RESEARCH DOSSIER
Automation Function Allocation and Supervisory Human Control
Abstract. Automation function allocation defines the distribution of information acquisition, interpretation, decision selection, and action execution among human and automated agents. Supervisory human control concerns the capacity to observe, understand, direct, override, and recover these operations. This critical narrative monograph reconstructs the field's historical origins and psychological concepts; examines automation taxonomies, empirical experiments, meta-analyses, and contemporary human–AI research; and develops a framework for evaluating routine efficiency, exception detection, transfer readiness, authority, and recovery. The evidence shows that automation's routine benefits can coexist with impaired supervisory awareness and failure recovery. Outcomes depend on stage and level of automation, reliability, task demands, interface design, training, institutional authority, and intervention opportunities. In the historical horizon of the Artificial Era, the work extends these questions to artificial systems that initiate decisions and actions, while keeping conceptual analysis distinct from validated evidence.
Keywords: automation function allocation; supervisory human control; human factors; out-of-the-loop performance; situation awareness; automation bias; adaptive automation; human oversight; trust calibration; takeover performance; human–AI interaction; artificial agency.
Definition and Scientific Scope
Automation function allocation is the distribution of information acquisition, interpretation, decision selection, and action implementation across human operators and technical systems. Supervisory human control denotes the operator's empirically demonstrable ability to comprehend, evaluate, direct, interrupt, and recover the behavior of an automated process. The scientific object is therefore an arrangement of cognitive work, authority, interfaces, and opportunities for effective intervention. It extends from conventional industrial process control to aircraft, medical decision support, autonomous vehicles, and artificial agents.
Control is meaningful when an operator can change an outcome within a useful time window. A formal approval right cannot establish effective oversight if the person lacks actionable evidence, relevant competence, or sufficient time. The scientific analysis must separate an operator's legal responsibility, institutional permission, subjective sense of agency, and measurable control capacity. This distinction explains why a system's advertised degree of autonomy is a poor substitute for detailed task analysis.
The foundational model of Parasuraman, R., Sheridan, T. B., & Wickens, C. D (2000) connects function allocation to human performance. The associated psychological constructs include situation awareness, vigilance, workload, mode awareness, trust calibration, reliance behavior, response selection, procedural knowledge, and recovery proficiency.
Research Questions and Methodological Approach
The central questions concern how function allocation changes routine output, monitoring, error recognition, intervention, and recovery. What type of automation improves performance under normal conditions? When does allocation reduce the operator's awareness of state changes? Which forms of feedback and authority support timely correction? How do system reliability, operator experience, task risk, and organizational arrangements change these effects? What constitutes an adequate human fallback when control must be reassigned?
This study is a critical integrative narrative monograph based on original experimental publications, foundational models, selected meta-analyses and systematic reviews, and official professional standards. It distinguishes the evidence supplied by controlled experiments from conceptual proposals, domain-specific engineering standards, and interpretive hypotheses. Quantitative values are reported only where they can be tied to identifiable source reports. It is not a newly conducted systematic review, a registered evidence synthesis, or a new primary experiment.
The scope is the psychological and operational organization of automation and supervisory control. Trust calibration, human–Artificial collaborative agency, and long-term skill retention are examined where they affect the allocation of control; their independent theoretical breadth belongs to neighboring research subjects. Such boundaries preserve a unique scientific object within Human Factors and Engineering Psychology.
Historical Foundations of Supervisory Control
Industrial automation changed the human role from directly adjusting every process variable to monitoring automatic control loops, diagnosing unusual conditions, modifying goals, and intervening during exceptions. Aviation similarly shifted many flight tasks toward systems capable of managing routine trajectories. These changes made the supervisory task a distinct human-factors concern because the conditions demanding human action became less frequent but potentially more complex.
Lisanne Bainbridge's 1983 analysis90046-8) formulated the ironies of automation: the human may retain the difficult, irregular, and failure-related work after routine tasks are transferred to a controller. This is a conceptual account of design constraints, not a numerical demonstration that every automation necessarily degrades safety. Later experimental studies supplied specific mechanisms. Endsley, M. R., & Kiris, E. O (1995) investigated the out-of-the-loop problem; Muir, B. M., & Moray, N (1996) studied supervisory trust in simulated process control; Sarter, N. B., & Woods, D. D (1997) analyzed operational experience and automation surprises on an advanced flight deck.
By 2000, the functional taxonomy of automation had become explicit in the work of Parasuraman and colleagues. Meta-analytic research then examined patterns across experimental tasks. Contemporary machine learning extends the same problem into recommendations, delegated digital actions, and artificial systems capable of initiating multi-step work.
The classic engineering history has a specifically documented teleoperation strand. Sheridan, T. B., & Verplank, W. L (1978), archived by the U.S. National Technical Information Service, examined human and computer control of undersea teleoperators, including the allocation of manipulation and sensory functions. Sheridan, T. B (1992) brought together experiments, control models, and applications in hazardous remote environments. This lineage makes clear that human supervisory control developed as an integrated question about sensing, remote action, and decision authority, well before contemporary conversational AI.
Norman, D. A (1990) argued that automation problems often arose from inadequate feedback and interaction with the humans responsible for the broader task. His analysis refined a simple question about the amount of automation into a question about information architecture and shared control. In the historical sequence, these studies complement Bainbridge: a machine can be technically capable of taking over regular operations while the surrounding interface leaves the human unable to infer why a process is departing from its intended course. The issue is visible in process control, aircraft, robotics, and modern agentic workflows, although each domain requires independent evaluation.
The Four Functions of Automation
Parasuraman, R., Sheridan, T. B., & Wickens, C. D (2000) identify four broad classes of functions. Information acquisition includes sensing, searching, selecting, and filtering observations. Information analysis includes classification, diagnosis, estimation, and synthesis. Decision and action selection includes choosing among possible responses. Action implementation includes executing commands and altering the external process. Each class can be automated to a different degree.
Automated acquisition can broaden coverage but also conceal unusual signals through filtering. Automated analysis can reduce calculation effort while changing the evidence available for human interpretation. Automated selection makes the machine's preferred response a psychologically salient reference point. Automated implementation changes the timing and reversibility of the resulting intervention. The allocation of a diagnostic function thus raises different questions than the allocation of a final command.
A defensible analysis describes every function in operational terms and asks what the human actually sees and can do. For instance, a human who receives only a system-generated summary may still formally own the final decision but lacks direct access to evidence used in the earlier analytical stages. The resulting arrangement should be characterized as an entire chain of dependencies rather than by the single label "human decision-maker."
