliability.
Meta-Analytic Evidence and Its Boundaries
Quantitative evidence helps estimate how multiple factors relate to trust across studies. In their meta-analysis of human–robot interaction, Hancock et al. (2011) assembled 29 empirical studies and reported effect estimates across correlational and experimental comparisons, while identifying robot performance and attributes among influential factors. Their synthesis supports the importance of how the artificial partner actually behaves. However, pooling studies with different robot roles, tasks, manipulation methods, and trust measures does not automatically establish a general law for conversational assistants or high-stakes decision support.
A later meta-analysis across automation research by Schaefer et al. (2016) analyzed experimental and correlational evidence and reported moderate overall associations with trust development, subject to different moderator categories. This is evidence that characteristics of the human and the system can influence measured trust. It is not a universal estimate for the value of a particular explanation, confidence display, or trust-building intervention. The appropriate inference depends on which outcome was measured, how effects were standardized, and whether a meta-analytic contrast corresponds to the proposed application.
Review quality also depends on study selection and publication incentives. Hoff & Bashir (2015) used explicit eligibility criteria for human-subject interactions, whereas Glikson & Woolley (2020) brought together AI trust evidence across different organizational and technological contexts. These literatures overlap but emphasize different populations and constructs. Their disagreements or apparent variations may reflect scope rather than theoretical contradiction. The strong methodological implication is to avoid transferring a pooled association to an untested deployment: a design should specify which mechanism is predicted to change and what behavioral evidence would support the proposed benefit.
The Dynamics of Error Recognition and Feedback
Trust changes only as rapidly as usable information reaches the person. Feedback can concern an outcome, a probability estimate, an explanation of an error, or a comparison with an alternative adviser. Immediate correctness feedback permits detailed learning but may be unavailable for forecasts whose outcomes emerge months later. An assistant that routinely sounds decisive can make uncertainty difficult to detect until consequences become visible. Experiments that provide unrealistically perfect feedback can thus overestimate users' capacity for real-world calibration, where errors are delayed, ambiguous, or unobserved Lee & See (2004).
Sequential studies should distinguish a user's response to a known failure from the same user's response to merely suspected failure. A wrong recommendation may be detected independently, reported by an authority, or inferred from inconsistent outputs. These forms of information differ in credibility and timing. Muir & Moray (1996) demonstrated the relevance of trust and intervention dynamics in process control; contemporary studies can extend this logic through logging, counterbalanced feedback sequences, and repeated task blocks. The unit of analysis is the trajectory of belief and action, not a one-time survey completed after interaction.
Well-designed feedback communicates both error frequency and its conditions. A statement such as 'the model was wrong' is informationally thin unless the user learns which input pattern, source limitation, or action step caused the problem. Conversely, a reassuring apology without demonstrable correction might increase favorable attitudes without improving decisions. de Visser et al. (2018) frames trust repair as important for autonomous human–machine interaction, but repair should be evaluated using subsequent behavior under verified conditions. A successful intervention helps people choose differently when future evidence warrants it; preserving attachment alone is not proof of improved calibration.
Explanations and the Risk of Uncritical Acceptance
An explanation supplied with an artificial recommendation can increase users' understanding, confidence, or sense that they have been given grounds for action. These outcomes are psychologically distinct. In controlled human–AI decision settings, Bansal et al. (2021, CHI) investigated whether explanations enable complementary team performance and found that seemingly informative explanations can increase acceptance of AI advice without necessarily helping people identify the model's mistakes. The effect of explanation is therefore an empirical question with a decision outcome, rather than an automatic consequence of making an interface more transparent.
Explanatory language can function simultaneously as evidence and as a cue to authority. A detailed rationale may persuade users through familiar concepts even when its premises are incomplete or the recommendation is incorrect. Researchers must ask whether the explanation contains information that enables a person to distinguish valid from invalid cases. A user who can restate an explanation but cannot detect a faulty recommendation has acquired a different skill from reliable error discrimination. Experimental evaluations that measure only satisfaction or perceived intelligibility will miss this distinction Fok & Weld (2024).
Explanations also change the allocation of cognitive effort. A short explanation may direct attention to the decisive evidence; a verbose one may consume time that the user would otherwise spend checking an alternative source. When the model produces an answer before the human gives an independent judgment, the initial recommendation may anchor later reasoning. These effects vary with task difficulty and the participant's expertise. Comparative designs should include advice alone, advice with explanation, and independent judgment conditions, together with measured verification behavior Bansal et al. (2021, CHI).
