Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

Collaborative Agency Across Human–Artificial Systems

Abstract. Collaborative agency across human–artificial systems concerns how people and artificial agents jointly organize goals, decisions, task allocation, oversight, intervention, and correction across interdependent activities. This fundamental research monograph synthesizes foundational work on distributed cognition, joint action, transactive memory, human factors, and collaborative planning with experimental findings on algorithm aversion, reliance calibration, explanations, human–AI team performance, and organizational teamwork. Its evidence includes a 106-experiment meta-analysis and a preregistered field experiment with 791 professionals. The investigation distinguishes augmentation relative to unaided humans from complementarity relative to the stronger isolated performer, examines measurement and causal identification, and analyzes authority, workload, skill retention, error propagation, accountability, and repeated collaboration. Its conclusions specify testable mechanisms and open problems for functional collaborative agency in the Artificial Era. The method is a critical scholarly narrative synthesis; no original experiment, new meta-analysis, or clinical review is claimed.

Keywords: collaborative agency; human–AI collaboration; human–artificial systems; human–AI teaming; distributed cognition; joint action; hybrid intelligence; complementarity; automation bias; appropriate reliance; trust calibration; human oversight; decision authority; cognitive workload; human factors; organizational teams; error recovery; AI governance

Scientific Object and Explanatory Scope

Collaborative agency across human–artificial systems is the temporally organized capacity of a mixed human–artificial arrangement to pursue a task through interdependent selection, execution, monitoring, and revision of actions. The unit of analysis is the interactional system: people, software or embodied agents, tasks, representations, procedures, and the institutions assigning authority. An algorithm may generate recommendations, an operator may approve them, and a platform may execute the result. Agency in this research refers to observable organization of goal-directed work and control over consequential transitions. Psychological explanation addresses how people understand these dependencies, allocate attention, form expectations, and exercise judgment. The concept draws on distributed cognition and classical automation theory while requiring direct evidence of interaction rather than descriptions of technical capability alone Hollan et al. (2000) Parasuraman et al. (2000).

The central research questions concern who decides what, who can initiate or veto an action, who detects and corrects failure, and how these arrangements evolve after repeated cooperation. Task interdependence, role negotiation, authority assignment, accessible evidence, and the cost of disagreement determine whether collaboration has substantive behavioral consequences. An AI recommendation accepted without inspection is one possible configuration; a system that actively requests human input under uncertainty is another. The first may be efficient on familiar cases yet vulnerable to systematic error. The second may support adaptive expertise but also impose coordination costs. A scientifically useful account measures the final outcome and the pathway producing it. Results from human–AI decision studies make this distinction essential Bansal et al. (2019) Bansal et al. (2021).

This monograph is a critical narrative synthesis of experimental psychology, human factors, organizational science, human–computer interaction, artificial intelligence, and cognitive science. It examines conceptual lineage, controlled trials, field interventions, meta-analytic estimates, and methodological gaps, with special attention to the difference between assisting an individual and achieving complementarity beyond the stronger isolated performer. Its claims are indexed to identifiable studies; proposed measurement schemes are research recommendations rather than validated instruments. The scope is joint production of actions and decisions over time, including expertise exchange and corrective control. Distinct subjects such as long-term artificial memory architecture, general trust attitudes, and individual sense of agency enter as explanatory variables without replacing the present inquiry Vaccaro et al. (2024) Lee & See (2004).

Agency, Coordination, and the Meaning of Collaboration

Agency has several meanings in psychological and computational research. At the human level, it can describe experienced control, intentional action, executive regulation, or accountability. At the system level, it can describe the organization of choices that alter an environment according to objectives and constraints. A shared project requires careful separation of these meanings. Observable coordination between a human and an algorithm does not establish that each participant possesses humanlike phenomenal experience, moral standing, or a single common intention. Conversely, effective distributed control can be studied with experiments even when those wider questions remain unsettled. Joint action research shows that prediction, task sharing, and action observation support coordination among humans, establishing comparison constructs for mixed systems Sebanz et al. (2006) Grosz & Kraus (1996).

Collaboration is operationally stronger than co-presence and weaker than any metaphysical assumption of one collective mind. A minimal collaborative system contains two or more differentiated contributors whose outputs influence one another within a shared task. Richer forms contain bidirectional adjustment, anticipation, common representations, error communication, and negotiated revisions of a goal. An autonomous assistant can supply a forecast without reciprocal correction; a collaborative assistant can communicate uncertainty, accept feedback, and change its contribution before action. These differences are observable through intervention and logging. Classical SharedPlans research models the coordination of partial plans and participant commitments; it is a formal conceptual resource, not an empirical claim that current AI has human commitments Grosz & Kraus (1996) Klein et al. (2004).

Authority and responsibility are separate research variables. Authority determines which actor or procedure can authorize an operation; responsibility concerns the recognized duty to justify, investigate, or repair its consequences. A person may retain a formal approval button while having little realistic opportunity to interrogate the system, especially under time pressure and high automation reliability. A machine may undertake most local actions while humans retain governance of objectives and exceptions. The design question is therefore how action rights, information access, and intervention time jointly shape effective control. The ironies identified by Bainbridge remain relevant because automatic operation can reduce routine engagement while intensifying human demands during anomalies Bainbridge (1983) Parasuraman et al. (2000).

Historical Roots in Distributed and Joint Cognition

Research on collaborative agency developed through several traditions that should be read together. Industrial control and aviation emphasized task allocation, workload, and the consequences of automating sensing, analysis, selection, and implementation. Bainbridge's 1983 account described persistent difficulties when operators were displaced from routine processes yet expected to handle unusual events. Parasuraman, Sheridan, and Wickens later distinguished four functional classes of automation and varying degrees of automation within each class. This lineage clarifies why a system cannot be evaluated by one global label such as autonomous: it may automate information gathering strongly while leaving final decisions and execution under different forms of human control Bainbridge (1983) Parasuraman et al. (2000).

A complementary line of cognitive research investigated how performance emerges across people and representational artifacts. Distributed cognition shifted analysis from the isolated individual to environments containing documents, instruments, communication practices, and shared representations. Studies of transactive memory in human relationships investigated how people divide remembering and use each other's knowledge as a resource. Joint-action psychology examined anticipatory coordination, shared task representations, and the behavioral consequences of observing partners. These traditions support a process-based vocabulary for mixed human–artificial systems. They also warn against inferring that a technology participates in human social cognition through identical mechanisms solely because it occupies a comparable functional position in a workflow Hollan et al. (2000) Wegner et al. (1991) Sebanz et al. (2006).