Levels of Automation and Control Rights
Within each function, degrees of automation range from manual performance through machine assistance to increasingly automatic execution. A system can automate data collection almost completely while requiring explicit human authorization for consequential actions. Another can offer human review only after an action has already been initiated. These architectures differ even when their general interface appears similar.
The rights relevant to control include who may initiate work, propose alternatives, choose a response, approve it, veto execution, pause or stop the system, reverse a completed action, and assume fallback control. Rights must be paired with constraints: which conditions trigger the permission, what information accompanies the request, and how much time remains. Formal authority and effective authority are distinct measurement targets.
The interaction of stage, level, and control rights generates an allocation configuration. Experimental studies should manipulate these dimensions independently when possible. A comparison called "low versus high automation" can otherwise mix multiple causes, including display changes, speed, reduced practice, and changed responsibility. Good scientific inference depends on a specification precise enough for another research group to reconstruct the same allocation.
Supervision as a Human Feedback Process
Supervisory control can be represented as a sequence of observation, state inference, comparison with goals, intervention selection, execution, and evaluation of consequences. The operator observes a subset of system data and uses memory, expertise, and contextual cues to infer what is occurring. If deviations matter, the operator chooses whether and how to intervene. Feedback then determines whether the intervention succeeded.
Every link can fail. Sensor information may be delayed, summarized, or inaccurate. A user may interpret the data incorrectly, misjudge the controller's mode, or fail to anticipate the effects of a command. A stopping action may be technically available yet produce little benefit after an irreversible execution. Supervisory control therefore depends on observability and controllability together.
An adequate design distinguishes preventive control before execution, concurrent intervention during operation, corrective recovery after an error, and retrospective audit. The same interface may permit one but not another. Evaluating a so-called human-in-the-loop system requires evidence about the actual intervention window and authority at each phase, not simply the presence of a person on a workflow diagram.
Dynamic Allocation Across Operating Conditions
A static plan assigning tasks to humans and machines rarely captures the full dynamics of real operations. Operator workload, system reliability, environmental variability, and task risk change over time. A human able to review a routine recommendation may be unable to scrutinize a burst of conflicting alarms during a cascading failure. Conversely, a system capable of reliable operation under familiar input conditions can encounter observations outside its validated envelope.
Allocation can be fixed, adjustable by authorized participants, or adapted automatically according to monitored conditions. Dynamic reassignment may relieve temporary overload, preserve human engagement, or provide a path out of degraded automation. It also introduces a transition task requiring attention and shared understanding of why authority changed. Unannounced transfers can undermine predictability and create a new class of error.
Kaber, D. B., & Endsley, M. R (2004) examined intermediate and adaptive levels of control in dynamic tasks. Their research motivates measuring the complete path of assignments and handovers rather than assuming that every dynamic allocation improves performance. Adaptivity is a design variable with costs and potential benefits.
Routine Benefits and Failure Costs
Automatic control can reduce repetitive effort, increase consistency, and improve routine throughput. These effects are important and should be measured directly. They coexist with new demands associated with monitoring, exception diagnosis, and recovery. A system evaluated only under typical operating conditions may look superior even though its exceptional-state behavior is poorly understood.
An appropriate experiment distinguishes normal automation performance, degraded performance, unexpected failure, and post-failure recovery. The same operator should not be credited with safe supervision merely because a controller ran successfully while the operator observed it. Relevant endpoints include fault-detection latency, diagnostic accuracy, first response, final system state, and unintended consequences of intervention.
The integrated Onnasch, L., Wickens, C. D., Li, H., & Manzey, D (2014) found increasing routine-performance benefits with increasing degree of automation, with weaker workload benefits during proper operation and adverse patterns for situation awareness and failure performance. The results are not a universal engineering limit. They show why total-system evaluation must cover the full outcome distribution, including low-frequency events that may dominate the consequences of an operational failure.
Out-of-the-Loop Performance
The out-of-the-loop performance problem arises when automation displaces human participation in active control or information processing, leaving the person less prepared to recognize and manage an unexpected state. The mechanisms can include loss of situation awareness, reduced attentional engagement, insufficient feedback, and decay of practiced control skills. Each mechanism operates on a different timescale and requires different measurements.
In a navigation-task study, Endsley, M. R., & Kiris, E. O (1995) linked reduced situation awareness with poorer decision timeliness after an automated expert-system failure. The operator's degree of control moderated the effect. The study's interpretation emphasized the shift from active to passive processing. It supports a context-specific explanatory mechanism but cannot establish an inevitable loss of all skills under every form of automation.
An experimental program should vary the degree of prior engagement, length of reliance, available feedback, and response requirements independently. An immediate takeover after passive observation probes awareness and reorientation; a delayed test after extensive automation exposure can additionally probe retained competence. Treating these distinct phenomena as one generic "automation dependence" measure obscures useful causal distinctions.
Situation Awareness and Future-State Prediction
Situation awareness, in Endsley, M. R (1995), includes perceiving relevant elements of an environment, understanding their meaning, and anticipating their future status. A supervisor must represent not only the external process but also the automation's goals, mode, information sources, pending actions, and limits of operation.
A display that reveals the latest state variable without explaining how it was derived may preserve surface awareness while limiting causal understanding. A detailed display can create the opposite problem by increasing attentional demand. The design goal is task-appropriate support for accurate state reconstruction when intervention becomes necessary. This can involve visible mode transitions, salient deviations, access to underlying evidence, and forecasts tied to the controller's current policy.
Situation awareness can be probed through recall of task state, identification of relevant changes, prediction of near-term events, and demonstrated response to unexpected failures. These measures should be interpreted separately from confidence and subjective workload. Successful performance during a stable automated episode cannot alone establish readiness for an unfamiliar exception.
Mode Awareness and Automation Surprises
A mode specifies a controller's operating logic, interpreted inputs, and permitted actions. Mode awareness is the operator's understanding of the currently active configuration and the conditions that will cause it to change. Automation surprises occur when the system behaves in ways that operators did not anticipate, even if the behavior is consistent with its internal rules.
In research on the Airbus A-320, Sarter, N. B., & Woods, D. D (1997) linked operational difficulties to operators' limited awareness of current and future automation behavior. Their work demonstrates that supervising a system requires a mental model of what it will do next, not only knowledge of the external environment.
Tests of mode awareness can manipulate transition frequency, display labels, warning delays, and unexpected controller responses. Measures include correct mode naming, prediction of the next system action, time to notice a mismatch, and ability to restore a stable state. Designs that perform well on routine mode recognition still require tests of ambiguous and rapidly changing transitions.
Vigilance, Alarm Design, and Rare Events
Supervisors of reliable automation encounter an unusual attentional challenge: important signals may be both consequential and infrequent. Normal operation generates relatively few opportunities to practice anomaly discrimination. Monitoring can become less active or compete with other work, and false alarms may shift the operator's willingness to respond to future warnings.