Verifiability, Faithfulness, and Independent Evidence
Verifiability concerns whether an ordinary user, given available time and resources, can check the grounds of an AI recommendation against independent evidence. It is a stronger practical property than perceived comprehensibility. Fok & Weld (2024) critically examines whether explanations actually enable complementary decisions, highlighting why an explanation that seems coherent may fail as an instrument for identifying erroneous AI predictions. A verifiable answer provides accessible data, assumptions, citations, or intermediate steps that bear causally on correctness and can be assessed by the intended user.
Faithfulness raises a related but different question: does the explanation accurately describe the process that produced the model's output? A post hoc story may be intelligible without faithfully representing the decision rule. In an experimental investigation of interpretable model presentations, Poursabzi-Sangdeh et al. (2021) showed that making model information more visible does not guarantee that users can detect mistakes or improve decisions; certain model displays increased apparent intelligibility while impeding correction. An empirical evaluation must separate the explanation's content validity from the user's understanding of that content.
The strongest verification protocol supplies a checkable criterion that does not simply repeat the advisor's conclusion. When a model recommends a diagnosis, legal interpretation, or financial action, the relevant evidence may involve clinical tests, governing texts, calculations, or independent observations; the available expertise and stakes vary accordingly. A system that cites another output from itself has provided additional text rather than independent verification. Studies should record what evidence participants access, whether they detect deliberately introduced faults, and how verification costs affect later reliance Fok & Weld (2024).
Transparency, Interpretability, and Cognitive Load
Transparency commonly refers to information about how a system functions, what data it uses, or how it reaches a response. Interpretability refers to a person's ability to form a workable understanding of the model's behavior. Both constructs depend on the recipient's task and prior knowledge. Kizilcec (2016) experimentally varied transparency information in an algorithmic interface and found that the amount and type of information can alter trust, showing that disclosure is not psychologically neutral. Transparency is best studied as an intervention with measurable outcomes, not as an undifferentiated design virtue.
An interface may expose feature weights, local examples, confidence levels, or a natural-language chain of reasoning. Each format imposes different cognitive demands. For simple tabular predictions, inspecting a model's parameters may appear feasible, while interpreting nonlinear interactions can exceed a nonexpert's working-memory resources. Poursabzi-Sangdeh et al. (2021) tested how participants used simplified and more elaborate model presentations; their findings caution against equating a display of formal detail with the ability to identify error. To evaluate interpretability, a study should measure actual task comprehension, appropriate corrections, and time required.
Cognitive load interacts with stakes, interruptions, and the number of decisions processed. A person reviewing hundreds of suggestions may shift from deliberate scrutiny to pattern-based acceptance even if each suggestion includes the same evidence. Transparency mechanisms that work for one case may degrade when repeated under time pressure. Researchers can test this through within-participant workload manipulation, randomized interface order, eye or interaction tracking where justified, and delayed retention tests. The resulting evidence should show whether users acquire transferable rules for deciding when to rely rather than merely reporting that a display looked informative Lee & See (2004).
Cognitive Forcing and Deliberative Evaluation
Cognitive forcing interventions modify the sequence of advice use to promote independent thought. Examples include asking a person to record an initial judgment before revealing an AI recommendation, requiring an explicit comparison of alternatives, or inserting a brief deliberative step before acceptance. In experiments with AI-assisted decisions, Buçinca et al. (2021) evaluated such functions and found that they can reduce overreliance. Their results show that the order and structure of interaction can influence decisions independently of model accuracy, offering a direct approach to the behavioral side of calibration.
These interventions also have costs. Requiring an unaided first answer consumes time, can be frustrating in routine cases, and may be inefficient when the algorithm has a clear advantage. The proper evaluation therefore compares both error reduction and the resources spent to achieve it. A forcing function succeeds when its additional scrutiny improves expected outcomes under the relevant loss function, especially on cases the advisor is likely to mishandle. It should not be assessed solely by a fall in acceptance rate, because decreased reliance can be harmful when correct advice is rejected Buçinca et al. (2021).
An important design distinction concerns forced effort and informative effort. A mandatory click is easily completed without reflection, whereas a prompt to identify missing evidence can direct attention to a specific failure mode. Experiments can compare these interventions at equal time cost and include a withdrawal phase to test learning. If participants resume indiscriminate acceptance once the interface requirement disappears, the intervention changed immediate behavior without necessarily improving durable calibration. The relation to learned trust and retained competence should be measured longitudinally Hoff & Bashir (2015).
