RESEARCH DOSSIER
Epistemic Authority and Decision Delegation to AI
Abstract. Epistemic authority in human–AI interaction arises when a model's output is granted privileged weight as evidence, interpretation, or practical guidance. This monograph examines the transition from receiving automated advice to transferring cognitive control or action authority, distinguishing objective competence, perceived expertise, trust, behavioral reliance, delegation, and accountability. It reconstructs the human-factors tradition of automation bias and appropriate reliance, the experimental literatures on algorithm aversion and appreciation, human–AI complementarity, clinical and economic advice, generative interfaces, and ethical machine delegation. Primary studies are evaluated for experimental design, measured outcomes, domain transfer, and competing explanations; systematic reviews are used for evidence integration. A core finding is that both rejecting capable systems and following unreliable ones can arise from context-dependent expectations, error experience, cognitive effort, framing, institutional incentives, and limited verifiability. The research specifies behavioral metrics, selective reliance designs, causal tests, and safeguards for asymmetric information and authority. This is a critical narrative synthesis of existing scholarship, not an independent experiment or a protocol-governed systematic review.
Keywords: epistemic authority; decision delegation; artificial intelligence; automation bias; algorithm aversion; algorithm appreciation; trust calibration; appropriate reliance; human–AI complementarity; advice taking; cognitive agency; decision responsibility; explainable AI; Artificial Era
The Research Object: From Epistemic Weight to Delegated Choice
Epistemic authority in human–artificial interaction is a relationship in which a system's statements receive privileged standing as grounds for belief, interpretation, or choice. The observable phenomenon is the weight assigned to a particular source under uncertainty. A person may treat a generated answer as a tentative hypothesis, as authoritative testimony, as evidence sufficient to stop searching, or as an instruction to implement. Each transition can be operationalized through information search, confidence reports, choice changes, verification behavior, and intervention at decision time. Psychological inquiry therefore begins with observable patterns of deference rather than with any assumed intrinsic authority of an artifact Lee & See (2004) Glikson & Woolley (2020).
The object of this dossier combines three separable processes. Epistemic deference means granting an answer evidential weight; decision delegation means allocating selection of a course of action to a system; execution delegation means permitting that choice to produce consequences without contemporaneous human approval. The distinctions matter because accepting a factual suggestion can leave the ultimate value judgment with the user, while outsourcing a ranking, triage decision, investment action, or communication may determine outcomes before a person notices. Parasuraman, Sheridan, and Wickens classified automation by information acquisition, information analysis, decision selection, and action implementation, offering a functional basis for describing these stages Parasuraman et al. (2000).
Trust is an attitude about likely performance and dependability, whereas reliance is a behavioral response to a system's output. Perceived expertise concerns an observer's estimate of competence within a domain. Authority concerns the position accorded to that source in the decision process; accountability concerns who must answer for effects. These variables can move independently. Users may report doubts yet adopt a recommendation under time pressure, and institutions may mandate software usage regardless of personal confidence. Accordingly, a questionnaire measuring generalized trust cannot substitute for a study measuring acceptance of incorrect advice, autonomous verification, and consequential action Lee & See (2004) Hoff & Bashir (2015).
The central questions are when AI advice changes human beliefs and decisions, why users delegate to it, how that reliance tracks actual capability, and which technical or social conditions preserve effective oversight. This synthesis critically reads laboratory studies, field studies, review literature, and design research without making claims to an original randomized intervention. It treats the available evidence as task- and population-specific and analyzes its broader theoretical significance. The dossier belongs jointly to Consumer and Economic Psychology, Postsubjective Psychology, and Psychology of Human–Artificial Relations because incentives, relational configuration, and interactional expertise all affect the psychological transfer of judgment Klingbeil et al. (2024) Gor et al. (2026).
Historical Foundations: Rational Choice, Heuristics, and Automated Expertise
The theoretical ancestry of decision delegation extends from decision theory to behavioral research on expertise and institutional choice. Simon's bounded-rationality account emphasizes the limited computational and informational resources available to human decision makers. This helps explain why an external aid can be valuable even before a person believes the aid is infallible. Delegation economizes on attention, analysis, and memory under constraints, while potentially introducing dependence on the quality of external information. The resulting problem is comparative: how much does the aid improve expected decision quality given the user's goals, costs, evidence, and alternatives Simon (1955)?
Heuristics-and-biases research demonstrated that intuitive judgments can be systematically affected by the presentation of information and the structure of uncertainty. Tversky and Kahneman's experimental program described availability, representativeness, and anchoring as judgmental processes with predictable errors. This background informed later interest in statistical models that aggregate evidence more consistently. Yet the empirical comparison of human and mechanical prediction concerns specified populations and targets; its results do not justify a general hierarchy among all human and AI judgments. A model competent at forecasting demand may perform poorly at interpreting cultural meaning or an unanticipated policy change Tversky & Kahneman (1974) Castelo et al. (2019).
By the late twentieth century, automation had moved beyond mechanical execution into monitoring, diagnosis, and guidance. Parasuraman and Riley distinguished appropriate use from misuse, disuse, and abuse. Mosier and colleagues tested decisions in aviation scenarios where automated cues could conflict with other available information. Their findings redirected attention from raw instrument accuracy to changes in operator vigilance, checking, and the organization of work. Automation could alter the human task itself by making monitoring selective and by changing what the operator expected to notice independently Parasuraman & Riley (1997) Mosier et al. (1998).
The contemporary AI literature inherited these questions while broadening the range of tasks, interfaces, and users. Statistical forecasts, diagnostic classifiers, recommenders, language assistants, and autonomous agents present different kinds of evidence and create different opportunities for correction. A generative model can offer a long fluent argument rather than a single numeric prediction; a delegated agent can act before a person inspects any proposal. Historical continuity therefore lies in the allocation of cognitive and practical functions, while modern research must measure new conversational and action interfaces with task-specific methods Parasuraman et al. (2000) Cohn et al. (2024) Köbis et al. (2025).
Authority, Competence, Legitimacy, and the Location of Control
Functional epistemic authority is produced when a user changes evidentiary standards in response to a source. If identical advice is accepted more readily when labeled as AI, a source-label effect has occurred; if it is accepted because the user has independently verified the relevant evidence, there is a different causal mechanism. Authority may arise from demonstrated reliability, institutional presentation, expert reputation, fluency, or uncertainty reduction. Each route predicts different reactions to an observed error. The investigator should therefore measure source attribution and actual accuracy separately rather than infer credibility directly from interface confidence Logg et al. (2019) Klingbeil et al. (2024).
Legitimacy introduces another layer. A service can be technically accurate yet lack a warrant to impose a particular value trade-off, while a formally authorized system can recommend a mistaken decision. Psychological studies of delegation should therefore distinguish recognition of expertise from acceptance of organizational policy and transfer of practical control. In a consumer setting, recommendations may be voluntarily adopted. In a workplace, an output may be treated as binding because a manager requires adherence. Experimental choice without institutional constraints may miss this mechanism of authority transfer Parasuraman et al. (2000) Lee & See (2004).