Performance depends on event probability, discriminability, payoff structure, workload, and the informational quality of alarms. Signal detection analysis can distinguish discriminating real faults from adopting a liberal or conservative response criterion. Hit rates without false-alarm rates give an incomplete picture. The eventual safety effect also depends on what happens after detection.
Parasuraman, R., & Manzey, D. H (2010) integrated complacency and automation bias through attentional processes, highlighting concurrent-task competition. A realistic monitoring experiment should test how alarms interact with competing tasks and evaluate correct intervention after alarms, rather than assuming that assigned supervisory responsibility guarantees sustained attention.
A particularly useful early experiment was conducted by Parasuraman, R., Molloy, R., & Singh, I. L (1993). The published abstract reports two experiments using a flight simulation with monitoring, tracking, and fuel-management tasks. Forty participants performed four 30-minute sessions, and automation reliability was either stable or alternated between low and high levels. Operators were markedly worse at detecting automation failures under constant reliability after roughly 20 minutes in the automation condition; this difference did not arise in the single-task monitoring condition. The result connects under-detection to sustained multitask monitoring and the reliability experience established by the system.
The interpretation requires care. The manipulation of reliability in a short simulation does not show that variable unreliability should be engineered into safety-critical products. It reveals that attention to monitoring changes under predictable reliability and competing demands. The findings support studying variation in reliability, monitoring incentives, independent feedback, and the temporal development of reliance, while preserving the actual hazard consequences in each application.
Workload and Multiple Resources
Automation redistributes cognitive work. It can reduce routine calculation or physical control while increasing the costs of checking recommendations, switching modes, and diagnosing abnormal behavior. Average workload during successful operation may therefore understate the peak burden imposed by failures. This temporal redistribution is one mechanism connecting apparently efficient automation with poor exceptional-state performance.
Wickens, C. D (2008) provides a framework for studying interference between concurrent tasks according to their perceptual, cognitive, and response demands. A visual explanation competing with visual fault recognition can produce different interference from an appropriately designed auditory cue. System designers should therefore evaluate not only the amount of information supplied but also when and how it is presented.
Subjective instruments such as the NASA Task Load Index62386-9) provide one perspective. Objective performance, secondary-task costs, physiological data, and situation-awareness measures offer complementary evidence. A lower workload rating does not demonstrate safer supervision unless the system retains adequate detection, decision, and recovery performance.
Trust and Appropriate Reliance
Trust refers to expectations about a system's competence and behavior under uncertainty; reliance refers to observable acceptance or use of its outputs. The two are related but distinct. A person may express trust while routinely verifying results; another may depend on a system despite low trust because alternative decision procedures are unavailable.
Lee, J. D., & See, K. A (2004) framed the design objective as appropriate reliance, meaning that acceptance and rejection should track actual system capability in context. Muir, B. M., & Moray, N (1996) experimentally examined trust and intervention in simulated process control, while Hoff, K. A., & Bashir, M (2015) synthesized factors influencing trust into a layered model.
A good calibration experiment manipulates true recommendation validity, apparent source, uncertainty disclosure, verification costs, and feedback. Its outcome is not trust maximization or simply fewer overrides. It measures correct acceptance of valid outputs and correct rejection of invalid ones, relative to the cost and consequences of each action. Trust is one psychological mediator among several, and confidence reports require validation against actual behavior.
Omission and Commission Errors
Automation bias has two well-studied behavioral forms. An omission error occurs when a person fails to recognize a problem because an automated aid did not signal it. A commission error occurs when the person follows an erroneous automated recommendation despite contradictory evidence. The two outcomes imply different points of failure in the supervisory sequence.
In a controlled simulated flight task, Skitka, L. J., Mosier, K. L., & Burdick, M (1999) documented omission and commission errors associated with decision support. Mosier, K. L., Skitka, L. J., Heers, S., & Burdick, M (1998) studied automation bias in high-technology cockpits, including differences related to operators' self-reported accountability and double-checking strategies. These findings are evidence of particular task-specific vulnerabilities rather than proof that automated assistance always impairs decision making.
To identify a mechanism, researchers can manipulate the reliability of automated cues, visibility of independent indicators, error direction, task stakes, and opportunity to verify. Failure to detect an incorrect output must be distinguished from unwillingness to challenge an output recognized as uncertain. Different interface and organizational interventions may be appropriate for these two situations.
Selective Nonuse and Algorithm Aversion
Reliance may also be inappropriately low. Users can reject an accurate aid after a salient error, distrust unfamiliar computational methods, or resist recommendations that threaten professional autonomy. A supervisor may be both over-reliant under some conditions and under-reliant under others. The scientifically relevant quantity is correspondence between behavior and competence, conditional on the task and its consequences.
Override frequency alone cannot establish quality. If a machine is correct on most cases, a low override rate might reflect competent reliance or passive compliance; if errors are common and detectable, the same rate may be unsafe. Conversely, a high override rate can indicate appropriate expert correction or needless rejection. Analysis requires validity-labeled recommendations and comparison with counterfactual human-only and machine-only performance.
Atanassova, I., Khan, H., & Khan, Z (2026) propose distinguishing override rates from override accuracy as part of a trust-calibration routine. Their contribution is a practice-oriented synthesis; the proposed routine is not itself evidence that every organizational deployment will improve outcomes. Its measurement logic is valuable for experimental research.
Transfer of Control and Handover Protocols
A handover involves transferring authority, context, and responsibility for a defined task. Taking over includes recognizing the request, reconstructing the relevant situation, selecting an appropriate response, initiating action, and stabilizing the resulting operation. Measuring the first physical input as the entirety of takeover time ignores essential phases of cognitive recovery.
Handover protocols should specify trigger conditions, advance warning, the information to be passed, acknowledgment procedures, transfer authority, permitted refusal, and fallback if transfer fails. The protocol must address the possibility that the recipient is unavailable or currently engaged in other demanding tasks. Technically successful transfer does not establish that the recipient has acquired effective control.
A defensible time budget is derived from the distribution of response and stabilization performance in realistic operating circumstances, with uncertainty and critical scenarios represented. A single average reaction-time threshold is an inadequate guarantee. Human supervisors require opportunities to regain context and act before a system reaches an unrecoverable state.
Automated Driving as a Takeover Laboratory
Automated driving supplies unusually explicit definitions of responsibility for the dynamic driving task and fallback. The SAE International (2026), current in its September 2026 revision, distinguishes assistance requiring continuous human supervision from conditional automation that performs the driving task within a defined domain while expecting a human response to an intervention request. This is a specialized technical taxonomy, not a classification to apply without modification to general-purpose AI.