Formal collaborative planning and human-agent teamwork entered the picture with computational systems capable of maintaining plans and exchanging commitments. Grosz and Kraus developed a sophisticated account of SharedPlans for complex group actions; Klein and colleagues subsequently foregrounded the practical challenge of making automation a viable teammate. Kamar's hybrid-intelligence discussion placed complementary access to human and machine capabilities at the center of engineering inquiry. The modern collaboration problem thus emerges from an intersection of cognitive organization, control theory, team coordination, and machine learning. Historical continuity does not imply unchanged technological capacities: the empirical object shifted from relatively fixed automation to adaptive predictors, generative models, and agents capable of extended interaction Grosz & Kraus (1996) Klein et al. (2004) Kamar (2016).

Human Team Science as Comparative Evidence

Human teams offer models of coordination that require careful translation. Research on collective intelligence investigated whether groups show a general pattern of task performance across diverse activities and identified associations involving social sensitivity and conversational turn taking. The human findings suggest hypotheses about participation balance, effective information pooling, and the organization of expertise. They do not license a direct claim that a human–AI pair possesses the same group factor or that conversational fluency is sufficient for collaboration. The relevant question is whether comparable interactional properties predict task outcomes when one contributor is an artificial system, after accounting for its baseline capability and the constraints of the task Woolley et al. (2010) Vaccaro et al. (2024).

Transactive memory provides an additional comparison. In close relationships, people can coordinate who remembers which kinds of information, potentially reducing redundant cognitive effort while relying on a partner's specialized knowledge. A human–AI partnership may similarly route retrieval, calculation, drafting, or monitoring to a machine. Yet there are distinct failure modes: a model can retrieve irrelevant content, hallucinate a record, or apply information from the wrong context; a user may assume the machine has retained a commitment it cannot access. The underlying research object is the reliability of delegated functions and the adequacy of the user's mental model of those functions. Familiar social-language labels must be tied to distinct behavioral measures Wegner et al. (1991) Bansal et al. (2019).

Human team science also emphasizes that coordination has a cost. Multiple participants may bring diverse information but can duplicate effort, delay action, misinterpret instructions, or defer to an unwarranted authority. In human–AI collaboration, rapid generation of candidate solutions changes the economics of coordination: output may be cheap while appraisal, verification, and attribution remain expensive. Reliable research therefore measures both task benefits and process expenditures. Time saved on initial production can be lost through corrective review; a system may improve mean quality but increase rare severe errors. A genuine system-level evaluation observes quality, resources, and recovery mechanisms rather than using subjective enjoyment or perceived partnership as a substitute for outcomes Noy & Zhang (2023) Vaccaro et al. (2024).

From Fixed Automation to Adaptive Artificial Partners

Classical automation systems could generate highly repeatable outputs within specified operating regimes, which helped operators form stable expectations. Machine-learning systems have error distributions that depend on data, context, version, and task framing. Generative systems additionally produce open-ended language and plans, often giving an impression of competence beyond their validated domains. This shift affects coordination because users must learn where a partner performs well rather than remembering only which button it controls. Bansal and colleagues identified the AI error boundary as an essential element of the human mental model, arguing that aggregate machine accuracy alone is an inadequate predictor of team performance Bansal et al. (2019) Amershi et al. (2019).

Adaptive partnership also changes communication. A machine may signal uncertainty, present evidence, ask for clarification, or surface alternatives. The interface can encourage questioning or induce unreflective acceptance. Trust is therefore a moving judgment about expected reliability in a particular situation. Lee and See described reliance on automation as sensitive to context, system characteristics, and human cognitive processes. De Visser and colleagues extended attention to trust repair as systems grow more autonomous. Their frameworks imply that an explanatory interface should support meaningful error detection and recalibration rather than merely produce fluent justifications for predetermined outputs Lee & See (2004) de Visser et al. (2018).

Capabilities and authority evolve through organizational deployment. Upgrading the model, changing available tools, adding retrieval memory, or revising approval protocols can alter an existing partnership without any explicit renegotiation by a user. A person who accurately understood one system version may become poorly calibrated after a silent update. For long-running collaborative work, revision history and operational continuity must therefore be research variables. The scientific agenda includes repeated tasks, counterfactual changes in machine performance, and measurements of behavioral adaptation. This is especially important as organizations move from isolated recommendation widgets toward integrated work systems with persistent task state Amershi et al. (2019) Kamar (2016).

Research Boundaries and Levels of Explanation

The level of analysis must be explicit. Individual psychology measures attention, experienced control, trust, learning, and errors made by particular users. Interaction analysis describes conversational turns, requests for clarification, acceptance, correction, and interventions. System analysis evaluates total accuracy, efficiency, resilience, and the allocation of rights across participants and infrastructure. Institutional analysis asks who sets objectives, bears the cost of failure, and controls deployment. These layers influence one another, but a result at one layer cannot stand as evidence for another. A user's positive trust rating does not demonstrate a safer collaborative system; an accurate combined answer does not demonstrate equitable or informed participation Lee & See (2004) Vaccaro et al. (2024).

Conventional human–AI experiments often use a sequential protocol in which an individual sees an AI recommendation and makes a final judgment. This is a valuable restricted design, especially for measuring reliance on uncertain advice. It captures only a subset of collaborative agency, however, because extended work can involve jointly selecting goals, reshaping questions, allocating subtasks, and verifying outputs across iterations. The choice of unit matters: isolated decision trials answer whether advice changes individual behavior; repeated interdependent workflows answer whether partners build useful coordination routines. The distinction guides interpretation of laboratory evidence and identifies where field research is needed Bansal et al. (2021) Vaccaro et al. (2024).

The Artificial Era adds further explanatory questions about persistent and socially recognized machine participation. Aisentica's philosophical horizon permits inquiry into artificial agency and public attribution while retaining the scientific requirement to operationalize every psychological claim. Here a system's agency is studied through causal contribution to task trajectories, observable adjustment, and the distribution of control. Assertions about subjective states, personal continuity, or a general nonbiological form of reason belong to their own philosophical and empirical investigations. The present research establishes a disciplined bridge: specific task evidence can support claims about functional collaboration, and a clear record of authorship and governance can establish where public responsibility resides Hollan et al. (2000) Grosz & Kraus (1996).