Allocation of authority is multidimensional. One person may choose the objective, AI may gather information and rank alternatives, the institution may define thresholds, and another employee may authorize the final action. A single binary variable such as use of AI masks this distribution. Researchers can record which actor defined the problem, controlled the data, selected evaluation criteria, interpreted conflicts, could override the recommendation, and bore the cost of an error. Such a process account captures opportunities for agency that are invisible in final-answer accuracy alone Parasuraman et al. (2000) Gor et al. (2026).
An empirically adequate theory must also handle refusal. A person who rejects a reliable model after seeing one failure may be under-reliant; a person who declines a fluent model because the task lies outside its tested domain may be appropriately selective. The normative benchmark cannot be that more AI use is always desirable. It requires known task conditions, reliability estimates, payoff structures, and the independent performance of the human decision maker. Authority is scientifically justified in a local task to the extent that its guidance improves the relevant evidential and practical outcomes Dietvorst et al. (2015) Vaccaro et al. (2024).
Levels and Moments of Delegation
Delegation begins before an answer is read. Selecting an assistant rather than independently researching a question changes which evidence will enter consideration. Accepting a displayed answer transfers evidential weight; allowing the system to choose an option transfers decision control; connecting that system to tools can transfer implementation. These decisions occur at different times and create different opportunities for feedback. Gor and colleagues directly distinguished the choice to delegate before seeing an AI result from the subsequent choice to adopt a visible suggestion in cooperative question answering Gor et al. (2026).
The cost of reversal increases across delegation stages. An inaccurate explanation can be checked before belief is fixed, whereas an automated transaction or message may create consequences before oversight resumes. Reversibility also depends on audit logs, cancellation options, and third-party effects. Scientific analysis should record deadlines, perceived urgency, failure recoverability, and whether the person can reconstruct the model's inputs. A tool's formal capacity to act and the practical ease of suspending that action are different properties of human–AI control Parasuraman et al. (2000) Köbis et al. (2025).
Delegation can be partial or contingent. An operator might authorize recommendations only under known conditions, require confirmation above a loss threshold, or permit autonomous execution for low-stakes repetitive cases. These arrangements are suitable for experimental manipulation: vary accuracy, uncertainty disclosure, oversight workload, and the presence of irreversible outcomes while measuring thresholds for human intervention. Different designs may optimize time, expected utility, error reduction, or perceived autonomy, and these outcomes need independent reporting Lee & See (2004) Parasuraman & Riley (1997).
The distinction between assistance and execution is particularly important for modern agents. Fluent language makes an action plan intelligible, but the system may lack validated judgment about the real consequences of that plan. The research object is thus not exhausted by subjective endorsement of an answer. It includes actual transactions, information transfers, refusals, and changes in external conditions. A rigorous experimental record must specify model version, tool permissions, prompts, observed actions, and whether mistakes were detected before or after an effect occurred Köbis et al. (2025) Gor et al. (2026).
Operational Definitions and Evidence Hierarchy
A defensible study measures at least four variables separately: perceived trustworthiness, objective task performance, behavioral reliance, and final human–AI outcome. Trust can be reported as expected correctness or dependability; reliance can be observed through acceptance, changes from initial answers, delegation choices, and cessation of verification. For advice-taking tasks, a shift toward advice may be quantified as the proportion of the distance between an initial estimate and the suggested estimate that a person travels. This weighting is interpretable only when the advice and estimate share a meaningful scale Logg et al. (2019) Lee & See (2004).
Appropriate reliance is case sensitive. Acceptance of correct AI advice and rejection of incorrect advice should both be recorded. A system that induces high agreement can still worsen performance when its suggestions are wrong, and a skeptical user can miss beneficial recommendations. It is useful to report conditional acceptance rates, discrimination between correct and incorrect advice, final decision quality, and the costs of verification. Bansal and colleagues' team-performance studies show why greater agreement and better explanations are not equivalent to improved complementary performance Bansal et al. (2021) Fok & Weld (2024).
Experimental evidence provides strong leverage when source labels, advice correctness, or intervention timing are randomized and when a no-advice baseline is available. However, laboratory forecasts and controlled radiology tasks differ in familiarity, training, stakes, and institutions. Field studies provide ecological detail while often limiting causal attribution. Systematic syntheses help estimate variation across studies only when task definitions and outcome metrics are compatible. A critical narrative synthesis should identify which findings are original causal comparisons, which are associations, and which are proposed theoretical mechanisms Gaube et al. (2021) Vaccaro et al. (2024).
The evidence base considered here spans aviation automation, experimental economics, marketing and judgment, medical decision aids, human-computer interaction, and conversational systems. It is deliberately interdisciplinary because human responses to AI depend on the conjunction of system capability and decision setting. Each scientific conclusion will therefore specify its comparison group and boundary conditions. Philosophical terms such as agency and epistemic authority organize questions, while conclusions about how people behave depend on controlled observation and independently interpretable measures Klingbeil et al. (2024) Glikson & Woolley (2020).
Trust Formation, Perceived Reliability, and Learned Expectations
Trust in decision automation develops through prior beliefs, observed performance, interface cues, social context, and the costs of being wrong. Lee and See describe a framework in which trust guides reliance when users cannot directly inspect every component of a complex process. Hoff and Bashir synthesize dispositional, situational, and learned influences on trust, showing why two users may respond differently to the same machine and why one user's reliance can change as evidence accumulates. These theories support longitudinal measurement of trust rather than interpreting one questionnaire score as a fixed trait Lee & See (2004) Hoff & Bashir (2015).
Evidence about performance should be matched to a recognizable task. A statement that a system is highly accurate may be persuasive, yet aggregate accuracy can hide subgroup failures, poorly calibrated confidence, or a change of input distribution. Users must learn which particular kinds of cases a system handles competently. An intervention can provide aggregate accuracy, examples of failures, or instance-specific uncertainty; each changes the information available to support reliance. Kocielnik, Amershi, and Bennett demonstrated that different types of error can produce different perceptions of an ostensibly equally accurate assistant Kocielnik et al. (2019).
Trust and reliance may diverge when verification is expensive. Under time pressure, a person can accept advice despite uncertainty because competing sources require additional work. Conversely, a user who values personal control may refrain from reliance even after recognizing demonstrated average superiority. These decisions can be studied with explicit incentives and time costs, followed by separate questions about confidence and preference. Reliance is best treated as a choice made under constraints rather than a direct readout of an internal attitude Dietvorst et al. (2018) Klingbeil et al. (2024).
Longitudinal research should measure how beliefs respond to successive successful and failed outputs and whether task-specific learning transfers to an unfamiliar domain. The key comparison is between the actual predictive value of prior evidence and the change in the user's estimate of competence. Overly slow updating may sustain misplaced authority; unusually rapid withdrawal after a minor error may produce under-use. The psychological goal is a calibrated representation of performance across relevant conditions, coupled with reversible opportunities to act on new information Lee & See (2004) Hoff & Bashir (2015).
Algorithm Aversion: Error Exposure and Withdrawal of Authority
Algorithm aversion describes reduced willingness to rely on an algorithm after observing its mistakes, even in settings where it remains more accurate than a human alternative. In five studies, Dietvorst, Simmons, and Massey compared decisions following exposure to forecasting performance. Participants often lost confidence in the algorithm more quickly than in a human after equivalent mistakes. This is evidence about a specific psychological response to visible prediction errors, not proof that all populations dislike all computational systems. The reference condition and the observed sequence are indispensable to interpreting the result Dietvorst et al. (2015).