Merat, N., Jamson, A. H., Lai, F. C. H., Daly, M., & Carsten (2014) used a simulator to study the transition from automated to manual driving. Their results distinguish an initial resumption of control from subsequent stabilization, with context-specific times that should not be used as universal safety thresholds. A Weaver, B. W., & DeLucia, P. R (2022) found impaired takeover performance during certain non-driving tasks, especially when cognitive resources overlapped with driving. A Jin, L., Liu, X., Guo, B., Han, Z., Wang, Y., Cao, Y., Yang (2025) examined non-driving activity and takeover-request modalities, describing advantages associated with multimodal requests under its studied conditions.
These studies show how temporary removal from a task changes the response to an unexpected request. Generalization to aircraft, process plants, or software agents requires separate analysis of the available time, task complexity, consequences, and interface.
Adaptive Automation and Adjustable Autonomy
Adaptive automation reallocates functions when conditions change. Possible triggers include workload estimates, detected errors, environmental disturbances, operator actions, or system confidence. Adjustable autonomy allows authorized participants to change the allocation explicitly. Both architectures create new tasks: recognizing the need for a change, explaining why it occurred, coordinating the transition, and restoring a shared representation of who currently controls what.
Kaber, D. B., & Endsley, M. R (2004) explored adaptive automation in dynamic control. A rigorous evaluation should manipulate transfer triggers and thresholds, compare them with a fixed-allocation baseline, and measure changes in routine output, situation awareness, false transitions, fault recovery, and subjective workload. It should also estimate the price of unpredictable reassignment.
Adaptivity should therefore be treated as an empirical policy hypothesis rather than a design virtue in itself. A complex controller that changes authority too often may reduce operators' ability to predict its next action. A fixed controller that never responds to overload can fail for different reasons. The appropriate approach depends on the entire sequence of work and conditions under which reassignment is necessary.
Mixed Initiative and Negotiated Human Machine Roles
In mixed-initiative systems, either the human or the machine may propose actions, ask for information, or request a transfer of control. This differs from a simple recommendation interface because initiative is distributed throughout a work episode. A digital agent might gather evidence, suggest a plan, seek confirmation, execute approved steps, and revise its approach when new information appears.
The psychology of this arrangement depends on who initiates an action, whether the proposal is transparent, how disagreement is represented, and what costs accompany refusal. A design that automatically executes after a short timeout may preserve a nominal approval step while making dissent difficult. Conversely, requiring confirmation for every trivial operation may encourage habitual acceptance and divert attention from consequential decisions.
Research can vary initiation rights, consent defaults, reversibility, and access to supporting evidence. The primary outcomes are correct delegation, rejection of invalid proposals, response time, system performance, and the operator's demonstrated understanding of the current authority arrangement. Interaction logs enable process-level analysis of proposals, objections, retries, and recovery.
The conception of automation as a participant in coordinated activity also has a documented human-centered computing history. Klein, G., Woods, D. D., Bradshaw, J. M., Hoffman, R. R., & (2004) described ten challenges involved in making automation a genuine team participant. Their contribution is a conceptual analysis of coordination requirements rather than an experiment establishing that autonomous agents achieve those requirements. Its relevance lies in the need for aligned expectations, shared information, and action coordination when the machine can make independent proposals.
A design that permits machine initiative must support the operator's awareness of what the artificial participant has already done, what it intends to do, and which obligations it assumes. Conversely, the machine needs information about human commitments and the possibility of refusal or interruption. Experiments could test coordinated versus uncoordinated initiative by introducing ambiguous instructions, changed priorities, or conflicting alerts and measuring repair sequences. Successful collaboration becomes a property of traceable joint activity, not simply a label applied to a system capable of independent action.
Explanation and Operational Transparency
Supervisory interfaces should help users identify what a system proposes, what evidence supports it, how uncertain the proposal is, and what the consequences of action may be. Transparency is useful when it improves discrimination between valid and invalid machine behavior and supports correct action. The mere length or fluency of an explanation is not an adequate measure.
An automated recommendation may appear persuasive because its reasoning is well worded while still being wrong. An interface can also be transparent about a confidence score yet obscure how that score changes under unfamiliar conditions. Evaluations should therefore use controlled incorrect recommendations, varied explanation types, and observable measures such as error detection, correct override, time to diagnose, and post-intervention outcomes.
Explanations should be available at a level appropriate to time pressure and expertise. Operators need a concise actionable summary during a critical event and access to deeper evidence when investigation is feasible. The appropriate interface provides both while avoiding conflicts with other sensory and attentional demands.
Fault Detection and Escalation
An automated system can fail through sensor errors, classification mistakes, flawed estimates, software faults, communication loss, inappropriate action selection, or operation outside a validated environment. A fault taxonomy should distinguish failures that the controller can resolve autonomously from conditions requiring external intervention. This partition forms part of the allocation itself.
Escalation rules should be evaluated for detection sensitivity, false alarms, delay, clarity, and ability to place the system in a safer state. A high numerical confidence score does not guarantee accuracy under distribution shift, and a low confidence score does not specify the correct human response. More useful are task-specific indicators of uncertainty, contradictions, unusual observations, and constraints that the machine can no longer meet.
Experiments should include both obvious and subtle failures, multiple fault sequences, and conditions where automation produces a plausible but incorrect result. The operator's ability to respond after the first attempted correction fails is also an important measure. Effective supervision requires a sequence of workable alternatives, not just a single veto action.
Manual Skill and Longitudinal Competence
Function allocation changes what operators practice. Repeated delegation of routine control may reduce opportunities to acquire or maintain diagnostic and motor skills needed for emergency fallback. Immediate loss of situational awareness and long-term change in practiced competence are different hypotheses and must be studied separately.
Longitudinal experiments can provide initial training, randomize periods of automated versus manual participation, and introduce delayed unassisted recovery tests. Measurements should include first-response correctness, stable execution, fault diagnosis, procedural fluency, and retention across time. Practice intervals, expertise, task complexity, and simulation fidelity may all moderate results. Workplace studies require attention to selection effects because employees with different skill levels may be assigned different automation arrangements.
The adjacent dossier Automation-Related Skill Retention and Human Competence Continuity addresses this long-duration mechanism as its primary object. Here, the issue is whether retained competence supports a realistic supervisory fallback at the moment control must return to the person.
Team and Organizational Supervisory Control
Organizational structures shape effective authority. A supervisor may possess a technical right to stop a process but face productivity targets, hierarchical pressure, or procedural ambiguity that discourage intervention. In complex work, information may also travel through several specialized teams before anyone can authorize a response. These conditions alter both response time and the probability of independent verification.