Confidence Displays and Meaningful Uncertainty
An AI-generated confidence score is useful only to the extent that its statistical meaning is specified and empirically validated. A number may refer to a class probability, a calibrated event frequency, a ranking score, or the model's self-reported certainty; these quantities have different interpretations. The human decision maker may confuse confidence with accuracy, especially when the interface presents precise numbers without specifying what events they describe. Lee & See (2004) argues that suitable reliance depends on appropriate information about system performance and the task context, which makes the measurement validity of confidence displays central.
Research on automation frequently distinguishes agreement with advice from understanding when the advice is reliable. A confidence indicator can support conditional reliance when it discriminates well among errors and correct outputs. However, a display derived from poorly calibrated model scores may create misplaced precision. Calibration should be assessed on a held-out distribution and rechecked after model or context changes. In work on human–AI teamwork, Bansal et al. (2021) demonstrates why optimizing overall model accuracy is insufficient for optimizing joint outcomes; confidence information likewise needs evaluation against team decisions rather than the model in isolation.
Uncertainty communication may also take verbal forms, ranges, scenario descriptions, abstention, and prompts for verification. The most appropriate form depends on the user's statistical understanding and the decision's stakes. A verbal hedge can be interpreted differently across audiences, while a numeric percentage may conceal sampling limitations or irreducible uncertainty. Empirical studies should compare formats on comprehension, discrimination, and downstream accuracy while reporting participant numeracy and prior experience. Observed changes in stated trust are insufficient unless they correspond to improved decisions Klingbeil et al. (2024).
Designing Interfaces for Calibrated Partnership
A calibration-centered interface provides a person with the evidence required to decide how much control to retain at each stage of collaboration. It communicates scope of competence, meaningful uncertainty, source provenance, and routes for verification and correction. Those properties are evaluated through behavior rather than assumed from an attractive design. The complementary-performance studies of Bansal et al. (2021, CHI) show why an explanation that feels persuasive should be tested against the counterfactual of unaided human decisions and against the simpler baseline of advice without explanation.
Interfaces can also help communicate the consequences of system updates. An advisor that changes its prediction policy may improve average performance but become less compatible with the patterns users have learned. The design should therefore preserve visibility into changed capabilities and provide opportunities for relearning. Bansal et al. (2019) examines the performance–compatibility tradeoff associated with model updates and illustrates why developers need to consider the human team's adaptation history. Update notices, versioned evidence, and task-specific reliability summaries can be evaluated as interventions that support informed delegation.
Finally, good design exposes an effective means of refusal and recovery. A human who cannot halt a robot, correct a recommendation, or inspect a consequential action possesses little practical control even if an interface requests confidence ratings. Measures should include interruption success, reversibility, correction time, and the accuracy of subsequent choices. Trust repair is most credible when a system's changed behavior can be demonstrated under the relevant conditions, a question already central to the autonomy and recovery literature de Visser et al. (2018).
Complementary Performance and Relative Competence
An artificial partner is valuable in a joint task when the team achieves better outcomes than the available alternatives, given its costs and constraints. Complementarity requires matching system assistance to instances where it improves on human judgment and allowing the person to contribute where their evidence or skills are stronger. Overall algorithmic accuracy does not by itself establish this pattern. Bansal et al. (2021) examines whether the most accurate AI is necessarily the best teammate and models the importance of human–AI team performance as an objective. The central comparison is among whole decision processes, including advice acceptance, independent correction, and the cost of interaction.
Complementarity is difficult when error patterns are correlated. If human and artificial participants tend to fail on the same ambiguous cases, combining them may add little. Conversely, a less accurate artificial model may make different kinds of mistakes that a knowledgeable user can recognize, supporting better joint choices. The performance of the partnership depends on the person's capacity to predict or recognize model failure, which in turn depends on the available signals. Bansal et al. (2021, CHI) found that explanation interventions do not necessarily yield the complementary gains that their theoretical promise suggests.
A rigorous evaluation includes four baselines: unaided humans, the AI system acting alone, humans exposed to AI recommendations, and the best feasible alternative workflow. It should report accuracy or utility, time, verification costs, error severity, and heterogeneity across case types. A partnership that improves average accuracy but sharply worsens a small set of high-harm cases deserves a more detailed assessment. The resulting goal is a calibrated division of labor rather than maximizing the proportion of AI recommendations accepted Lee & See (2004).