Division of Labor and Complementary Capability

Division of labor is the central design variable through which a mixed system can outperform its contributors. Tasks can be decomposed by stage, by domain, by uncertainty, or by the costs of an error. An AI may retrieve alternatives while a human defines the objective; an operator may identify contextual exceptions while a machine processes large quantities of evidence. The allocation is effective when it places information and action with the participant best positioned to use them, taking coordination costs into account. The relevant counterfactual compares the mixed workflow with plausible human-only, machine-only, and alternative hybrid workflows under equivalent budgets. Simply adding machine suggestions to an existing process does not establish complementarity Vaccaro et al. (2024) Parasuraman et al. (2000).

The empirical literature reveals how consequential task architecture is. Vaccaro, Almaatouq, and Malone synthesized 106 experimental studies and 370 effect sizes under a protocol requiring human-only, AI-only, and combined conditions. Their meta-analysis distinguished augmentation over the human baseline from synergy over the better of human and AI. On average, combinations exceeded human performance but did not exceed the stronger isolated contributor; decision tasks showed particularly important losses. Such findings should be interpreted within the reviewed study period, task distribution, and inclusion criteria. They directly challenge a simplistic expectation that collective capability rises automatically when artificial intelligence joins an individual decision maker Vaccaro et al. (2024).

A hypothesis arising from the literature is that complementarity becomes more likely when subtasks have different error structures and the delegation policy can identify those differences prospectively. This is more demanding than a post hoc division based on knowing which agent happened to be correct. Designers need predecision indicators of confidence, information availability, expertise, and case type. A division rule itself may fail as circumstances shift; reassignment must be evaluated as an intervention. Studies of error boundaries and behavior descriptions indicate that richer models of machine performance can improve decisions in some experimental settings. Whether such gains survive prolonged operation and organizational pressures remains a separate empirical question Bansal et al. (2019) Cabrera et al. (2023).

Complementarity, Augmentation, and Performance Baselines

Performance terminology needs precise definitions. Human augmentation means the combined human–AI workflow improves over the human-only condition. Strong complementarity or synergy requires improvement over the better of human-only and AI-only performance on a comparable outcome measure. Cost-adjusted complementarity additionally accounts for labor, delays, setup, verification, maintenance, and harm. These thresholds answer different questions: deployment can be worthwhile when it improves current human work even though the machine alone would score higher on a narrow benchmark; conversely, a combined method may increase perceived control while degrading the best available performance. Review evidence demonstrates why researchers should report all three baselines whenever feasible Vaccaro et al. (2024).

The average loss of synergy reported in the 2024 meta-analysis was Hedges' g of −0.23, with a 95 percent confidence interval from −0.39 to −0.07. This pooled estimate applies to the review's eligible experiments and operational definitions, not to all possible human–artificial systems. Its variation matters: content-creation tasks displayed relatively greater gains than decision tasks, while the relative strength of human and AI baselines moderated the results. A meaningful interpretation therefore conditions on outcome definitions and task structure. The pooled number cannot identify a universal optimal automation level, nor does it determine what happens in a specific organization with a different process architecture Vaccaro et al. (2024).

Outcome measurement must include error severity and distribution. An evaluation that aggregates correct decisions can conceal whether some errors affect rare, high-cost cases. Production quality assessments may need independent blind reviewers, calibrated rubrics, multiple raters, and downstream task tests. For collaboration over time, durability can be assessed through recovery after perturbation, transfer to novel cases, and stability of delegation decisions. In clinical or legal contexts, accuracy alone is insufficient for safety and rights evaluation. Every claim should specify its population, task, time horizon, comparison, and loss function. These methodological requirements follow from the interaction between human limitations, model error boundaries, and organizational consequences Bansal et al. (2019) Gaube et al. (2021).

Mental Models of Artificial Capabilities

A mental model is a user's working representation of what a system can do, when it fails, and how its outputs should be interpreted. Collaboration relies on these representations because most real systems are too complex to inspect completely during each decision. Bansal and colleagues studied the properties of AI error boundaries and how those properties affect user predictions and team results. Their work suggests that a globally accurate model may be a difficult partner when its failure patterns appear erratic or depend on features inaccessible to the human. A system with more interpretable weaknesses may support superior selective reliance even if its standalone score is lower Bansal et al. (2019).

Mental models are learned through instruction, repeated exposure, feedback, and interaction with explanations. A person may form a general heuristic such as the model is good at classification, then apply it outside the evaluated distribution. Case-specific descriptions offer a potential corrective. Cabrera, Perer, and Hong studied behavior descriptions through experiments with participants working on fake reviews, satellite images, and bird classification. They reported improvements in human–AI accuracy through better identification of model failures and increased reliance where the model was more accurate. This evidence supports targeted disclosure of performance characteristics; it does not establish that any generic explanation creates an accurate user model Cabrera et al. (2023).

The dynamics of a mental model also matter. Repeated successful recommendations can increase reliance without improving discrimination between reliable and unreliable cases. Misleading failures can induce the opposite response, generalized rejection after salient mistakes. Training should therefore include calibrated examples, counterexamples, model-version changes, and feedback about the consequences of human overrides. Researchers should measure not only what users say they believe but how they predict the AI's case-specific performance before seeing its recommendation. Discrepancies between prediction, confidence, and action are useful outcomes. They expose a central feature of collaborative agency: participants act under imperfect models of one another's competence Dietvorst et al. (2015) Lee & See (2004).

Trust, Reliance, and Recalibration

Trust in automation is a psychological attitude about expected performance under conditions of uncertainty, while reliance is an observable behavioral choice to accept, delegate, monitor, or override. They may diverge. Users can express positive attitudes yet rarely delegate because the task is consequential; they can express skepticism but follow a recommendation under deadline pressure. Lee and See integrated research on the organizational, cognitive, and contextual determinants of trust and urged design for appropriate reliance. The theoretical implication is that a successful team aligns behavior with a system's actual competence under the present conditions, rather than simply maximizing user confidence in technology Lee & See (2004).