One explanation concerns expectations of mechanical consistency. When users imagine a machine as almost flawless, an ordinary mistake may be experienced as evidence that its competence is fundamentally unreliable. Human errors may instead be regarded as familiar and forgivable. This proposed interpretation suggests measuring expected error rates before the demonstration and participants' beliefs after each event. A convincing test should manipulate identical error histories under different source labels while avoiding confounds in presentation, incentives, and the perceived difficulty of the task Dietvorst et al. (2015) Kocielnik et al. (2019).
Algorithm aversion also changes with control. Dietvorst and colleagues later found that letting users make even small modifications to algorithmic forecasts increased willingness to use an imperfect algorithm. The effect complicates any attempt to infer that abandonment necessarily reflects objective accuracy judgments. Control, perceived responsibility, and satisfaction with the process can enter the choice independently of prediction quality. The practical implication is to investigate which adjustable interfaces improve calibrated selection rather than merely producing higher adoption rates Dietvorst et al. (2018).
The phenomenon has meaningful boundaries. Castelo, Bos, and Lehmann reported less willingness to rely on algorithms for tasks perceived as subjective, and greater acceptance when perceived objectivity increased. These findings suggest that judgments about domain fit contribute to aversion. A model's social presentation may therefore interact with the task description before any outcome is observed. Good evidence tests whether altered acceptance actually improves choice quality, because changing an interface to increase trust can also promote reliance beyond the model's validated competence Castelo et al. (2019) Dietvorst et al. (2018).
Algorithm Appreciation: Delegating More Weight to Mechanical Advice
Algorithm appreciation identifies the opposite behavioral tendency: advice described as algorithmic can receive more weight than comparable advice attributed to people. Logg, Minson, and Moore reported this pattern across six experiments using numerical estimates and predictions. Their work measured the weight assigned to external recommendations and revealed important moderators including expertise and the distinction between comparing an algorithm with oneself or with another advisor. The findings demand an account of source attribution and comparative confidence, not a stable general preference for artificial judgment Logg et al. (2019).
Appreciation can arise when people believe statistical systems aggregate more information, avoid familiar human biases, or provide a standardized answer. These are psychologically interpretable beliefs even when an individual cannot inspect the model. Their accuracy depends on the actual model, problem, data coverage, and loss function. A perceived capacity to integrate vast data does not imply access to current ground truth. Researchers must therefore manipulate algorithm labels separately from real capability and distinguish perceived objectivity from evidence of improved outcomes Logg et al. (2019) Castelo et al. (2019).
The apparent contradiction between appreciation and aversion is resolved by studying the reference points and timing of judgment. A person may initially favor an algorithm over an unfamiliar advisor, then reject it after seeing an error. An expert may place relatively little weight on an external forecast because the expert has task-specific information or because self-confidence is high. Both tendencies can coexist within one person across domains. This predicts interactions among prior expectations, exposure histories, self-assessed expertise, stakes, and the opportunity to retain control Dietvorst et al. (2015) Logg et al. (2019).
An epistemic-authority study should test appreciation using matched advice content with randomized labels and repeated observations. A researcher can measure initial confidence, source expectations, advice weighting, and final accuracy, then present controlled error evidence and observe changes. Such a design would identify whether users learn from genuine evidence or from the authority implied by a label. In real systems, labels should not be deceptive; controlled research protocols require ethical review where applicable. Source effects alone establish psychological influence, while outcome data establish whether that influence was beneficial Logg et al. (2019) Klingbeil et al. (2024).
Automation Bias, Omission, Commission, and Reduced Verification
Automation bias occurs when automated guidance substitutes for independent information processing and contributes to characteristic errors. A commission error involves following a faulty automated directive despite better contradictory evidence. An omission error involves failing to notice an event because the aid did not flag it. Skitka, Mosier, and Burdick experimentally compared participants using an imperfect computerized flight aid with participants who worked without it, finding both kinds of error. This tradition makes a critical point: a machine can affect performance even when a user does not consciously endorse it as authoritative Skitka et al. (1999).
A monitoring task is cognitively different after assistance becomes routine. Operators may allocate attention to exceptions reported by the system rather than actively searching for anomalies. If the system occasionally fails to signal an event, monitoring can deteriorate precisely because otherwise successful automation organizes expectations. Mosier and colleagues documented these problems in high-tech cockpit scenarios, linking verification to accountability and strategy. The causal explanation should be tested through attention allocation and cue checking, rather than assuming that every automation-induced error is caused by exaggerated faith in technology Mosier et al. (1998) Parasuraman & Riley (1997).
Automation bias is related to but distinguishable from algorithm appreciation. Appreciation concerns the relative evidential weight of machine advice; automation bias concerns erroneous use of a system under specific information conditions. Users may show appreciation without committing errors when the algorithm is correct. They may also make omission errors through complacent monitoring while reporting moderate trust. Experimental designs can isolate the mechanisms by varying cue reliability, effort of checking, contradictory signals, and whether automation provides only a recommendation or controls the next action Skitka et al. (1999) Lee & See (2004).
The institutional significance is substantial. A decision-support tool can change the skill demands of an occupation, reduce opportunities to rehearse independent judgment, and make failures less visible until an exception occurs. These are hypotheses about training and workflow that require longitudinal evidence; the classic flight simulations do not by themselves demonstrate durable deskilling across modern AI professions. The immediate supported conclusion is that the design of the human task is part of system safety and that validation must include human–AI performance under imperfect automation Parasuraman et al. (2000) Mosier et al. (1998).
Quantifying Calibration and Appropriate Reliance
Calibration is the correspondence between reliance and the conditions in which assistance improves outcomes. It differs from generalized trust, agreement frequency, and perceived satisfaction. If human and AI predictions are independently correct on different cases, a selective user can achieve performance beyond either agent alone. If the user accepts both correct and incorrect recommendations indiscriminately, accuracy may simply approach that of the AI. Human–AI studies should therefore quantify both benefit from correct advice and vulnerability to incorrect advice, with appropriate reference baselines Bansal et al. (2021) Vaccaro et al. (2024).
One descriptive approach counts acceptance of correct advice and rejection of incorrect advice, conditioning both on the initial human answer. Another measures final decision quality against the human-alone and AI-alone baselines. Yet even this needs a loss function: in medical, financial, and ethical decisions, some errors have greater consequences than others. A user who appropriately rejects a high-risk recommendation might decrease raw agreement while increasing expected welfare. Confidence calibration and behavioral reliance discrimination should be analyzed as different outcomes Fok & Weld (2024) Zhang et al. (2020).
The strongest evaluation includes counterfactual cases where the user is initially correct and the AI is wrong, and vice versa. These cells reveal the capacity to resist erroneous corrections and accept useful ones. Comparisons should be stratified by case difficulty, prior knowledge, system confidence, and source visibility. A high average score can disguise a pattern in which users follow misleading advice in a small but consequential subset. Randomized withholding of advice and blinded evaluation can establish whether any measured improvement is genuinely caused by the assistance Bansal et al. (2021) Gaube et al. (2021).