An organizational task analysis should identify the persons responsible for detection, diagnosis, recommendation, approval, implementation, and post-event investigation. It should also document who may act when another role is unavailable, how conflicting opinions are resolved, and which records permit retrospective reconstruction. These details matter as much as the interface of a single device.
Field observations, incident reviews, interviews, and longitudinal logs can reveal interaction patterns that short laboratory tasks omit. Controlled experiments remain valuable for isolating specific mechanisms, but institutional effectiveness requires validating findings under real incentives, competing duties, staffing conditions, and consequences.
Experimental Design for Function Allocation
A strong causal design specifies the function being automated, the level of automation, human control rights, machine reliability, failure modes, and the complete information available to each participant. Factorial experiments can cross decision authority with information access to determine whether poor oversight arises from lost control, missing evidence, or their interaction.
Studies should predefine routine and exceptional events, outcome measures, and contrasts. Within-subjects experiments can compare several configurations efficiently but require counterbalancing to address practice and carryover; between-subjects experiments reduce some contamination but may need larger samples. Reliability manipulations must ensure that a purported machine failure is observable as a failure under the stipulated task criteria.
The experiment should include human-only, machine-only, and combined-system baselines where feasible. Otherwise, the claim that collaboration improves performance may simply mean that human users become better than an unaided human baseline without surpassing the machine alone. Comparative baselines are indispensable for assessing the value of allocation.
Consider a preregistered experiment with four conditions: human analysis and human choice; machine analysis and human choice; machine analysis and machine choice with a human veto; and machine analysis, choice, and implementation with a human emergency-stop right. The experiment can hold task content, available training, and probability of faulty analysis constant while varying the control architecture. After normal trials, a controlled set of cases contains wrong but recognizable analyses, valid analyses leading to wrong selected actions, and correct planned actions that fail during implementation. Those faults reveal whether available supervisory rights match the stage at which error occurs.
Primary estimands can include the difference in correct final outcomes between configurations during normal and failure conditions, the difference in time to detect and correct faults, and the interaction between control rights and information accessibility. A mixed-effects analysis can account for repeated participants and tasks where the design supports it. Researchers should state in advance how they treat timeouts, failed interventions, and cases where no safe alternative exists. Power analysis must reflect the relevant effect and clustering structure rather than relying on arbitrary participant counts.
Such a study can show whether a human veto is effective only when the reviewer sees the evidence behind the machine's chosen action, or whether other arrangements achieve comparable performance with lower workload. It cannot by itself establish long-term effects on professional competence, which require a separately powered longitudinal design.
Outcome Measurement and Construct Validity
Human supervisory outcomes can be organized along a task trajectory. Acquisition is represented by detection of relevant signals; analysis by diagnostic accuracy; selection by the appropriateness of chosen action; implementation by successful execution; and recovery by restoration of a stable operating state. Additional outcomes include recognition of automation modes, time to identify faults, decision confidence, workload, and the cost of incorrect intervention.
Override frequency is not equivalent to override accuracy. Takeover initiation is not equivalent to safe stabilization. An operator's reported sense of being in control is not identical to demonstrated influence over a process. Construct validity depends on keeping such measures separate and interpreting them within clearly defined scenarios.
Instruments require evidence of reliability, scoring consistency, and sensitivity to relevant changes. Results should include effect estimates and uncertainty, not only statistical significance. For rare adverse events, simulation permits targeted fault injection but leaves open questions of transfer to natural environments. The entire measurement system should be reproducible enough to support independent comparison.
Evidence from the Automation Meta Analysis
Onnasch, L., Wickens, C. D., Li, H., & Manzey, D (2014) combined data from 18 experiments examining degree of automation, routine performance, workload, situation awareness, and performance when automation fails. The synthesis reports increasing routine benefits as automation advances and an adverse pattern for failure performance and situation awareness. The authors identify a consequential transition between automation of information analysis and automation of action selection.
The significance lies in the interaction of task stage and degree of authority. A machine may support a human by presenting analysis without requiring the person to reconstruct an entire control strategy when automation fails. Moving into automatic action selection can change how operators attend to and evaluate alternatives. These mechanisms are plausible explanations, and the precise relationship varies by task design.
This synthesis should not be flattened into an assertion that one level is always superior. It pools specific experiments with differing procedures, samples, and outcomes. Application requires knowledge of failure probability, stakes, recovery feasibility, and the nature of the human alternative. A safety-relevant allocation decision should evaluate expected routine benefits and the distribution of exceptional harms.
Human AI Performance and Comparative Baselines
Human–AI systems represent a contemporary extension of automation research. A machine can analyze information or recommend a choice, while a person reviews context and decides whether to rely on the suggestion. The joint system's performance depends on error patterns of each participant and on whether the interface allows correct discrimination between cases that warrant acceptance and those that warrant rejection.
Vaccaro, M., Almaatouq, A., & Malone, T. W (2024) conducted a preregistered systematic review and meta-analysis of 106 experimental studies with 370 effect sizes. The analysis compared humans alone, AI alone, and combined human–AI arrangements. Against the better of the two independent baselines, the pooled effect for the combined system was negative, Hedges' g = −0.23 (95% confidence interval −0.39 to −0.07). Against the human-only baseline, combinations showed improvement on average. Decision tasks and content-creation tasks exhibited different patterns.
These results demonstrate why comparator selection matters. An AI aid can improve a person's performance while producing an arrangement inferior to autonomous machine output in the same tasks. Conversely, the machine can make critical errors that only the human recognizes. Research must locate the conditions under which the combined control architecture yields genuine added value and avoid using increased subjective satisfaction as a substitute for correctness.
Source Labels and High Stakes Expert Oversight
The perceived identity of a recommendation's author can affect human scrutiny independently of the recommendation's objective quality. In a preregistered randomized grading experiment, Goulas, S., Megalokonomou, R., & Sotirakopoulos, P (2026) asked education experts to evaluate work paired with intentionally inaccurate initial scores labeled as human- or AI-generated. The researchers also manipulated whether the score was too harsh or too lenient.
For harsh recommendations, the published abstract reports a fairness gap 22% larger under the AI label. For lenient recommendations, differences between AI and human source labels were not statistically distinguishable. The finding is conditional on error direction and task setting. It supports a psychologically differentiated theory of deference rather than a blanket claim that experts invariably trust AI more than humans.
In high-consequence domains, successful oversight requires more than a signed confirmation. Researchers should test whether reviewers can identify errors under realistic time constraints, whether evidence is independently accessible, and whether correct rejection changes the final outcome. Medical, judicial, employment, and financial deployments have distinct professional, legal, and ethical requirements that must be assessed within their own domains.