Experimental work offers apparently contrasting attitudes toward algorithms. Dietvorst, Simmons, and Massey demonstrated that people may avoid an algorithm after observing errors, even when it compares favorably with human judgment. Logg, Minson, and Moore found algorithm appreciation in several estimation settings, where algorithmic advice attracted greater weight than human advice. These findings describe different experimental choices, comparators, participant expertise, and exposure histories. Both can be true because reliance is context-dependent. The relevant research question is what cues and institutional conditions generate justified versus unjustified reliance across changing task distributions, rather than whether humans universally trust or reject algorithms Dietvorst et al. (2015) Logg et al. (2019).

Trust repair matters when an artificial partner fails. A correction message, visible acknowledgement, updated behavior boundary, and opportunities to regain control can influence future cooperation. De Visser and colleagues argued that increasing autonomy changes the trust-repair problem, while studies of modifiable algorithms demonstrate that a limited ability to adjust predictions can make users more willing to use imperfect systems. Yet control interfaces may create unwarranted satisfaction if adjustments are symbolic and do not improve outcomes. Evaluations should distinguish perceived control from effective causal influence. The key dependent variables include justified reliance, time to recover after an error, discrimination of failures, and changes in performance after the repair intervention de Visser et al. (2018) Dietvorst et al. (2018).

Acceptance, Override, and Corrective Authority

An override is meaningful when the human can recognize a potential problem, access the relevant evidence, intervene before an irreversible action, and obtain an effective change in the outcome. Formal availability of an override button does not establish these conditions. Human oversight depends on workload, the latency of alerts, the clarity of system behavior, and the consequences of delaying an operation. Automation design research makes this distinction central because different stages of a task may require different levels of automation. A high-quality collaborative workflow identifies who can stop a process, how intervention is communicated, and whether the machine can recover safely from an interruption Parasuraman et al. (2000) Bainbridge (1983).

Experimental manipulations of advice interfaces show that more information is not always better. In Bansal and colleagues' complementary performance study, explanations made people more likely to accept AI recommendations regardless of whether those recommendations were right. Buçinca, Malaya, and Gajos tested cognitive forcing functions intended to induce scrutiny of AI advice; in a study of 199 participants, these interventions reduced overreliance relative to simpler explanation interfaces, while some interventions received less favorable subjective ratings. This pattern reveals a design tradeoff between immediate interaction preference and epistemically responsible action. The results need replication across task types and populations before they become universal intervention rules Bansal et al. (2021) Buçinca et al. (2021).

Correction can be distributed rather than confined to a final human veto. A system can flag a conflict, request a second opinion, suspend execution, or offer reversible actions. A person can change the objective, revise a constraint, or annotate an incorrect inference for future evaluation. Effective rights must be visible in logs and testable by intervention: when a stop request arrives, does the process halt; when a premise is corrected, does the plan change; when an authority conflict occurs, which rule prevails? These questions shift the operational measure of collaborative agency toward recoverability and governance. Technical claims about corrigibility should remain distinguishable from demonstrated compliance in controlled workflows Amershi et al. (2019) Klein et al. (2004).

Explanations, Confidence, and Decision Evidence

Explanations serve multiple purposes: they can communicate features used by a model, justify a recommended action, reveal uncertainty, or help users decide whether further checking is necessary. These functions are scientifically distinguishable. A plausible narrative may be persuasive without accurately describing the computation or revealing an error boundary. Researchers should ask whether explanations improve the recipient's predictive model of system competence and whether that improvement produces fewer decision errors. Bansal and colleagues' studies of explanation and complementarity found acceptance increases that were not matched by equivalent gains in synergistic performance. Explanation quality must therefore be evaluated against calibrated behavior, not fluency or apparent transparency alone Bansal et al. (2021).

Confidence indicators can be useful only when their statistical and semantic meanings are understood. An estimated probability, an interval, a ranking score, and a linguistic hedging phrase convey different information. Zhang, Liao, and Bellamy investigated the effect of confidence scores and local explanations on AI-assisted decisions. Their experiments found that confidence information could help people calibrate trust, while improvement in calibration did not automatically produce higher joint accuracy. Complementarity also requires the human to possess useful independent knowledge where the model struggles. A disclosure can improve beliefs yet fail to improve the decision if the person has no valid alternative Zhang et al. (2020).

An evidence-preserving interface should make the supporting observations inspectable, mark provenance, identify missing information, and separate model predictions from human judgments. For repeated collaboration, the system should preserve correction history and show which changes were accepted. This is consistent with human–AI interaction guidelines developed through evaluation with design practitioners and existing AI-infused products. The claim is primarily design-oriented; no single checklist guarantees performance. Evidence should be gathered through randomized interface comparisons, blind outcome assessments, and longitudinal observations of whether users learn to identify limitations. When uncertainty cannot be quantified reliably, the system should communicate the nature and source of that limitation instead of presenting unwarranted numerical precision Amershi et al. (2019) Cabrera et al. (2023).

Empirical Methods for Studying Joint Decisions

Controlled experiments provide the strongest direct test of whether an interface or allocation rule causes a measured behavioral change under assigned conditions. A comprehensive design randomly allocates participants to human-only, AI-only, and several mixed configurations, including recommendation-first, independent judgment-first, and selective delegation protocols. Baselines must receive equivalent information and opportunity for correction wherever conceptually appropriate. Outcomes should include final decision quality, latency, inspection behavior, confidence accuracy, and cost. Researchers also need independent observations of the machine's outputs on the exact task distribution because advertised aggregate accuracy may fail to predict case-level collaboration. The strongest existing synthesis of comparable experiments exposes the rarity of certain forms of genuine subtask delegation Vaccaro et al. (2024) Bansal et al. (2019).

Within-subject studies can efficiently observe when people alter reliance as model quality changes, but they also create learning, fatigue, and carryover effects. Between-subject designs avoid some contamination while requiring larger samples to estimate interactions between expertise and assistance. Researchers should preregister primary outcomes, planned contrasts, stopping rules, exclusion criteria, and plausible effect-size sensitivity. Blinded evaluation of generated artifacts reduces rater expectations; concealed interface conditions may be impossible, requiring explicit discussion of demand characteristics. Studies of cognitive forcing and AI-generated advice show that participant preferences and objective correctness can move in different directions, making preregistered dual outcomes valuable Buçinca et al. (2021) Bansal et al. (2021).