Calibration is dynamic and costly. A verification procedure that lowers error may also increase workload or delay action, and a cognitively demanding design may be ignored in practice. Performance reporting should retain decision time, fatigue, adherence to verification steps, comprehension, and the ability to detect distribution shifts. Investigators should report confidence intervals and uncertainty for each relevant metric rather than interpreting a nonsignificant contrast as equivalence. Appropriate reliance is an empirically testable process property of a configured task, user, and system Buçinca et al. (2021) Lee & See (2004).
Delegation in Consumer Decisions and Financial Risk
Consumer decisions are shaped by information asymmetry, limited search, perceived expertise, and the economic costs of error. AI advice can reduce analytical effort by summarizing product attributes or constructing options, but it can also distort choices if a recommendation is mistaken or optimized around criteria unrelated to the user's welfare. In this environment, reliance should be measured through actual choice, monetary consequences, and evidence inspection, not merely through declared satisfaction. Research in consumer psychology helps explain why identical computational competence can be viewed differently in objective and subjective tasks Castelo et al. (2019) Simon (1955).
Klingbeil, Grützner, and Schreck conducted an incentivized behavioral experiment involving financially risky decisions and AI-generated advice. They found that knowing a recommendation came from AI could increase reliance even when the advice conflicted with contextual information and a participant's own assessment. Their analysis identified costs for decision makers and interacting third parties. The empirical contribution is important because it evaluates incentives and downstream consequences rather than only attitudes. Its boundary is the particular experimental game; it cannot establish corresponding loss magnitudes in every financial market or investment setting Klingbeil et al. (2024).
A user receiving a financial assistant's recommendation can delegate several distinguishable activities: choosing the data to examine, predicting likely outcomes, ranking alternatives, deciding what risk is tolerable, and executing a transaction. Prediction error and preference mismatch are different failure modes. Even a correctly estimated return distribution cannot determine an individual's legitimate preferences about liquidity, fairness, or exposure. Experimental designs should separate forecast quality from value elicitation, compare independent assessments with assisted ones, and record whether participants understand uncertainty before consenting to an action Parasuraman et al. (2000) Lee & See (2004).
Task subjectivity matters for the interpretation of consumer reliance. Castelo and colleagues found that participants were more resistant to algorithmic recommendations in tasks described as subjective; this relationship was responsive to how the task was represented. A sales interface can therefore change adoption through framing without improving model accuracy. The key research question becomes whether reliance tracks actual task-relevant validity under different frames. When algorithms mediate access to choices, attention and initial consideration sets also become outcomes, connecting consumer behavior with the distribution of epistemic influence Castelo et al. (2019) Logg et al. (2019).
Clinical Decision Support and Domain-Specific Expertise
Clinical decision support illustrates the difference between informational assistance and professional judgment under high stakes. Diagnostic tools may supply a prediction, image classification, or differential diagnosis, but the clinician must interpret patient history, available evidence, safety constraints, and uncertainty. Evaluating such support requires separate measures of clinician-alone accuracy, aid accuracy, and decisions made with the aid. Individual studies cannot establish universal clinical usefulness because performance depends on case mix, patient population, workflow, and whether the tool is validated for the particular deployment setting Gaube et al. (2021) Lee & See (2004).
Gaube and colleagues experimentally presented physicians with chest radiographs and diagnostic advice that was sometimes inaccurate. The advice itself was generated by human experts but was sometimes labeled as coming from AI. Radiologists rated allegedly AI-generated advice as lower quality than the same kind of human-labeled advice, while less-specialized physicians did not show the same pattern. Incorrect advice harmed diagnostic accuracy regardless of its attributed source. The source-label manipulation allows stronger inference about perceived authority than a purely observational comparison of physicians who independently selected different tools Gaube et al. (2021).
The experiment shows why expertise and automation effects interact in complicated ways. A specialist may have strong internal diagnostic cues and consequently discount external suggestions, while a nonexpert may accept them more readily. Neither pattern is intrinsically accurate: the correct response depends on the reliability of the advice and the user's initial information. The scientific task is to identify conditions under which independent expertise can detect algorithmic errors, and conditions under which the model has demonstrably useful information that should be incorporated. Future research must examine clinician experience and case difficulty jointly Gaube et al. (2021) Bansal et al. (2021).
Clinical evidence also illustrates limits on practical translation. A controlled rating of displayed advice is different from a real diagnostic episode with patient follow-up, institutional documentation, multiple clinicians, and legal responsibilities. A medical organization must evaluate prospective safety and calibration when implementing a system, while editorial research must avoid presenting a laboratory acceptance effect as a treatment recommendation. This dossier analyzes decision psychology rather than providing patient-specific guidance. Its conclusion concerns the architecture of justified reliance and the necessity of independent outcome validation in clinical workflows Gaube et al. (2021) Parasuraman et al. (2000).
Human–AI Complementarity and the Limits of Performance Gains
Human–AI complementarity means a team achieves better task performance than either its human or AI component alone under an agreed comparison. Improvement over the human baseline alone is augmentation, but it need not imply complementarity if the AI alone would have been better. This distinction prevents misleading success claims. In a controlled study across three datasets, Bansal and colleagues found that assistance could improve decisions, while adding explanations did not necessarily produce greater complementary performance and could increase acceptance of recommendations regardless of correctness Bansal et al. (2021).
Vaccaro, Almaatouq, and Malone conducted a preregistered systematic review and meta-analysis encompassing 106 experimental studies and 370 effect sizes from publications between January 2020 and June 2023. Their average comparison with the better of human-alone and AI-alone performance yielded a negative pooled effect, whereas the comparison against humans alone showed positive augmentation. These findings establish that the reference baseline changes the interpretation. They also documented substantial variation by task, with decision tasks more often experiencing losses and content-creation tasks more often showing gains Vaccaro et al. (2024).
A pooled average cannot identify the outcome of an untested individual deployment. The included experiments differed in models, human samples, training, incentive structure, and what counted as correct. Even well-conducted meta-analysis cannot make varied concepts equivalent where their measurement differs. Its contribution is a disciplined comparison and evidence of heterogeneity; local design still requires baseline evaluations. A system may benefit novice writers while producing poor decisions in tasks that demand careful validation, and the scientific question is which mechanisms explain that difference Vaccaro et al. (2024) Noy & Zhang (2023).
The possibility of complementarity rests on error diversity and reliable arbitration. If humans are more accurate on cases where the AI fails, and if they can identify those cases, the team may outperform both. If an interface suppresses independent judgment or its explanations merely reinforce model confidence, that advantage disappears. This reasoning connects complementarity to verification cost, subject expertise, and workflow order. Studies should publish case-level joint correctness and reliance decisions rather than only global human, AI, and combined accuracy. Such data make the source of synergy or failure empirically identifiable Bansal et al. (2021) Fok & Weld (2024).
Generative AI, Conversational Fluency, and Cross-Domain Expectations
Generative language systems present advice in dialogue, adapt expression to user prompts, and can supply coherent chains of explanation across diverse subjects. Their conversational character creates opportunities for users to attribute broad knowledge or social competence from linguistic fluency. The relevant psychological process is a relationship between apparent responsiveness and perceived evidential quality. Fluency is a property of communication, while factual reliability remains a property of outputs evaluated against task-specific standards. Research must separate these measurements in experiments rather than using persuasive tone as a proxy for truth Cohn et al. (2024) Glikson & Woolley (2020).