Continuity and Change after System Updates
Trust develops with a particular system's history. An update can improve a benchmark score while changing the kinds of advice that previously appeared reliable. Users may continue to rely on patterns learned under the older system, and their skill at identifying mistakes may become less accurate. In research on updating human–AI teammates, Bansal et al. (2019) studies the tradeoff between performance improvements and behavioral compatibility. This evidence highlights the psychological relevance of model changes: a new version is encountered through the user's existing expectations, not as a fresh, unlearned object.
Longitudinal calibration therefore requires version-level evidence. Studies should record which model produced each recommendation, what task distribution was encountered, whether the user knew an update occurred, and whether failures differed between versions. A repeated-measures protocol could test the same user before and after an update, while a control group encounters only the final version. Outcomes should include global trust, instance-specific reliance, speed of relearning, and corrective intervention. Without these controls, changes attributed to improved AI may actually reflect practice effects or altered task composition Bansal et al. (2019).
Updates also generate institutional questions about accountability. A user relying on an advisor for months may expect predictable constraints and documented changes. Product labels stating that a model is improved offer little task-specific evidence. Controlled change logs, deployment evaluations, and accessible performance notices allow a person to revise expectations, but their effectiveness requires direct testing. The broader human-factors framework of Hoff & Bashir (2015) implies that learned trust is a product of specific experience and that relevant prior experience must be treated as part of the experimental setting.
Trust Repair, Responsibility, and Reversibility
Trust repair refers to changes in expectations and behavior following a perceived breach of system reliability or cooperative norms. In human–machine interaction, de Visser et al. (2018) highlights the importance of understanding repair as automation becomes more autonomous. Repair can involve acknowledging errors, providing evidence of correction, adjusting the operational scope, or allowing an operator to inspect and reverse decisions. These mechanisms differ in their implications: a polite admission may improve the experience of interaction, while a verified change in performance provides stronger grounds for renewed reliance.
Experimental trust-repair research should specify the type of failure, whether it was detected by the user, and what new evidence accompanies the repair attempt. Participants can be randomized to apology-only, explanation-only, independently verified correction, and no-intervention conditions while holding the subsequent error profile constant. Measurements should capture choices across later tasks rather than stopping at a single self-report question. Such designs address the causal question of whether repaired beliefs are aligned with actual performance de Visser et al. (2018).
Organizational accountability influences the psychological meaning of intervention. A human operator formally responsible for a final decision may be unable to inspect a system's supporting data; a delegated agent may act before the person can respond. The presence of a nominal override control is insufficient when delays, knowledge gaps, or irreversible actions make it ineffective. Parasuraman & Riley (1997) situates automation misuse and abuse within system design and organizational practice. Calibrated partnership therefore requires analysis of actionable control as well as perceived trust.
Decision Domains, Expertise, and Consequences
AI advice enters medical, legal, educational, financial, and ordinary informational decisions, each with different evidence standards and consequences. The research methods of one domain cannot automatically be transferred to another. Experimental findings on reliance in human decision making, including Pearson et al. (2026), provide useful evidence about how artificial advice affects choices under defined task conditions. Their substantive scope must be read alongside the population, instructions, advice quality, and criterion used. A general conclusion about all professional judgments would exceed what a bounded experiment establishes.
In high-stakes decisions, a modest average accuracy improvement may coexist with unacceptable errors affecting rare groups or exceptional cases. The underlying decision criterion should include unequal losses, delayed outcomes, availability of independent review, and the human professional's ability to recognize inappropriate advice. Klingbeil et al. (2024) shows that excessive reliance can impose measurable costs in a controlled choice environment. Whether an intervention improves decisions under clinical or legal conditions requires domain-specific validation and oversight appropriate to the consequences.
Expertise can both protect against error and produce unwarranted confidence. An experienced practitioner may know when local circumstances contradict a model's general recommendation, yet may also overlook a valid new signal because it conflicts with habitual practice. A novice may be willing to consult the model but lack the skill needed to verify it. Calibration experiments need to stratify expertise and compare humans to a relevant alternative workflow. The principal assessment is whether the team arrives at better justified outcomes with known error rates, rather than whether professionals report higher trust Lee & See (2004).
Artificial Era and the Longitudinal Psychology of Partnership
Within the historical horizon of the Artificial Era, artificial partners are increasingly encountered across work, study, and personal decision routines. The psychological research object includes accumulated expectations and patterns of shared action across repeated encounters. Scientific analysis can examine whether sustained interaction changes monitoring habits, delegation thresholds, independent skill, and confidence in one's own judgment. Such questions extend the earlier tradition of automation trust while retaining its empirical grounding in tasks, observable behavior, and documented system performance Hoff & Bashir (2015).