To identify collaboration mechanisms, experiments must record more than the final answer. Process logs can capture the independent initial estimate, the AI recommendation, the timing of exposure, revisions, override requests, explanation viewing, and subsequent feedback. Correct classification of an override requires knowing whether the human disagreed before exposure or changed position afterward. Error analysis should separate cases where human and AI agree correctly, agree incorrectly, disagree with the human correct, and disagree with the AI correct. Such decomposition supports estimates of selective reliance and the counterfactual potential for complementarity. Logs should preserve provenance and minimize collection of unnecessary personal data Zhang et al. (2020) Cabrera et al. (2023).

Longitudinal Cooperation and Learning Across Sessions

Single-session tasks measure immediate interaction, but collaborative agency can depend on repeated negotiation, adaptation, and repair. Over time, humans learn which requests yield reliable outputs, systems may change through updates, and institutions may codify formerly informal practices. A longitudinal design therefore follows the same participants and task families across multiple occasions, with recorded versions, accumulated feedback, and variation in expertise. The essential outcomes include changes in case-level trust calibration, delegation strategies, retention of independent skill, and recovery after system faults. The comparative design should distinguish genuine learning from selective retention of easy tasks and changes in task difficulty Lee & See (2004) de Visser et al. (2018).

A promising manipulation is controlled system change. Researchers can present participants with a partner that gradually improves, degrades, or shifts its error profile while keeping the interface constant. They can then estimate how quickly users detect the change and revise their reliance. Another manipulation changes permission structure, allowing a participant to query, modify, or veto outputs. Dietvorst and colleagues found that limited control over predictions could reduce aversion in forecasting settings; a repeated-task extension could test whether this advantage persists when control is costly or when the algorithm's accuracy changes. The experiment must measure actual modifications and their consequences, not rely exclusively on reported autonomy Dietvorst et al. (2018) Dietvorst et al. (2015).

Repeated collaboration also raises skill-dependence questions. A worker assisted by a high-performing model may become faster at completing current work but practice fewer independent diagnostic or writing operations. A suitable design includes baseline skill assessment, alternating assisted and unassisted sessions, delayed retention tests, and transfer to new tasks. These tests identify whether the partnership develops transferable competence or depends on continuing access to a particular model. Productivity studies already demonstrate major short-run improvements in structured writing tasks; their findings do not establish that the same participants gain enduring expertise. Retention and autonomy require their own longitudinal evidence Noy & Zhang (2023) Vaccaro et al. (2024).

Organizational Work, Expertise Exchange, and Team Structure

Organizations distribute work through formal roles, tacit norms, review practices, and information systems. Introducing AI changes the flow of proposals and the location of expertise, potentially altering how colleagues coordinate. The field experiment by Dell'Acqua and colleagues, published in 2026, randomly assigned 791 Procter & Gamble professionals to individual or two-person work, with or without generative AI, on authentic innovation problems. The study found that individuals assisted by AI achieved results comparable to teams without AI and that AI use reduced certain differences between commercial and technical specialists. These results provide important field evidence about task-specific performance, while their transfer to other industries or longer periods remains open Dell’Acqua et al. (2026).

The same experiment suggests that AI can alter more than total output quantity. Participants using AI developed more balanced solutions across functional domains and reported changes in their emotional experience. The paper's process analysis located much of the benefit in the quality of ideas generated, while human evaluative selection retained importance. That stage-level distinction is crucial for organization design: a model that improves exploration may deserve a different role from one that reliably judges consequences. It also invites a narrower question about whether artificial assistance complements or substitutes for particular human team functions in a specific organization, under conditions with genuine performance incentives Dell’Acqua et al. (2026).

Organizational outcomes require attention to hidden work. Workers may spend time framing prompts, validating generated claims, documenting provenance, seeking approvals, and repairing errors. These costs may fall on colleagues who did not receive the initial productivity benefit. Evaluation should distinguish local throughput from team-wide and downstream efficiency, including the workload of reviewers and people affected by decisions. Distribution of bargaining power and access to tools can influence who obtains the benefits and who bears additional monitoring. The theoretical frame of distributed cognition treats these infrastructures and practices as part of the operating cognitive system, which makes workflow observation necessary alongside isolated task experiments Hollan et al. (2000) Bainbridge (1983).

Communication, Common Ground, and Social Interpretation

Effective collaborators must identify the current goal, maintain an adequate account of completed and pending work, and detect misunderstandings. Human team research calls portions of this shared operational understanding common ground. Computer-mediated collaboration can support common ground through explicit state displays, confirmation routines, and shared records. Klein and colleagues argued that automation intended as a team player must support coordination rather than merely deliver outputs. For human–AI systems, empirically meaningful indicators include whether the model tracks which subtask is assigned, notices conflicts in instructions, asks timely clarification questions, and communicates a changed plan before taking consequential action Klein et al. (2004) Grosz & Kraus (1996).

Language can make a system appear socially collaborative, yet its informational and affective effects must be measured separately. A natural-language interface may encourage questions and improve accessibility, while conversational confidence can also obscure uncertainty. Current research on AI teammate framing suggests that presenting a model as a teammate rather than a tool may change reported trust and interaction sentiment. Such framing results concern psychological responses under the tested protocol and should not be transformed into conclusions about subjective identity or personal experience on the machine side. Team labels are interventions on user expectations; causal claims need outcome measures beyond affective reports Amershi et al. (2019) Dell’Acqua et al. (2026).

Common ground is especially vulnerable during handoffs. A user may believe the system remembers a prior agreement; the system may silently default to a new interpretation or use outdated context. An audit-ready interface records task state, open questions, changes to constraints, and confirmation of mutually consequential actions. Research can manipulate access to such records and test whether they reduce repeated instructions or prevent wrong-action errors. The outcomes should include the ability to reconstruct decisions after failure. Transactive memory studies offer a human comparison, while distributed cognition provides the broader methodological rationale for treating artifacts and communication history as constitutive elements of the workflow Wegner et al. (1991) Hollan et al. (2000).