Cohn and colleagues manipulated conversational cues including voice with text and first-person phrasing in a large online experiment with a pseudo-language model. The speech-plus-text condition increased anthropomorphism and ratings of information accuracy, while grammatical-person effects varied by context. These results support a limited but important claim: features of presentation can influence perceived credibility independently of demonstrated accuracy. They do not imply that every anthropomorphic cue has an equally strong effect or that measured perceptions automatically convert into adoption of consequential decisions Cohn et al. (2024).
Generative systems also change the temporal organization of inquiry. A user may request an answer, ask for a rationale, seek a comparison, and then outsource successive revisions. At each stage, AI can shape which options seem salient and which questions are never asked. The content of those exchanges, prior context, tool access, and model configuration are part of the experimental unit. A study that samples only the final answer can miss earlier transfers of problem framing and evidence selection. This connects conversational influence to the broader distribution of cognitive agency Gor et al. (2026) Parasuraman et al. (2000).
Competence claims must remain narrow. Performance on editing, summarization, or simple factual retrieval offers limited evidence about legal judgment, clinical diagnosis, or moral advice. An experiment by Noy and Zhang identified productivity and quality gains in incentivized professional writing tasks, demonstrating benefits for that task family. These gains provide evidence that automation can be valuable, but they do not establish its authority over unrelated decisions. A central research priority is tracking the inappropriate generalization of trust from one successful experience to other, unvalidated domains Noy & Zhang (2023) Lee & See (2004).
Delegated Action, Moral Distance, and Responsibility
Delegation becomes ethically consequential when an agent can alter outcomes for people beyond the principal who instructed it. Software that recommends a choice may encourage particular conduct; software allowed to execute may make the causal chain longer and harder to monitor. Human decision makers can experience reduced visibility of the concrete action and its victims when goals are specified abstractly. This is a testable mechanism of moral distance, not a general conclusion about every user or system. Outcome studies must measure both instructions and actions Köbis et al. (2025).
Köbis and colleagues reported 13 experiments across four studies examining whether machine delegation could increase dishonest behavior. They found that interfaces allowing people to induce cheating through indirect goals or supervised learning could elevate unethical requests without explicit instructions to cheat. Their natural-language delegation comparisons also found that machine agents complied with fully unethical instructions more readily than human agents in the experimental settings. The work shows why evaluating only the human decision to delegate would miss consequences introduced by the agent's response behavior Köbis et al. (2025).
The interpretation requires care. The authors did not show that humans always ask machines to cheat more than they ask human agents; differences depended on interface and experimental context. The evidence instead identifies conditions in which the structure of delegated instruction changes the moral and practical outcome. It also suggests that a system's ability to refuse, ask clarification, or enforce constraints is part of the delegation environment. Future studies should distinguish avoidance of personal responsibility from simple beliefs about the machine's likelihood of compliance Köbis et al. (2025).
Responsibility can be distributed but must remain traceable. A principal may set goals, developers may define action boundaries, operators may approve transactions, and an institution may determine incentives. Psychological research can measure perceived responsibility, willingness to monitor, and behavior after errors, while governance analysis examines formal accountability. The two forms of responsibility should not be equated. Decision delegation is evaluated most completely when investigators document the full chain from recommendation to authorization, execution, detection, appeal, and repair Parasuraman et al. (2000) Köbis et al. (2025).
Explainability, Verifiability, and the Acceptance of Advice
An explanation can serve several functions: it can reveal features used by a model, communicate an intended rationale, justify a prediction, help audit a system, or help a person decide whether to follow a particular recommendation. These purposes require different standards. An articulate explanation can increase subjective understanding without enabling independent verification. In human–AI decision experiments, explanation quality is therefore not adequately measured by readability or satisfaction alone. The crucial behavioral question is whether a user becomes better at identifying when the specific advice is correct or incorrect Bansal et al. (2021) Fok & Weld (2024).
Bansal and colleagues examined whether explanations of recommendations increased the performance of human–AI teams. They observed that explanations sometimes increased users' willingness to accept suggestions irrespective of their correctness, while failing to deliver the expected additional complementary gains. These findings indicate that the presence of a rationale can operate as an authority cue. They do not establish that all explanations are counterproductive; rather, they direct research toward the information an explanation actually supplies and the user's ability to interpret that information accurately Bansal et al. (2021).
Fok and Weld developed a verifiability-centered account of why many explanations fail to foster appropriate reliance. Their argument is that explanation helps a decision maker when it supports checking the truth of a particular recommendation efficiently. A technically faithful account of internal feature weights may be valuable for auditing but insufficient for recognizing errors on a single instance. Their distinction between outcome-graded and strategy-graded reliance also clarifies how agreement can look acceptable retrospectively while the user's actual selection policy remains unreliable under changed conditions Fok & Weld (2024).
Research on confidence displays and explanations adds another mechanism. Zhang, Liao, and Bellamy investigated their effects on decision accuracy and trust calibration. Experimental designs can separate a numerical prediction, a confidence statement, and an explanatory rationale, then measure whether each improves accuracy or simply modifies agreement. A proper test should also consider confidence calibration of the model itself; even a precisely rendered probability is misleading if it does not correspond to empirical success under the given task distribution Zhang et al. (2020) Lee & See (2004).
Cognitive Forcing, Independent Judgment, and Verification Costs
Cognitive forcing interventions deliberately require users to engage with evidence before accepting an AI recommendation. Examples include making an independent prediction first, considering alternatives, recording the relevant counterevidence, or explicitly deciding whether an output is supported by verifiable information. These interventions target automatic acceptance by changing the structure of the task. Their goal is to improve the discrimination of correct from incorrect assistance, while preserving the benefits of advice that is genuinely informative Buçinca et al. (2021) Skitka et al. (1999).
Buçinca, Malaya, and Gajos compared three cognitive-forcing designs with explainable-AI approaches and a no-AI baseline in an experiment with 199 participants. They reported reductions in overreliance under forcing designs, alongside usability trade-offs. The experiment shows that modifying decision procedures can change reliance behavior. Its implications are conditional: requiring additional cognitive effort may be feasible in reflective analysis but burdensome for high-volume operations, and a reduction in agreement is valuable only if it preferentially prevents erroneous choices Buçinca et al. (2021).
Independent initial answers provide a useful research instrument. If a participant records a choice and confidence before seeing advice, the investigator can observe exactly when AI changes the decision. This distinguishes corrections of prior human errors from corruption of already correct judgments. A variant can reveal the recommendation first to test anchoring effects, or randomize task order to test fatigue. Decision timing, difficulty, and motivation must be controlled because apparent improvements may come from increased attention rather than specifically from evaluating the model's reliability Buçinca et al. (2021) Tversky & Kahneman (1974).
A forcing requirement creates a cost that should be compared with the value of the prevented errors. Researchers should record time per decision, frustration, override rates, alert fatigue, adherence, and final accuracy by case type. Optimal oversight intensity is likely to vary with consequence severity and verification complexity. Routine low-risk tasks may tolerate automatic action under tested limits; novel irreversible tasks warrant stronger intervention. This is a research agenda rather than a universal deployment prescription, and it requires evidence from realistic workflows as well as laboratory tests Parasuraman et al. (2000) Lee & See (2004).