Artificial Agency and Initiative Allocation
Increasingly capable artificial systems can propose plans, initiate tool use, request information, and execute connected actions. The supervisory problem consequently extends from evaluating isolated recommendations to governing a changing sequence of initiatives. Allocation needs to specify who may begin an operation, alter its objective, use external tools, approve irreversible effects, and terminate the process.
The Artificial Era provides an intellectual horizon for comparing organized cognitive action across Homo and Artificial. For empirical study, claims about artificial agency should be expressed through observable policy selection, action traces, continuity across tasks, feedback integration, and capacity to adapt under controlled conditions. Philosophical categories supplied by the Aisentica authorial system can guide questions; empirical conclusions about particular systems require operational definitions and independent tests.
Researchers can experimentally compare human-initiated, machine-initiated, and negotiated transfers while holding task difficulty and competence stable. The dependent variables include correct intervention, error propagation, user understanding, reliance calibration, and successful completion. Such studies translate conceptual inquiry into testable findings without assuming that human and artificial cognition have identical underlying mechanisms.
Human Centered Design and Governance
Human-centered system development investigates real users, goals, operational environments, and the effects of interfaces on work. International Organization for Standardization (2019) provides a relevant design standard, while the National Institute of Standards and Technology (2023) identifies governance, mapping, measurement, and management functions relevant to AI risks and human–AI configurations. These documents specify important processes; neither proves that a particular product's human supervision is effective.
Evaluation must accompany deployment. Training, software updates, changes in model behavior, and changing work routines can alter the user's understanding of the automation and the opportunities for effective intervention. A first-session usability test may miss long-term reliance patterns and rare failures. Controlled fault-injection trials, field monitoring, and incident learning supply complementary forms of evidence.
Good governance assigns concrete responsibilities for testing, escalation, documentation, and review. It also establishes clear boundaries around automatic execution, conditions for human veto, and procedures when the human is unavailable. The empirical measure of such safeguards is their demonstrated effect on detection, prevention, recovery, and harm in representative conditions.
Research Limitations and Generalizability
Evidence across this field is heterogeneous. Some foundational works are conceptual models, some involve controlled laboratory tasks, some rely on high-fidelity simulation, and others synthesize results across specialized contexts. Their findings cannot be assigned one universal level of certainty. A cognitive effect observed in a navigation simulation may be informative for industrial control while its magnitude and boundary conditions remain unsettled.
Important limitations include small experimental samples, inconsistent construct definitions, artificial rates of rare failures, short exposure periods, and outcomes that prioritize first reaction over stable recovery. Field studies can improve ecological validity but introduce selection effects and confounding. Systematic reviews can identify regularities while masking substantial variation in training, authority arrangements, and task consequences.
Cross-population validity also deserves attention. Operators differ by expertise, accessibility needs, age, experience with automation, and institutional authority. A system optimized for expert supervisors may not generalize to public-facing technology. Transparent reporting of task conditions, failed replications, uncertainty intervals, and competing explanations is therefore indispensable.
Research provenance should specify precisely what investigators and reviewers could inspect. Some landmark articles are available only through publisher abstracts or restricted full text, whereas others have accessible primary accounts and numerical results. A bibliography can accurately identify a study while leaving portions of its method or effect estimation unverified. Readers and editors should be told which level of source access underlies a claim. In this monograph, carefully identified abstracts support the limited assertions attributed to several studies; detailed effect magnitudes are used only for the accessible primary reports or clearly identifiable published abstract statements.
An integrative narrative monograph also lacks the sampling properties of a comprehensive systematic review. The chosen bibliography is purposive, not the product of a documented database search with predefined inclusion and exclusion criteria. This limits claims about exhaustiveness and publication bias, although the monograph can still critically compare established conceptual work, original experiments, and formal meta-analyses. An independent systematic review would require a registered protocol, search strings, screening flow, structured risk-of-bias assessment, and reproducible extraction procedures.
Open Problems and Research Program
Several unresolved questions organize future work. How much active engagement is required to preserve takeover readiness? Which information representation makes mode changes understandable without imposing excessive workload? How do trust and perceived institutional responsibility jointly influence correct override? Which adaptive policies reduce overload without creating unpredictable transitions? Under what conditions can an artificial agent safely initiate consequential actions?
These questions can be pursued with preregistered factorial studies manipulating automation stage, control rights, feedback, and failure type. Longitudinal training studies can test delayed competence. Field research can examine whether formal intervention authority functions under organizational incentives. Designs should quantify correct acceptance, correct rejection, false alarms, recovery quality, and the costs of intervention rather than relying on one overall satisfaction score.
Research must also distinguish technical feasibility from socially authorized control. An architecture can be capable of making a decision without being legitimately empowered to do so. Empirical inquiry should document these boundaries and analyze their effects on behavior. Scientific progress will come from comparing explicit configurations and their consequences rather than treating automation as a single explanatory variable.
Allocation as a Risk-Weighted Optimization Problem
Function allocation is a design decision under uncertainty. If an automated controller performs better than a human in common conditions, transferring that function may raise average performance. The same transfer can make a rare failure more difficult to recognize or correct. A scientifically defensible comparison must include the probability and severity of different operating conditions, rather than evaluating automation only where it performs best. An arrangement can be optimal for throughput and unsuitable for a safety-critical process if its low-frequency adverse outcomes are catastrophic.
The objective is usually multi-dimensional. Relevant criteria include correct decisions, total task completion time, human workload, intervention reliability, recoverability, operator comprehension, maintainability, and social or institutional constraints. These outcomes cannot always be condensed into a single score without explicit value judgments about their relative importance. A risk-weighted analysis specifies which consequences matter, how they are measured, and which uncertainty remains in their estimated distribution. Sensitivity analysis can reveal whether a preferred architecture depends on uncertain assumptions about failure frequency or takeover readiness.
An important distinction concerns the difference between the best independent performer and the best combined configuration. Allocating every stage to the participant with the higher average isolated accuracy can produce poor joint performance when outputs are difficult to verify, errors are correlated, or switching costs are large. The design problem is therefore systemic. The evidence-based performance trade-offs described by Onnasch, L., Wickens, C. D., Li, H., & Manzey, D (2014) make the joint distribution of routine benefits and exceptional losses a primary scientific target.
The Operational Design Envelope and Boundary Recognition
A system's operating envelope specifies the circumstances in which its performance and constraints have been evaluated. In automated driving, the operating design domain makes such limits especially explicit; in other domains, comparable boundaries may include population characteristics, environmental conditions, input quality, communication availability, or permitted transaction types. A controller that performs reliably within a defined envelope may offer little guarantee when a case lies outside it.