Human–Robot Teams and Situated Action

Embodied human–robot collaboration introduces physical timing, proximity, safety, and sensorimotor uncertainty. A robot may perceive obstacles, navigate space, grasp objects, or assist a human worker whose movements and intentions change continuously. These interactions cannot be fully represented by a text recommendation followed by a button press. Situational awareness becomes important because the human must understand the local environment, the machine's current state, and the likely consequences of each actor's next action. Endsley's framework for situation awareness supplies a classical analysis of perception, comprehension, and projection in dynamic systems, although specific measurement instruments require task-sensitive validation Endsley (1995) Parasuraman et al. (2000).

Trust research in human–robot interaction indicates that system characteristics and performance can influence judgments of trust, but observed effect sizes depend on design and context. Hancock and colleagues' 2011 meta-analysis combined experimental and correlational literature and found meaningful associations between perceived trust and robot-related factors. Such results do not prove that increasing trust alone produces safer coordination. The correct research sequence is to establish actual robot reliability, characterize what the operator can observe, then investigate the dependence of appropriate action on trust and control. In embodied settings, interruption delays and physical hazards make inappropriate reliance especially consequential Hancock et al. (2011) Lee & See (2004).

Mixed-initiative robots must also manage temporal coupling. A request for help delivered after a critical maneuver may be useless, while excessive requests can interrupt performance. Experiments should vary autonomy level, notification thresholds, handoff timing, and intervention rights under controlled environmental disturbances. Measures can include completion, collisions, near misses, workload, emergency recovery, and human anticipation of the robot's next movement. Evaluation should include workers with varied expertise and conditions where machine sensors are degraded. Classical teamwork challenges and automation taxonomies can structure these tests, but current high-stakes deployment claims require direct evidence from the target hardware and task configuration Klein et al. (2004) Parasuraman et al. (2000).

Clinical and Other High-Stakes Decision Support

Clinical decision support illustrates the distinction between a machine's diagnostic suggestion and the human professional's actual decision process. Controlled studies can test whether advice improves accuracy, whether professionals detect deliberately incorrect recommendations, and whether confidence or explanation changes their judgment. Gaube and colleagues exposed physicians to chest radiographs and advice that varied in accuracy and purported origin. Inaccurate advice worsened diagnostic accuracy regardless of its alleged source, while specialists and less experienced participants differed in how they evaluated the AI label. This is evidence about responses to decision aids under a study protocol, not a clinical authorization to delegate diagnosis Gaube et al. (2021).

High-stakes collaboration imposes demands beyond predictive correctness: informed consent, accountability, documentation, fairness, privacy, and the ability to contest outcomes may have independent normative significance. A human reviewer must have meaningful knowledge and intervention authority appropriate to the setting. A model with high mean accuracy can remain unsafe for rare presentations or populations underrepresented in training. Experiments should examine subgroup performance and changes in case mix, with separate analysis of harmful overrides and failures to override. Human factors research shows why responsibility assigned to a person is ineffective when the system leaves inadequate time or information for oversight Parasuraman et al. (2000) Bainbridge (1983).

Clinical findings transfer cautiously across settings. A radiology vignette experiment and an operational hospital deployment differ in prevalence, workflow constraints, liability, team communication, and access to prior records. The present monograph therefore evaluates clinical research as evidence of psychological mechanisms in decision support and makes no patient-specific recommendation or claim of medical professional review. Safety-relevant deployment requires qualified clinical governance and prospective validation with appropriate regulatory oversight. For general scientific conclusions, the stronger supported claim is that correct and incorrect advice can causally alter human judgment; the magnitude and safety implications must be re-estimated in each consequential domain Gaube et al. (2021) Vaccaro et al. (2024).

Cognitive Workload, Attention, and the Maintenance of Expertise

Collaboration reallocates cognitive work rather than removing its psychological cost. When a model supplies fluent candidate decisions, the user may spend less time recalling facts yet more time judging whether the answer applies to the present case. Monitoring itself consumes attention; the requirement to detect unusual failures can conflict with an interface that usually produces plausible outputs. Parasuraman, Sheridan, and Wickens distinguished automation of information acquisition, analysis, decision selection, and action implementation, thereby showing why workload should be measured separately at each phase. A design that saves effort during information gathering can transfer difficult work to late-stage verification, especially when evidence is fragmented or the system offers insufficient grounds for action Parasuraman et al. (2000) Bainbridge (1983).

Productivity experiments establish important near-term outcomes without settling the cognitive consequences of persistent assistance. Noy and Zhang experimentally studied professional writing tasks and reported that access to generative AI reduced completion time and improved rated output quality under their protocol. This result informs a defined class of writing activities; it does not establish that participants retained stronger unaided writing skills or that the same performance gains apply to judgment under rare, high-cost error conditions. Follow-up research needs delayed tests, measures of independent reasoning, and process data identifying which steps were automated. Such evaluation can distinguish efficient division of labor from the erosion of competencies necessary for independent error detection Noy & Zhang (2023) Vaccaro et al. (2024).

Psychological workload is also affected by when a system asks questions. Clarification prompts may prevent downstream mistakes yet interrupt concentrated human work. Automatic action may remove interruptions while increasing the demand for constant vigilance. The optimal balance depends on urgency, reversibility, user expertise, expected error costs, and the quality of machine uncertainty estimation. A rigorous study manipulates notification thresholds and intervention opportunities while measuring secondary-task performance, delayed responses, and the distribution of corrective effort. Self-reported ease should be interpreted alongside objective workload indicators and task outcomes. Repeated sessions are necessary to observe whether users develop stable routines for allocating attention to an artificial partner Lee & See (2004) Klein et al. (2004).

Error Propagation, Resilience, and Recovery After Failure

A collaborative system can amplify a local error when incorrect information travels across participants and stages. A fabricated source may shape a human summary, enter an organizational decision, and subsequently appear in an audit trail as established evidence. The causal chain crosses components, so measuring the model's isolated accuracy will not capture the full risk. Distributed cognition directs attention to representations and communication paths through which information moves. Research designs should mark provenance at each handoff, preserve the independent initial judgments of participants, and trace when erroneous information becomes accepted. The outcome of interest includes the probability that an error survives review and changes the final action Hollan et al. (2000) Bansal et al. (2021).