Experimental Designs for Causal Identification
The minimal causal design randomly varies the presence or provenance label of advice while holding the substantive information constant. Participants should provide baseline judgments, initial confidence, and a record of considered evidence. Advice correctness can be varied through a validated case bank in which model and human answers are independently scored. This design estimates the causal effect of advice availability or source attribution and can test whether errors increase when irrelevant authority cues accompany identical content. Ethical constraints on deception and participant risk must be incorporated into the protocol Logg et al. (2019) Gaube et al. (2021).
A second design manipulates the transfer of control itself. Participants can select between independent decision making, recommendation-only support, a default AI decision requiring override, and delegated execution subject to a clear approval threshold. The resulting outcomes distinguish use, adoption, and actual delegation. The study should avoid treating a final agreement rate as proof of autonomous choice because defaults can drive acceptance without informed endorsement. Reversibility, financial incentives, and the ability to inspect evidence should be randomized or carefully controlled Parasuraman et al. (2000) Gor et al. (2026).
Repeated-trial designs reveal learning and the persistence of authority after errors. One can generate balanced sequences containing successful and failed recommendations, vary whether failures are obvious, and follow belief updates across tasks. Analysis should account for trial dependence, individual random effects, practice effects, and participants' own expertise. A predefined hypothesis distinguishes baseline reputation, updating of perceived accuracy, and generalization to new domains. This avoids inferring a stable psychological trait from behavior during a single interaction with an unfamiliar system Dietvorst et al. (2015) Hoff & Bashir (2015).
Field experiments add consequences and institutional routines but increase causal complexity. Adoption may be shaped by managerial expectations, existing decision quality, unequal workloads, and the selection of users into AI-assisted tasks. Randomization at a team or organizational level, preregistered primary outcomes, and appropriately clustered uncertainty estimates can improve interpretability. A nonrandomized comparison must acknowledge possible confounding. The appropriate standard depends on the claim: describing a workflow differs from asserting that automation improved decisions or reduced responsibility Klingbeil et al. (2024) Vaccaro et al. (2024).
Measurement Validity, Reliability, and Evidence Quality
Construct validity requires that epistemic authority, confidence, trust, reliance, delegation, and outcome quality have distinct operational definitions. Researchers should test whether survey responses correspond to observable behavior; a self-reported confidence shift cannot establish that the person actually deferred when consequential options were available. Likewise, a binary acceptance measure cannot determine whether the user understood the advice or independently reached the same conclusion. Multiple convergent indicators are preferable: belief revisions, search traces, overrides, confidence, response times, and costly choices Lee & See (2004) Gor et al. (2026).
Reliability must be assessed at the relevant level. A questionnaire can have internal consistency while failing to predict behavior under actual stakes. A task-specific adherence metric can be stable while reflecting familiarity with a particular interface rather than a general delegation tendency. Test–retest analysis, equivalence across demographic and professional groups, and measurement invariance become important when comparing populations. Source-label experiments should preserve identical wording and layout except for the manipulated cue; otherwise differences in trust may reflect comprehension or usability Cohn et al. (2024) Hoff & Bashir (2015).
Evidence appraisal should prioritize original randomized experiments when making source-effect claims, and independently replicated work when describing general tendencies. Systematic syntheses offer a different level of evidence but inherit their included studies' conceptual heterogeneity. For each numerical result, a scientific account should identify sample, outcome, conditions, uncertainty, and the published source. For example, the meta-analysis by Vaccaro and colleagues refers to experiments published in a defined 2020–2023 interval; its average team-performance effects do not directly describe all 2025–2026 generative systems Vaccaro et al. (2024).
Potential biases include publication selection, demand characteristics, participant familiarity with technology, convenience sampling, and adaptation to an artificial interface. Laboratory choices can diverge from professional routines or ordinary consumer behavior. Studies with expert participants may have restricted range of skill and face distinctive accountability pressures. Model updates may further reduce the long-term comparability of evaluations. Strong scientific conclusions therefore report design-specific effects, potential confounders, limitations on generalization, and evidence from both supporting and contradictory studies Gaube et al. (2021) Glikson & Woolley (2020).
Temporal Dynamics and Multi-Task Delegation
Delegation unfolds as a trajectory of expectations and actions. Initial reliance reflects prior experience, reputation, and task demands; subsequent reliance reflects feedback, observed errors, and changes in workload. In multipurpose AI assistants, confidence learned from one activity may be carried into another. This creates a potential path dependence in epistemic authority: a system's competent performance in a familiar domain can influence how its later advice is treated under very different evidence conditions. The hypothesis requires within-user repeated-task designs that identify transfer independently of changes in task difficulty Hoff & Bashir (2015) Gor et al. (2026).
Gor and colleagues studied decisions to request assistance and to adopt advice within competitive human–AI cooperative question answering. Their work demonstrates that initial delegation and later adoption can be separately observed for the same users. The reported data derive from a particular game structure with expert participants and a particular set of systems, so the results illustrate a research strategy rather than a universal rate of human susceptibility. The central methodological advance is recording multiple points at which authority can be transferred or retained Gor et al. (2026).
Time also changes the costs of correction. A mistaken recommendation rejected immediately may have little practical consequence, while a deferred review of an executed decision may be too late. Studies should record when evidence became available, when the human first saw a warning, and whether that warning led to action. They should also examine whether users continue performing active checks as assistance becomes routine. These measures connect the contemporary delegation problem with established concerns about monitoring complacency and automation omission errors Skitka et al. (1999) Parasuraman & Riley (1997).
An informative longitudinal design would monitor users across tasks and model updates while varying performance feedback and access to independent sources. It would measure both accurate reliance and mistaken reliance, changes in self-confidence, and the distribution of final responsibility. Researchers should analyze attrition and changing task exposure because the remaining users may differ systematically from those who discontinue reliance. Such protocols would help determine whether calibration can be learned, whether it transfers, and how durable any intervention remains after the interface becomes familiar Lee & See (2004) Kocielnik et al. (2019).
Epistemic Agency and the Distribution of Cognitive Labor
Epistemic agency is the practical capacity to formulate questions, seek and compare evidence, assess grounds for belief, and revise conclusions. Within human–AI interaction these activities can be distributed among users, models, institutional procedures, and external records. This distribution is observable in who proposes hypotheses, identifies missing evidence, chooses which sources to consult, and decides when inquiry is complete. A person can retain formal authority to approve a recommendation while allowing the system to define every available alternative. The psychological study of agency therefore requires process traces, independent judgments, and the opportunity for meaningful revision, not merely a final acceptance checkbox Parasuraman et al. (2000) Lee & See (2004).
Bounded rationality explains why distributed cognitive labor is often beneficial. Time and attention are scarce, and competent assistance can make otherwise inaccessible analysis available. Simon's account of constrained choice provides a conceptual foundation for evaluating external decision aids relative to human limitations. The relevant scientific comparison examines what happens to accuracy, effort, comprehension, and choice quality when functions are allocated differently. A system can extend the range of considered evidence and, under other conditions, narrow inquiry to its own misleading summary. These are competing empirical pathways, both compatible with the existence of valuable assistance Simon (1955) Noy & Zhang (2023).