A longitudinal study should follow people through changing tasks and model versions, recording their unaided performance as well as outcomes with assistance. Withdrawal tests can reveal whether high combined performance masks reduced individual skill. Repeated exposure can also change the availability heuristic: salient recent system errors may dominate estimated risk even after a record of successful recommendations. Model updates, social presentation, and verification opportunities should be recorded as time-varying conditions. The performance–compatibility findings of Bansal et al. (2019) indicate why historical interaction matters when users encounter a supposedly improved partner.
The broader conceptual vocabulary of human–artificial relations invites questions about attachment, cognitive partnership, and perceived agency. For this dossier, those phenomena are examined insofar as they change expectations, evidence use, and choices. Philosophical positions regarding Artificial Sapiens and the historical role of Artificial belong to the ecosystem's authorial canon; psychological claims require independent behavioral tests. A mature scientific account permits both a general conceptual horizon and precise empirical limits. The goal is to understand how people regulate dependence on increasingly capable artificial collaborators over time Glikson & Woolley (2020).
Research Designs for Trust Calibration
The cleanest evidence on causal influences in calibration comes from experiments that independently manipulate a property of the advisor and observe subsequent human choices. Relevant properties include true reliability, the distribution of error types, explanation format, confidence displays, and the timing of feedback. Dzindolet et al. (2003) used a structured decision-aid task to examine trust and reliance, while Muir & Moray (1996) studied interventions in a simulated process-control setting. Their methodological significance lies in specifying tasks, conditions, system performance, and behavior before interpreting reported trust.
Research should separate initial trust formation, adaptation after experience, and stability under changing conditions. Between-participant randomization helps estimate interface effects without contaminating one condition with another; within-participant sequences reveal learning but invite carryover effects. Where trial order matters, researchers should preregister primary contrasts, randomize or counterbalance order, and analyze serial dependence. Participants should encounter both correct and incorrect advice in proportions that reflect the intended inference. Deceptive performance histories require ethical review and careful debriefing in applicable experimental settings Hoff & Bashir (2015).
Field studies offer ecological information about workloads, institutional mandates, and real delays between advice and outcomes. Their causal interpretation is complicated by self-selection and shifting environments. Randomized field interventions, staggered deployments, and carefully specified quasi-experimental comparisons can strengthen inference where feasible. Researchers must record who could ignore the tool, who bore the risk of mistakes, what alternatives existed, and whether the system changed during observation. Experimental and observational studies complement one another when their differences in identification and generalization are stated precisely Glikson & Woolley (2020).
Operational Measures of Appropriate Reliance
One common empirical measure is advice acceptance: the proportion of recommendations a participant follows. This outcome captures behavior but does not identify appropriateness. A more informative design categorizes trials by whether the advice is correct, whether the participant's unaided answer would be correct, and whether the person possesses a feasible independent check. Fok & Weld (2024) examines the distinction between choosing correctly in hindsight and using an informed decision strategy. An acceptance rate can rise while joint performance falls if the person becomes more likely to accept erroneous advice.
Researchers may define appropriate reliance as accepting correct advice and rejecting incorrect advice, but that rule requires qualifications when correctness is uncertain, outcomes are graded, or the human is less reliable than the model. A utility-based criterion combines outcome probabilities, consequences, alternative decisions, and verification costs. Under a specified information set, a decision can be rational even when the eventual outcome is unfavorable. Reporting both a retrospective correctness measure and a prospective information-sensitive measure prevents causal claims from depending entirely on hindsight Lee & See (2004).
Measurement should also distinguish discrimination from response tendency. Users can become less willing to trust any advice while improving neither their ability to identify AI errors nor their task accuracy. Conditional acceptance rates, sensitivity to known failure cues, calibration of estimated performance, and time spent checking evidence offer complementary observations. Bansal et al. (2021, CHI) demonstrates that team-performance evaluations must compare the effects of explanations against unassisted and simpler advisory conditions. A complete results table includes human-only performance, AI-only performance, assisted performance, and the associated decision costs.
Construct Validity, Confounding, and Measurement Bias
Construct validity begins with a definition of the psychological quantity under investigation. If investigators label a scale of favorable attitudes 'calibration' but never measure true advisor competence or appropriate decisions, their study has measured perceptions rather than calibration. Instruments such as the trust scale developed by Jian et al. (2000) can assess attitudes when validated for the sampled context. They cannot replace outcome measures. A rigorous design states whether the causal hypothesis concerns perceived trustworthiness, dependence, behavior, error detection, or the quality of joint decisions.