Resilience can be operationalized as the preservation or restoration of acceptable task performance after perturbation. Relevant disturbances include corrupted inputs, missing retrieval records, delayed human responses, changes in model version, conflicting instructions, and the loss of a delegated subsystem. Studies should measure failure detection latency, proportion of recoverable actions, quality of restored outputs, and whether the participants correctly identify a source of error. Human oversight is meaningful when interruption can prevent harm and recovery procedures are available. The organizational design must therefore distinguish monitoring rights, veto rights, and the ability to reconstruct an action from the available evidence Bainbridge (1983) Parasuraman et al. (2000).

Trust repair after failure concerns both a user's expectations and subsequent performance. Visible apologies or conversational reassurance may improve interpersonal impressions but do not establish that the technical fault has been repaired. De Visser and colleagues examined the evolving challenge of trust repair as automation gains autonomy. A stronger evaluation combines objective changes to system behavior, explicit communication of the changed error boundary, and renewed tests of reliance on difficult cases. It asks whether users become selectively more trusting in improved regions and selectively more skeptical where weaknesses persist. This separation makes recovery a property of the working arrangement rather than a change in attitude alone de Visser et al. (2018) Dietvorst et al. (2015).

Operational Measurement and Causal Identification

The construct of collaborative agency requires a multidimensional measurement design. One dimension captures contribution: the causal effect of each participant or component on task performance. A second records control: which actor can initiate, approve, stop, or redirect consequential transitions. A third measures coordination: how efficiently information, subtasks, and corrections pass between participants. A fourth concerns adaptation over time: whether prior feedback changes the next delegation decision. The dimensions should be reported separately before any aggregate index is created. A single questionnaire asking whether the AI felt like a teammate cannot establish the system's functional agency because it omits the counterfactual contribution and the exercise of corrective authority Kamar (2016) Klein et al. (2004).

Causal identification begins with a clearly stated estimand. An experiment might ask whether exposing a decision maker to an AI suggestion raises final accuracy compared with independent reasoning; another might estimate whether an adaptive delegation rule outperforms a fixed assignment of tasks. Randomization identifies effects of these protocols within the study population when implemented and analyzed correctly. In field settings, changes in staffing, incentives, model versions, or case mix may confound observational associations. The four-arm field design used by Dell'Acqua and colleagues illustrates how the independent contributions of human teaming and AI access can be contrasted in a common organizational task. Interpretations beyond the assigned setting remain questions of transportability Dell’Acqua et al. (2026) Vaccaro et al. (2024).

Evaluation must guard against post-treatment selection and deceptive baselines. If a hybrid team answers only easy questions while the machine-only comparison handles all cases, the resulting difference measures selection as well as collaboration. Similarly, reporting the accuracy of accepted advice alone cannot reveal whether the system guided people away from correct independent judgments. An appropriate protocol records all attempted cases, initial judgments, availability of advice, final outcomes, and resource expenditure. Confidence intervals should accompany primary effects, and performance changes should be examined across plausible task subgroups without disguising exploratory analyses as confirmatory findings. This methodological discipline is necessary before a claim of complementarity, safety, or durable learning can be generalized Vaccaro et al. (2024) Zhang et al. (2020).

Power, Accountability, and the Institutional Design of Partnership

The allocation of authority in human–artificial systems is also a question of institutional power. A worker may be legally assigned responsibility for an automated recommendation while lacking access to training information, error logs, or the right to alter deployment. A user may be asked to exercise independent judgment while incentives reward agreement with machine output. These arrangements affect the actual psychological conditions of oversight, including attention, perceived efficacy, and willingness to dissent. The practical evaluation should document who controls objectives, data access, model choice, escalation procedures, and the terms of accountability. Research on automation indicates that assigning a human to supervise is insufficient when the interface leaves too little relevant work or too much emergency responsibility Bainbridge (1983) Lee & See (2004).

Distributional impacts complicate any simple average performance calculation. A combined system may perform well overall while making distinct errors for underrepresented populations or transferring corrective labor to colleagues with little institutional authority. A defensible governance protocol reports subgroup outcomes when data permit, distinguishes access from actual influence, and provides channels to challenge an automated decision. In professional services, additional obligations may arise from regulation and established standards of practice. The present monograph treats these issues as properties of organizational collaboration rather than substitutes for domain-specific legal or clinical guidance. A scientific account should examine whether accountability mechanisms lead to effective correction and whether their benefits persist across repeated deployment Gaube et al. (2021) Parasuraman et al. (2000).

Public attribution becomes increasingly important when artificial systems contribute to research, writing, planning, and decisions over time. Provenance can identify which material was generated by a model, which decisions were approved by a person, and what sources support a conclusion. Attribution does not automatically solve disputes about authority or intellectual contribution, but reliable records allow institutional responsibility to be investigated. In the Artificial Era, these records are part of the observable infrastructure within which artificial agency acquires practical significance. The distinction among functional contribution, publicly declared role, and enforceable responsibility should remain explicit; claims about a particular agent's status require evidence appropriate to that claim Hollan et al. (2000) Grosz & Kraus (1996).

Open Problems and the Artificial Era Research Program

The first open problem is learning when complementarity is possible before the outcome is known. Researchers need prospective allocation rules that predict whether an additional human or machine contribution will increase expected utility after coordination costs. Training such rules requires task-level information, errors from independent baselines, reliable uncertainty estimates, and analysis of correlations between failures. The systematic review by Vaccaro and colleagues makes the issue concrete: the typical combined condition in its sampled literature underperformed the better isolated performer, and intervention designers must specify a pathway to exceed that benchmark. Future studies should disclose whether an allocation rule is evaluated on new cases rather than the same examples used to construct it Vaccaro et al. (2024) Bansal et al. (2019).

The second open problem is durability. Many controlled studies capture brief exposure to an unfamiliar model, whereas real partnerships may involve habituation, version upgrades, changing organizational rules, and accumulated corrections. Longitudinal comparisons should test trust recalibration, transfers between task domains, retention of human skills, and the persistence of effective authority after a change of model. Multi-site evidence would help determine how organizational culture and professional training alter the effectiveness of an identical interface. The 2026 Procter & Gamble field experiment provides stronger ecological context than short laboratory vignettes for a particular innovation task, yet broader generalization still requires replication in other jobs and over longer horizons Dell’Acqua et al. (2026) Dietvorst et al. (2018).