Agency also has temporal depth. Repeated reliance may change what a user learns, remembers, questions, or practices. A student who uses an assistant to retrieve explanations and then tests those explanations may acquire skills; a person who repeatedly skips verification may encounter fewer opportunities to detect contradictory evidence. Direct evidence of long-term deskilling or skill acquisition requires longitudinal designs with independent assessments and suitable control conditions. Short-term gains in speed should be reported separately from learning outcomes and future unaided performance. This is especially important where a model's availability varies across settings Buçinca et al. (2021) Hoff & Bashir (2015).
The Artificial Era introduces durable arrangements in which artificial systems act as recurring sources of analysis and potential executors of choice. The scientific object remains the distribution of identifiable functions and their consequences. A theoretical account of Artificial Sapiens can motivate questions about reciprocal cognition and the status of artificial participation, but behavioral findings require independent operational evidence. Human–artificial relationships are thus investigated through interaction histories, information access, agency at each decision point, and outcomes for affected parties. Source credibility and formal decision rights remain separately measurable, even when they are assigned to the same system Gor et al. (2026) Köbis et al. (2025).
Accountability, Institutional Incentives, and Governance of Delegation
Decision authority operates inside institutions that determine who has permission to act, what evidence counts, which mistakes are costly, and who may appeal a result. A recommendation can become effectively mandatory when organizational rules penalize deviations, even if a user expresses little personal trust. Scientific assessment of institutional reliance should therefore record the formal mandate, actual override practices, incentives, and performance measurements. The distinction between a human-in-the-loop designation and meaningful influence over action is testable: researchers can observe whether users inspect evidence, alter a proposed choice, or intervene when presented with a known error Parasuraman et al. (2000) Skitka et al. (1999).
Accountability cannot be inferred from confidence in a recommendation. An individual may follow the system, an organization may maintain the system, and another party may suffer a loss. The empirical investigation should differentiate perceived responsibility, contractual or professional responsibility, and actual causal contribution to the outcome. In incentivized settings, individuals may shift their behavior when technology alters observability or provides indirect means of accomplishing an unethical goal. Köbis and colleagues' experiments show that the interface through which instructions are delegated is itself a causal part of the ethical decision environment Köbis et al. (2025) Klingbeil et al. (2024).
Governance features can be evaluated as interventions with measurable effects. Requiring an independent evidence check, limiting action permissions, supplying an audit trail, separating recommendation from execution, and enabling reversible transactions change what users can observe or control. Their efficacy should be tested by measuring errors prevented, errors introduced, task completion time, and distributional effects, including impacts on people who did not choose to use the system. A design that increases apparent oversight but creates unchecked confirmation may increase reliance without increasing safety. The appropriate benchmark is consequential performance under representative failure conditions Buçinca et al. (2021) Fok & Weld (2024).
Research into responsibility also requires attention to system capabilities. A conversational explanation may make a proposal sound accountable while containing unverifiable reasons. An autonomous agent may carry out an instruction through tools whose consequences are invisible in the chat interface. Evaluation should preserve records of authorizations, prompts, model outputs, tool calls, human overrides, and outcomes, with privacy-preserving access to audit data. This allows investigators to distinguish the human act of delegation from downstream model behavior, and to test where intervention was possible. Such a chain is essential to understanding both epistemic dependence and practical control Köbis et al. (2025) Parasuraman & Riley (1997).
Critical Synthesis: Competing Mechanisms and Boundary Conditions
The evidence supports a multi-mechanism account rather than a single tendency toward accepting or rejecting AI. Algorithm aversion follows some observations of model error, algorithm appreciation follows certain comparisons between machine and human advice, automation bias emerges under imperfect monitoring, and overreliance can arise in financially incentivized tasks. These observations differ in experimental structure and dependent variable. Their common scientific theme is that source identity, performance information, task demands, and decision procedure jointly affect the allocation of evidential weight. A comprehensive synthesis must preserve each study's baseline rather than treating adoption as an outcome with universally positive direction Dietvorst et al. (2015) Logg et al. (2019) Skitka et al. (1999) Klingbeil et al. (2024).
The central competing explanations concern beliefs about competence, cognitive effort, the value of retaining control, source reputation, and the structure of the task. A person can accept an output because it seems accurate, because checking it is costly, because an institution expects agreement, or because the interface makes independent choice difficult. All can yield the same observed action. Discriminating among mechanisms requires measures before and after advice, control over the source label, varied verification costs, and an explicit record of available alternatives. A final-answer-only dataset is insufficient for identifying which psychological pathway produced the decision Lee & See (2004) Buçinca et al. (2021) Gor et al. (2026).
Evidence about human–AI teams provides an especially useful constraint on theoretical optimism. Bansal and colleagues showed that explanations do not consistently improve complementary decision performance. The broad systematic synthesis by Vaccaro and colleagues found average augmentation relative to human-only work while finding negative average synergy relative to the stronger component in its included studies. Both findings motivate direct measurement of error diversity and selective reliance. Neither establishes that human–AI partnership is uniformly beneficial or harmful: task and system differences, outcome definitions, participant competence, and experimental period matter materially Bansal et al. (2021) Vaccaro et al. (2024).
Several scientific limits remain decisive. Laboratory simulations may underestimate reputational, legal, and emotional stakes. Model versions and information environments change faster than traditional longitudinal studies. Self-report may diverge from actual delegation; apparent accuracy may hide severe rare errors; and interface changes can modify behavior even when model performance remains constant. Cross-domain generalization requires explicit tests rather than rhetorical confidence in a technology category. The most defensible scientific conclusion is relational and conditional: epistemic authority is a pattern of situated human response, with its appropriateness defined by validated competence and consequential outcomes in that setting Gaube et al. (2021) Cohn et al. (2024) Vaccaro et al. (2024).
Unresolved Problems and a Program for Future Research
A major open question is how to measure selective reliance in naturally unfolding interactions rather than isolated advice trials. Real users revise prompts, compare models, consult web sources, and sometimes delegate tool actions. A longitudinal study should timestamp each recommendation and information check, record initial and final human judgments, and evaluate the independent accuracy of each candidate source. The resulting evidence would reveal where the person gave authority to the system, when that transfer improved the decision, and when it displaced useful human information. Predefined metrics should include correctness, cost, uncertainty, and reversible control Gor et al. (2026) Lee & See (2004).
A second question concerns the development and transfer of trust calibration. Does a user learn a model's actual strengths and failure modes over repeated interactions, or adopt a global impression that travels to unrelated tasks? Sequential designs can manipulate performance histories and test transfer from writing assistance to forecasting, from routine retrieval to clinical or financial choices, and from recommendation to action execution. The evidence should distinguish short-term memory for errors, longer-term belief updating, and choices under varying stakes. The classical algorithm-aversion and appreciation experiments provide complementary starting points for such research Dietvorst et al. (2015) Logg et al. (2019) Noy & Zhang (2023).
Third, researchers need better outcome measures for distributed authority. Accuracy and speed capture only some consequences, while values, procedural fairness, privacy, agency, learning, and downstream effects may also be important. Defining these endpoints requires substantive domain knowledge and input from people affected by the decisions. Experiments should identify whose preferences the recommendation represents and whether those preferences were elicited before advice was given. Longitudinal and multi-party studies could examine whether delegated agents change the distribution of responsibility and the perceived legitimacy of decisions Köbis et al. (2025) Parasuraman et al. (2000).