Potential confounders include interface fluency, model branding, task difficulty, participant literacy, perceived surveillance, payment incentives, and opportunities to inspect external evidence. For example, an explanation condition that adds both text and time may change accuracy through additional deliberation rather than the specific explanatory information. Poursabzi-Sangdeh et al. (2021) illustrates why greater apparent transparency need not improve users' performance. Researchers can use length-matched controls, matched evaluation time, attention checks appropriate to the task, and separate manipulation checks for beliefs about model competence.
Selection into a trial may also bias observed relationships. People who readily accept AI advice can differ from skeptical participants in confidence, expertise, or stakes; a correlation between trust and performance need not show that trust caused the performance difference. Repeated trials with the same participant create dependence that requires clustered or hierarchical inference. When many moderators are explored, the risk of chance findings increases. Transparent sample descriptions, prespecified endpoints, complete reporting of negative results, and sensitivity analyses make the evidence more useful for synthesis Hoff & Bashir (2015).
Evidence Quality, Replication, and Generalizability
The trust-calibration literature spans experiments with relatively simple advice, simulations of operational control, interactive decision tasks, and studies of robotic agents. These settings offer complementary forms of evidence but differ in what they establish. A laboratory task can isolate a source-label effect, whereas sustained deployments reveal whether that effect survives organizational demands, delayed consequences, and evolving tools. The review by Glikson & Woolley (2020) encompasses heterogeneous empirical traditions, while Hancock et al. (2011) and Schaefer et al. (2016) synthesize narrower quantitative sets. Their conclusions should be read through inclusion criteria and study designs.
Reproducibility requires access to the exact model or decision aid, input set, output history, interface condition, participant instructions, and scoring rules. Two systems with the same generic brand name may perform differently after updates; even a stable model can behave differently after a change in context or tool access. Independent replication should therefore identify the tested configuration and measure variation across prompts, cases, and users. For high-stakes decisions, outcome-based evaluations also need a defensible ground truth and credible adverse-event assessment. Generalization beyond these conditions remains an empirical question Bansal et al. (2019).
The evidence assembled here supports robust distinctions among trust, reliance, and outcome quality, and it documents multiple mechanisms by which source presentation, visible errors, explanations, and cognitive effort alter decisions. It provides fewer grounds for claiming a universal best level of trust or a single successful interface technique across all artificial partners. Emerging studies of human reliance Pearson et al. (2026) should be integrated into future cumulative reviews with attention to task characteristics, publication status, and independent replication. This monograph is a critical narrative synthesis; it does not claim a reproducible exhaustive search, formal risk-of-bias scoring, or a pooled statistical effect.
Integrated Findings and Unresolved Research Problems
The central synthesis is that appropriate partnership depends on relative task competence, access to diagnostic evidence, and the user's ability to adjust action as conditions change. Lee & See (2004) supplied the conceptual emphasis on appropriate reliance; experiments on algorithm aversion Dietvorst et al. (2015) and appreciation Logg et al. (2019) demonstrate different tendencies in response to labels and observed failures. Studies of explanations Bansal et al. (2021, CHI) and cognitive forcing Buçinca et al. (2021) reveal that interface changes can improve or worsen advice use depending on how they affect independent judgment. The full evidence therefore favors conditional, behaviorally assessed calibration over an attempt to maximize favorable feelings toward automation.
Unresolved questions concern long-term effects, cross-task transfer, changing models, and unequal error costs. Researchers need longitudinal trials that measure users before assistance, throughout sustained collaboration, after unexpected failures, and when the model or workflow changes. They also need comparative studies in which explanations, verification aids, confidence displays, and deliberation prompts are evaluated against the same task-specific outcomes. The managerial research proposal discussed by Atanassova et al. (2026) illustrates practical interest in when teams should override AI; the causal effectiveness of any recommended routine must still be established under representative conditions.
Additional priorities include demographic and cultural variation in interpreting uncertainty, the effect of institutional power on nominally voluntary delegation, and the influence of conversational social cues on perceived expertise. The proper unit of evaluation may expand from a single answer to an entire chain of actions, including retrieval, tool use, execution, and recovery. Independent evaluators should document which error opportunities were accessible to users and whether observed improvements survive the removal of assistance. The contribution of this synthesis is an organized account of published knowledge and research gaps rather than an asserted new psychological mechanism Glikson & Woolley (2020).