Human supervisory control requires awareness of where the operating boundary is located and whether the current situation is approaching it. A model may produce an apparently ordinary recommendation for unfamiliar inputs; a process controller may continue executing a valid local optimization after the organization's actual objective has changed. The interface should therefore represent boundary-relevant conditions, the provenance of important observations, and the assumptions supporting the automated operation. This is a design proposition subject to controlled evaluation.
Boundary testing can vary distance from validated conditions, error visibility, and the availability of an alternative control path. Outcomes include recognition that the system is outside scope, selection of a correct fallback, interruption before irreversible effects, and downstream recovery. Researchers should separate detection of an unusual input from deciding what action is warranted. The person may identify an abnormal case yet be unable to act because the control architecture has already committed a consequential action.
Reorientation, Readiness, and Supervisory Latency
Readiness for intervention has cognitive, procedural, and physical components. A person may be physically able to press an override button while lacking an adequate representation of the current process. The first observable movement after an alarm can therefore be a poor indicator of safe takeover. A complete measure includes recognition of the request, reconstruction of relevant state, choice of response, execution, and stabilization.
Weaver, B. W., & DeLucia, P. R (2022) included 48 articles containing 51 experiments. Their findings indicate that engagement in non-driving activities can degrade takeover performance, particularly where those activities compete for resources also required for driving. They describe weak evidence regarding short takeover budgets and found no general benefit for the information-support implementations aggregated in the reviewed studies. These qualifications matter because the existence of a warning does not establish that a warning design works in every task.
Jin, L., Liu, X., Guo, B., Han, Z., Wang, Y., Cao, Y., Yang (2025) synthesized 37 automated-driving studies and reported poorer response and vehicle-control performance during multiple concurrent non-driving tasks, with benefits associated with multimodal takeover requests in the reviewed conditions. The difference in findings about information support reflects different inclusion criteria and operationalizations, not necessarily a direct contradiction. A transferable research strategy reports warning modality, attention state, time budget, and quality of stable control separately and examines their interactions.
Recovery as a Sequence of Decisions
Fault recovery should be analyzed as a chain rather than a binary event. An operator who correctly identifies an automation failure may choose a poor response; an operator who initially applies an effective correction may subsequently lose control because the controller's hidden state remains unchanged. A more complete representation follows the process from first anomaly through verification, diagnosis, prioritization, intervention, feedback, and restoration of a stable operating condition.
The reliability of each transition matters. A well-designed display may improve anomaly detection but fail to communicate whether an override actually changed an actuator's state. A successful emergency stop may prevent immediate harm while leaving the operator unable to resume productive control. Evaluation should record both immediate and downstream outcomes, including the duration and consequences of any degraded state after intervention.
Researchers can introduce controlled faults at different points in the chain. In one condition the automated recommendation is incorrect but detectable; in another it is correct while the implementation mechanism fails; in a third the machine begins recovering automatically just as a human intervenes. These conditions can expose conflicting control inputs and reveal whether the architecture supports coordinated recovery. The analysis must document the system's ground-truth state and the information accessible to the operator at each step. The Bainbridge (1983)90046-8) critique of exceptional human duties remains relevant precisely because recovery is often the most cognitively demanding phase.
Feedback Timing and Causal Observability
An operator can act effectively only if relevant feedback arrives before the decision opportunity closes. Delayed feedback creates an epistemic problem: the person may know that a command was issued without knowing whether its intended effect occurred. In a digital workflow, the system can report that a transaction request was submitted while external services have not yet confirmed completion. In an industrial control loop, a display may show a target value without revealing that the controlled process has failed to reach it.
Supervisory interfaces should distinguish observation, model inference, intended action, attempted execution, and confirmed outcome. These distinctions permit the human to detect discrepancies between the controller's plan and the environment. The scientific hypothesis is that causal observability improves diagnosis and correction under defined information and time constraints. An interface can be tested by varying which of these elements appear and then introducing matched failures.
More feedback is not always more useful. Excessive event logs can obscure the variables that matter during an urgent intervention. The appropriate design question concerns the smallest set of timely, reliable observations that enables accurate state reconstruction and selection of an effective response. This requires task analysis, controlled fault experiments, and analysis of when operators request deeper explanations. Theoretical claims about transparency should be judged by discrimination and action quality rather than presentation volume.
Reversibility and the Architecture of Intervention
Intervention rights differ according to whether a system's actions can be paused, reversed, compensated, or merely investigated afterward. A reversible change may permit a supervisor to allow greater automatic initiative because errors can be corrected at acceptable cost. An irreversible external action may require earlier review, stronger preconditions, or a different allocation of execution authority. These are analytically separable attributes of the task rather than subjective labels attached to artificial intelligence.
The architecture should distinguish a pause that freezes further actions, an emergency stop that prevents immediate hazardous behavior, a veto that blocks a proposed action, and a rollback that restores a previous state. A user may have access to one of these mechanisms but not the others. Their performance should be tested under time pressure and after partial execution, where interface feedback and hidden dependencies become especially important.
A useful experiment changes the reversibility and consequence of the same underlying task while preserving the accuracy of machine advice. Researchers can then measure how perceived stakes, available authority, and action timing change verification and reliance. The result could reveal whether users calibrate decisions according to actual recoverability or overgeneralize from an automation label. In institutional settings, retrospective logs are valuable for accountability, but a log that only documents harm after occurrence cannot substitute for preventive control.
Error Correlation and Complementary Capabilities
Human–machine cooperation can improve outcomes when the participants make different errors and when the arrangement permits one participant to identify the other's weaknesses. Complementarity depends on the joint distribution of errors, not simply on the average accuracy of each participant. Two accurate decision-makers who make similar mistakes may produce little gain from combining their judgments. A less accurate human may still provide valuable oversight on a small, identifiable class of cases where the model fails systematically.
This can be studied with item-level paired evaluations. The researcher records human-only accuracy, machine-only accuracy, and joint-system accuracy on matched cases, then identifies overlap in errors. The next step is to test whether interface cues help users identify the cases where their judgment has additional value. If such cues are ineffective, a design that demands universal verification may impose workload without producing much incremental correction.
The Vaccaro, M., Almaatouq, A., & Malone, T. W (2024) meta-analysis illustrates why the combined configuration must be compared with both independent baselines. It reports gains relative to unaided humans alongside a pooled disadvantage compared with the stronger independent partner. The result reinforces a general methodological point: cooperation should be judged by the performance of the implemented interaction protocol under a precisely defined comparison, rather than by the attractive assumption that two sources of cognition must always outperform one.
Expertise, Training, and Selection Effects
Experience can change the quality of supervision in several ways. Experienced operators may identify anomalies more quickly, possess better representations of the controlled process, and be more willing to challenge implausible recommendations. They may also have stable habits that make unfamiliar automation modes harder to learn. Novices may rely more heavily on machine guidance yet lack the domain knowledge necessary to recognize rare model errors. These possibilities should be treated as hypotheses requiring data for the specific task and population.