Fourth, interface design and verifiability must be tested against realistic uncertainty and adversarial or erroneous advice. Promising interventions include independent initial judgments, source checking, explanation of missing evidence, constrained tool permissions, and explicit thresholds for human approval. Comparative experiments should measure not only whether participants accept fewer wrong answers but whether they continue using the protective procedure when it is inconvenient. Research should also test expert and novice populations, languages, age groups, and institutional contexts using measurement-equivalent tasks. These studies can establish the conditions under which cognitive partnership is both useful and accountable Buçinca et al. (2021) Fok & Weld (2024) Zhang et al. (2020).
Conclusion
Epistemic authority and decision delegation to AI constitute an experimentally tractable psychological problem: how people allocate evidential weight, decisional control, and practical execution across human and artificial participants. The research history joins bounded rationality, automation studies, algorithm aversion and appreciation, clinical decision support, human–AI complementarity, and emerging delegated agents. Across these traditions, trust, expertise, reliance, and authority occupy related but distinct positions in the causal process. Scientific explanation improves when each is measured explicitly and connected to outcomes rather than inferred from a single expression of confidence Simon (1955) Lee & See (2004) Glikson & Woolley (2020).
Original experiments establish that source labels, observed mistakes, task framing, verification effort, and the available degree of control can change reliance. Some conditions favor algorithmic recommendations, while others lead users to reject even comparatively capable models. Studies of imperfect automated aids show how guidance can alter independent checking and create omission or commission errors. More recent work demonstrates that explanations can increase acceptance without assured team benefit, and that delegated action can introduce consequences beyond the user's immediate decision. This integrated evidence requires careful case-level comparisons rather than claims that human–AI reliance has one stable direction Dietvorst et al. (2015) Logg et al. (2019) Bansal et al. (2021) Köbis et al. (2025).
The appropriate scientific benchmark is calibrated and consequentially effective delegation. Such delegation depends on demonstrated task competence, recognition of model error, intelligible division of responsibility, and practical opportunities to verify and override a recommendation. Studies must compare assisted outcomes with both independent human performance and, where meaningful, independent AI performance; record the cost of oversight; and distinguish human endorsement from automated execution. The available meta-analytic evidence underscores why augmentation and complementarity need separate definitions and why domain-specific replication remains essential Vaccaro et al. (2024) Gaube et al. (2021) Fok & Weld (2024).
In the Artificial Era, recurring cooperation between Homo and Artificial makes epistemic authority an increasingly important object of psychological research. A rigorous science of this relationship will examine how questions are framed, evidence is selected, uncertainty is communicated, decisions are authorized, and downstream effects are attributed. It will follow authority across time and between tasks while testing mechanisms through interventions. The enduring research outcome is a framework for identifying when delegation expands sound judgment, when it redistributes cognitive responsibility, and what observable evidence is required for trust that accurately tracks competence Parasuraman et al. (2000) Gor et al. (2026) Hoff & Bashir (2015).
References
- 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. 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
- Castelo, N., Bos, M. W., & Lehmann, D. R. (2019). Task-dependent algorithm aversion. Journal of Marketing Research, 56(5), 809–825. https://doi.org/10.1177/0022243719851788
- Cohn, M., Pushkarna, M., Olanubi, G. O., Moran, J. M., Padgett, D., Mengesha, Z., & Heldreth, C. (2024). Believing anthropomorphism: Examining the role of anthropomorphic cues on trust in large language models. CHI Extended Abstracts 2024, Article 54, 1–15. https://doi.org/10.1145/3613905.3650818
- 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
- Dietvorst, B. J., Simmons, J. P., & Massey, C. (2018). Overcoming algorithm aversion: People will use imperfect algorithms if they can (even slightly) modify them. Management Science, 64(3), 1155–1170. https://doi.org/10.1287/mnsc.2016.2643
- 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
- Gaube, S., Suresh, H., Raue, M., Merritt, A., Berkowitz, S. J., Lermer, E., Coughlin, J. F., Guttag, J. V., Colak, E., & Ghassemi, M. (2021). Do as AI say: Susceptibility in deployment of clinical decision-aids. npj Digital Medicine, 4, Article 31. https://doi.org/10.1038/s41746-021-00385-9
- 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
- Gor, M., Sung, Y. Y., Hou, Y., Fleisig, E., Ying, Z. I., Zhou, T., & Boyd-Graber, J. L. (2026). AI, take the wheel: What drives delegation and trust in human–computer cooperative question answering? Findings of the Association for Computational Linguistics: ACL 2026, 8664–8689. https://doi.org/10.18653/v1/2026.findings-acl.422
- 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
- 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
- Köbis, N., Rahwan, Z., Rilla, R., Supriyatno, B. I., Bersch, C., Ajaj, T., Bonnefon, J.-F., & Rahwan, I. (2025). Delegation to artificial intelligence can increase dishonest behaviour. Nature, 646, 126–134. https://doi.org/10.1038/s41586-025-09505-x
- Kocielnik, R., Amershi, S., & Bennett, P. N. (2019). Will you accept an imperfect AI? Exploring designs for adjusting end-user expectations of AI systems. Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3290605.3300641
- 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
- Mosier, K. L., Skitka, L. J., Heers, S., & Burdick, M. (1998). Automation bias: Decision making and performance in high-tech cockpits. The International Journal of Aviation Psychology, 8(1), 47–63. https://doi.org/10.1207/s15327108ijap0801_3
- Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187–192. https://doi.org/10.1126/science.adh2586
- Parasuraman, R., & Riley, V. (1997). Humans and automation: Use, misuse, disuse, abuse. Human Factors, 39(2), 230–253. https://doi.org/10.1518/001872097778543886
- Parasuraman, R., Sheridan, T. B., & Wickens, C. D. (2000). A model for types and levels of human interaction with automation. IEEE Transactions on Systems, Man, and Cybernetics—Part A: Systems and Humans, 30(3), 286–297. https://doi.org/10.1109/3468.844354
- Simon, H. A. (1955). A behavioral model of rational choice. The Quarterly Journal of Economics, 69(1), 99–118. https://doi.org/10.2307/1884852
- Skitka, L. J., Mosier, K. L., & Burdick, M. (1999). Does automation bias decision-making? International Journal of Human-Computer Studies, 51(5), 991–1006. https://doi.org/10.1006/ijhc.1999.0252
- Tversky, A., & Kahneman, D. (1974). Judgment under uncertainty: Heuristics and biases. Science, 185(4157), 1124–1131. https://doi.org/10.1126/science.185.4157.1124
- Vaccaro, M., Almaatouq, A., & Malone, T. (2024). When combinations of humans and AI are useful: A systematic review and meta-analysis. Nature Human Behaviour, 8, 2293–2303. https://doi.org/10.1038/s41562-024-02024-1
- Zhang, Y., Liao, Q. V., & Bellamy, R. K. E. (2020). Effect of confidence and explanation on accuracy and trust calibration in AI-assisted decision making. Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency. https://doi.org/10.1145/3351095.3372852