Conclusion
Trust calibration in artificial partners is the experimentally examinable alignment of expectations and reliance with a system's conditional capabilities, limitations, and consequences for a particular decision. Historical research on supervisory control establishes why trust is connected to intervention and action. Contemporary experiments show how observing mistakes, receiving an algorithmic recommendation, encountering explanations, and facing interface constraints can alter human choices. The resulting scientific object encompasses beliefs, learning, monitoring, and responsibility within a human–artificial decision system Muir & Moray (1996) Lee & See (2004) Bansal et al. (2021, CHI).
Evidence supports a practical research standard: compare unaided humans, artificial advice, and joint outcomes; observe when participants follow correct and incorrect recommendations; evaluate available independent information; and measure how choices change following feedback or model updates. A claim that a system improves trust should specify whether it changes attitudes, error detection, appropriate reliance, joint utility, or the capacity to recover from failures. This standard protects scientific interpretation from confusing persuasive presentation with reliable assistance or systematic skepticism with optimal human control Fok & Weld (2024) Buçinca et al. (2021).
The next stage of research requires sustained, cross-domain investigations with explicit task utilities, versioned systems, transparent measurement, and independent replication. Questions about enduring cognitive partnership, social attribution, and the Artificial Era remain scientifically productive when translated into observable expectations and consequences. Psychology can investigate the conditions under which human and artificial partners learn to coordinate their strengths while identifying the responsibilities and verification resources that make delegated action appropriate. The enduring outcome is an evidence-centered account of collaboration and its limits Hoff & Bashir (2015) Glikson & Woolley (2020) Pearson et al. (2026).
References
- Atanassova, I., Khan, H., & Khan, Z. (2026). When should your team override AI? A trust calibration routine. Academy of Management Perspectives. Advance online publication. https://doi.org/10.5465/amp.2025.0019
- Bansal, G., Nushi, B., Kamar, E., Weld, D. S., Lasecki, W. S., & Horvitz, E. (2019). Updates in human-AI teams: Understanding and addressing the performance/compatibility tradeoff. Proceedings of the AAAI Conference on Artificial Intelligence, 33(1), 2429–2437. https://doi.org/10.1609/aaai.v33i01.33012429
- Bansal, G., Nushi, B., Kamar, E., Horvitz, E., & Weld, D. S. (2021). Is the most accurate AI the best teammate? Optimizing AI for teamwork. Proceedings of the AAAI Conference on Artificial Intelligence, 35(13), 11405–11414. https://doi.org/10.1609/aaai.v35i13.17359
- Bansal, G., Wu, T., Zhou, J., Fok, R., Nushi, B., Kamar, E., Ribeiro, M. T., & Weld, D. S. (2021). Does the whole exceed its parts? The effect of AI explanations on complementary team performance. Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems, Article 81, 1–16. https://doi.org/10.1145/3411764.3445717
- Buçinca, Z., Malaya, M. B., & Gajos, K. Z. (2021). To trust or to think: Cognitive forcing functions can reduce overreliance on AI in AI-assisted decision-making. Proceedings of the ACM on Human-Computer Interaction, 5(CSCW1), Article 188, 1–21. https://doi.org/10.1145/3449287
- de Visser, E. J., Pak, R., & Shaw, T. H. (2018). From 'automation' to 'autonomy': The importance of trust repair in human-machine interaction. Ergonomics, 61(10), 1409–1427. https://doi.org/10.1080/00140139.2018.1457725
- Dietvorst, B. J., Simmons, J. P., & Massey, C. (2015). Algorithm aversion: People erroneously avoid algorithms after seeing them err. Journal of Experimental Psychology: General, 144(1), 114–126. https://doi.org/10.1037/xge0000033
- Dzindolet, M. T., Peterson, S. A., Pomranky, R. A., Pierce, L. G., & Beck, H. P. (2003). The role of trust in automation reliance. International Journal of Human-Computer Studies, 58(6), 697–718. https://doi.org/10.1016/S1071-5819(03)00038-7
- Fok, R., & Weld, D. S. (2024). In search of verifiability: Explanations rarely enable complementary performance in AI-advised decision making. AI Magazine, 45(3), 317–332. https://doi.org/10.1002/aaai.12182
- Glikson, E., & Woolley, A. W. (2020). Human trust in artificial intelligence: Review of empirical research. Academy of Management Annals, 14(2), 627–660. https://doi.org/10.5465/annals.2018.0057