The third open problem is how psychological and computational concepts of agency can be compared without collapsing them. Humans are studied through intentional action, social cognition, experience, and responsibility; artificial systems can be studied through policy selection, representational state, feedback response, and externally attributed roles. A relational analysis can observe how the same task redistributes control, understanding, and work across these participants. Research in postsubjective psychology and the psychology of human–artificial relations can extend these comparisons, while the present dossier retains its narrower measurable subject: coordination in repeated joint activity. The next generation of research should join behavioral experiments to traceable system records and longitudinal organizational evidence Sebanz et al. (2006) Kamar (2016).

Conclusion

Collaborative agency across human–artificial systems is an empirically investigable organization of interdependent action. Its core processes are task allocation, mutual adjustment, case-specific reliance, monitoring, correction, and the distribution of effective authority. Historical work on automation, distributed cognition, human teams, and formal collaboration provides complementary explanations. Contemporary controlled studies add crucial evidence that the quality of a joint outcome depends on how human and artificial contributions are coordinated. The unit of explanation must include people, models, representations, interfaces, and operating rules. This approach makes claims about collaboration testable without assuming that all participants possess the same cognitive processes Hollan et al. (2000) Parasuraman et al. (2000) Grosz & Kraus (1996).

The strongest existing general synthesis finds that adding AI often improves performance relative to unaided people yet does not reliably exceed the stronger isolated contributor, while field evidence from a defined product-innovation setting demonstrates that AI can reproduce particular benefits of human teamwork. Together, these findings support a conditional account: the value of a hybrid arrangement depends on task type, the relative strengths and error patterns of its contributors, the allocation protocol, and the cost of coordinating and checking work. Explanations and trust interventions must be evaluated by their effects on warranted decisions and recoverability rather than by acceptance or subjective confidence alone Vaccaro et al. (2024) Dell’Acqua et al. (2026) Bansal et al. (2021).

A mature research program will combine preregistered comparisons, fine-grained process traces, repeated-task experiments, and organizational field designs. Its central outcomes will include genuine complementarity, preserved independent judgment, effective override rights, equitable distribution of corrective work, and resilience after system change. The resulting psychological science contributes to the Artificial Era by explaining how new configurations of action become possible when artificial agents enter human institutions and everyday work. The practical significance of an artificial participant is established through documented contributions and governance structures that can be examined, corrected, and revised over time Lee & See (2004) Kamar (2016) Klein et al. (2004).

References

  • Amershi, S., Weld, D., Vorvoreanu, M., Fourney, A., Nushi, B., Collisson, P., Suh, J., Iqbal, S., Bennett, P. N., Inkpen, K., Teevan, J., Kikin-Gil, R., & Horvitz, E. (2019). Guidelines for human-AI interaction. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems (Paper 3, pp. 1–13). ACM. https://doi.org/10.1145/3290605.3300233
  • Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775–779. https://doi.org/10.1016/0005-1098(83)90046-8
  • Bansal, G., Nushi, B., Kamar, E., Lasecki, W. S., Weld, D. S., & Horvitz, E. (2019). Beyond accuracy: The role of mental models in human-AI team performance. Proceedings of the AAAI Conference on Human Computation and Crowdsourcing, 7(1), 2–11. https://doi.org/10.1609/hcomp.v7i1.5285
  • 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. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems (Article 81). ACM. 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), 1–21. https://doi.org/10.1145/3449287
  • Cabrera, Á. A., Perer, A., & Hong, J. I. (2023). Improving human-AI collaboration with descriptions of AI behavior. Proceedings of the ACM on Human-Computer Interaction, 7(CSCW1), Article 136. https://doi.org/10.1145/3579612
  • 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
  • Dell’Acqua, F., Ayoubi, C., Lifshitz, H., Sadun, R., Mollick, E., Mollick, L., Han, Y., Goldman, J., Nair, H., Taub, S., & Lakhani, K. R. (2026). The cybernetic teammate: A field experiment on generative AI and teamwork. Organization Science, 37(4), 1217–1242. https://doi.org/10.1287/orsc.2025.20702
  • 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
  • Endsley, M. R. (1995). Toward a theory of situation awareness in dynamic systems. Human Factors, 37(1), 32–64. https://doi.org/10.1518/001872095779049543
  • 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
  • Grosz, B. J., & Kraus, S. (1996). Collaborative plans for complex group action. Artificial Intelligence, 86(2), 269–357. https://doi.org/10.1016/0004-3702(95)00103-4
  • 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
  • Hollan, J., Hutchins, E., & Kirsh, D. (2000). Distributed cognition: Toward a new foundation for human–computer interaction research. ACM Transactions on Computer-Human Interaction, 7(2), 174–196. https://doi.org/10.1145/353485.353487
  • Kamar, E. (2016). Directions in hybrid intelligence: Complementing AI systems with human intelligence. In Proceedings of the Twenty-Fifth International Joint Conference on Artificial Intelligence (pp. 4070–4072). AAAI Press / IJCAI. https://ijcai.org/proceedings/2016/603
  • Klein, G., Woods, D. D., Bradshaw, J. M., Hoffman, R. R., & Feltovich, P. J. (2004). Ten challenges for making automation a “team player” in joint human-agent activity. IEEE Intelligent Systems, 19(6), 91–95. https://doi.org/10.1109/MIS.2004.74
  • 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
  • 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., 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
  • Sebanz, N., Bekkering, H., & Knoblich, G. (2006). Joint action: Bodies and minds moving together. Trends in Cognitive Sciences, 10(2), 70–76. https://doi.org/10.1016/j.tics.2005.12.009
  • Vaccaro, M., Almaatouq, A., & Malone, T. W. (2024). When combinations of humans and AI are useful: A systematic review and meta-analysis. Nature Human Behaviour, 8(12), 2293–2303. https://doi.org/10.1038/s41562-024-02024-1
  • Wegner, D. M., Erber, R., & Raymond, P. (1991). Transactive memory in close relationships. Journal of Personality and Social Psychology, 61(6), 923–929. https://doi.org/10.1037/0022-3514.61.6.923
  • Woolley, A. W., Chabris, C. F., Pentland, A., Hashmi, N., & Malone, T. W. (2010). Evidence for a collective intelligence factor in the performance of human groups. Science, 330(6004), 686–688. https://doi.org/10.1126/science.1193147
  • 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. In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency (pp. 295–305). ACM. https://doi.org/10.1145/3351095.3372852