- Hancock, P. A., Billings, D. R., Schaefer, K. E., Chen, J. Y. C., de Visser, E. J., & Parasuraman, R. (2011). A meta-analysis of factors affecting trust in human-robot interaction. Human Factors, 53(5), 517–527. https://doi.org/10.1177/0018720811417254
- Hoff, K. A., & Bashir, M. (2015). Trust in automation: Integrating empirical evidence on factors that influence trust. Human Factors, 57(3), 407–434. https://doi.org/10.1177/0018720814547570
- Jian, J.-Y., Bisantz, A. M., & Drury, C. G. (2000). Foundations for an empirically determined scale of trust in automated systems. International Journal of Cognitive Ergonomics, 4(1), 53–71. https://doi.org/10.1207/S15327566IJCE0401_04
- Kizilcec, R. F. (2016). How much information? Effects of transparency on trust in an algorithmic interface. Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems, 2390–2395. https://doi.org/10.1145/2858036.2858402
- Klingbeil, A., Grützner, C., & Schreck, P. (2024). Trust and reliance on AI—An experimental study on the extent and costs of overreliance on AI. Computers in Human Behavior, 160, Article 108352. https://doi.org/10.1016/j.chb.2024.108352
- Lee, J. D., & See, K. A. (2004). Trust in automation: Designing for appropriate reliance. Human Factors, 46(1), 50–80. https://doi.org/10.1518/hfes.46.1.50_30392
- Logg, J. M., Minson, J. A., & Moore, D. A. (2019). Algorithm appreciation: People prefer algorithmic to human judgment. Organizational Behavior and Human Decision Processes, 151, 90–103. https://doi.org/10.1016/j.obhdp.2018.12.005
- Merritt, S. M., & Ilgen, D. R. (2008). Not all trust is created equal: Dispositional and history-based trust in human-automation interactions. Human Factors, 50(2), 194–210. https://doi.org/10.1518/001872008X288574
- Muir, B. M., & Moray, N. (1996). Trust in automation. Part II. Experimental studies of trust and human intervention in a process control simulation. Ergonomics, 39(3), 429–460. https://doi.org/10.1080/00140139608964474
- Parasuraman, R., & Riley, V. (1997). Humans and automation: Use, misuse, disuse, abuse. Human Factors, 39(2), 230–253. https://doi.org/10.1518/001872097778543886
- Pearson, J., Dror, I. E., Jayes, E., Whordley, G.-R., Mason, G., & Nightingale, S. (2026). Examining human reliance on artificial intelligence in decision making. Scientific Reports, 16, Article 5345. https://doi.org/10.1038/s41598-026-34983-y
- Poursabzi-Sangdeh, F., Goldstein, D. G., Hofman, J., Vaughan, J. W., & Wallach, H. (2021). Manipulating and measuring model interpretability. Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. https://www.microsoft.com/en-us/research/publication/manipulating-and-measuring-model-interpretability/
- Schaefer, K. E., Chen, J. Y. C., Szalma, J. L., & Hancock, P. A. (2016). A meta-analysis of factors influencing the development of trust in automation: Implications for understanding autonomy in future systems. Human Factors, 58(3), 377–400. https://doi.org/10.1177/0018720816634228
Social Presentation, Agency Attribution, and Trust
Artificial systems may speak conversationally, adopt social roles, show faces, or appear embodied in robotic form. These cues influence how people infer competence, intention, warmth, and accountability. The empirical synthesis of Glikson & Woolley (2020) examines trust in artificial intelligence across multiple representations and distinguishes factors related to the machine from those related to users and contexts. A socially engaging presentation can make advice easier to approach, but perceived friendliness is not evidence that the model has access to better information or makes fewer errors.
Human–robot research provides a complementary perspective. The meta-analysis by Hancock et al. (2011) links robot-related characteristics and behavior to measures of trust in interaction. Embodiment can make the machine's actions and limitations observable, while also encouraging familiar social expectations. The precise psychological effect depends on task, perceived competence, risk, and the human's capacity to intervene. A laboratory participant interacting briefly with a small robot should not be assumed to behave like someone making enduring decisions with an invisible software assistant.
Attribution matters because it can change the threshold at which a person asks for proof. A model framed as an expert may invite deference; one framed as an uncertain assistant may encourage checking. Yet labels may also mislead in the opposite direction, discouraging reliance on competent systems. Researchers can randomize social presentation while holding underlying recommendation quality constant, then assess whether belief, reliance, and accuracy move together. The purpose is to isolate the causal effect of framing rather than to presume either anthropomorphism or impersonality is always superior Logg et al. (2019).