Training for supervision should cover normal interaction, failure recognition, mode transitions, escalation, and recovery. An effective curriculum needs opportunities to observe machine errors and practice appropriate intervention, including situations in which the best action is to retain reliance on a correct automated output. Teaching users simply to override more frequently would likely miscalibrate reliance.
Observational workplace comparisons are vulnerable to selection effects. Organizations may assign expert employees to difficult cases or high-risk systems, creating associations between expertise, automation exposure, and errors that do not directly reveal causality. Random assignment or carefully controlled matched designs can help distinguish the effects of training from the effects of task allocation. Longitudinal follow-up remains necessary to establish whether competence persists during prolonged automated operation.
Human Supervisory Control in Distributed Teams
In large work systems, supervisory control may be divided among several persons rather than concentrated in one operator. A frontline worker notices an anomaly, a specialist interprets it, a supervisor authorizes a response, and another team executes a corrective action. The path through these roles can be appropriate for complex decisions, but it creates additional opportunities for information loss, delay, and mistaken assumptions about responsibility.
A team-level allocation map should identify the recipient of every alert, the source of every decision, which role has authority to stop the process, and what happens when someone is unavailable. Communication protocols must include the status of automated operations and the point at which human ownership changes. Researchers can compare centralized and distributed arrangements using controlled team simulations, time-stamped task traces, and performance under communication disruption.
Human supervisory control is therefore partly an organizational design problem. The person's cognitive capabilities remain important, but the availability of relevant information and effective authority is mediated by the workflow. Measuring a single user's confidence cannot establish that the team as a whole can prevent or recover from failures. Institutional accountability should be tested by tracing where information was available, when intervention was feasible, and which role could have changed the outcome.
Fairness, Asymmetric Errors, and the Distribution of Harm
The impact of automated decisions is often asymmetric. A false-positive flag and a false-negative omission can affect people differently; a harsh grading recommendation and a lenient recommendation have distinct consequences. Evaluating supervisory control with one aggregate accuracy score can hide failures concentrated in particular categories of cases or affected populations.
Goulas, S., Megalokonomou, R., & Sotirakopoulos, P (2026) experimentally varied whether an inaccurate grading recommendation was attributed to a person or an AI system, independently manipulating whether its error was harsh or lenient. The reported fairness gap was larger under the AI label in the harsh condition, while the difference in the lenient condition was not statistically distinguishable. This demonstrates a context-specific interaction between source attribution and error direction. It motivates evaluation of which errors reviewers correct, rather than only whether they intervene at all.
Research should define outcome categories that correspond to the task's actual consequences, compare correction rates across those categories, and report how oversight changes the final distribution of harm. The authority to challenge a system should also be evaluated for users with different resources, knowledge, and institutional power. Formal equal access to a review button cannot by itself establish equal capacity to secure a correction.
Independent Replication and Transport Across Domains
General claims about automation frequently travel from aviation and simulated driving into clinical support, finance, education, and artificial-agent workflows. Such translation is plausible when the systems share a causal structure, but the differences can be decisive. A driver has seconds to regain physical control; a clinician may have minutes to examine evidence; a software supervisor may be able to freeze execution pending review. The information needed and the cost of a mistaken decision vary accordingly.
The scientific task is to identify which mechanisms transport. Attention diverted from an automated task may impair readiness across many domains, whereas a specific warning modality that works in a driving simulator may be poorly suited to a busy hospital. Trust effects may depend on the prestige of the automated source and the professional context. An empirical replication should therefore preserve the hypothesized mechanism while measuring the task characteristics that differ.
Evidence quality improves when researchers report independent replications, preregistered primary outcomes, variation in participants, realistic failure rates, and field follow-up. Null results and unsuccessful transfer are scientifically informative because they locate the boundary of a claim. The contrasting driving meta-analytic findings about information support illustrate why researchers must inspect outcome definitions, study samples, and intervention designs before asserting a universal design rule.
A Reproducible Supervisory-Control Evaluation Protocol
An evaluation protocol can organize the field into six linked decisions. First, specify the operating goal, hazardous outcomes, and full task structure. Second, map information acquisition, analysis, action selection, and implementation to named human and machine roles. Third, identify the information and control rights available to each role at each operational phase. Fourth, define representative routine, degraded, and failure conditions. Fifth, predefine observable human and system outcomes. Sixth, compare alternative configurations under a design that supports causal inference.
The outcome set should include routine accuracy or productivity, detection rate, false-alarm rate, fault recognition latency, diagnostic accuracy, correct acceptance and rejection of machine advice, first intervention, final stabilization, and unintended consequences. Situation awareness and workload may help explain outcomes but should not displace primary behavioral measures. Where rare events require simulation, the protocol must identify limitations to real-world generalization.
Analysis should preserve the distinction between human-only, machine-only, and combined arrangements whenever these baselines are feasible. It should report uncertainty around meaningful effect estimates, document operator expertise and preparation, and describe all system modes and exception paths. This protocol is a methodological synthesis inferred from the reviewed literature; its full effectiveness as a package remains open to empirical evaluation. Its value is to make the actual object of control scientifically inspectable rather than treating nominal human involvement as sufficient evidence.
Conclusion
Automation function allocation reorganizes the relationship between cognition, authority, and action. The foundational psychological problem is whether the human retains the information, time, competence, and practical permission required for effective supervision when automated systems take over routine work. Human control becomes empirically meaningful through demonstrated recognition, evaluation, intervention, and recovery.
The research literature shows that automation can improve routine performance while making failure supervision more demanding. Situation awareness, mode awareness, attention, reliance calibration, and the quality of handovers help explain this trade-off. The most reliable scientific conclusions remain conditional on which functions are automated, how authority moves, what failures are possible, and how operators participate.
In the Artificial Era, intelligent and increasingly agentic systems expand the range of functions that can be delegated and the variety of decisions that can be initiated by artificial processes. The discipline of human factors supplies an established empirical method for evaluating these arrangements. A sound allocation is the one whose whole human–machine configuration can be shown to perform effectively across the operational conditions and consequences that matter.
References
- Atanassova, I., Khan, H., & Khan, Z. (2026). When should your team override AI? A trust calibration routine. Academy of Management Perspectives. https://doi.org/10.5465/amp.2025.0019
- Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775–779. https://doi.org/10.1016/0005-1098(83)90046-8
- 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
- Endsley, M. R., & Kaber, D. B. (1999). Level of automation effects on performance, situation awareness and workload in a dynamic control task. Ergonomics, 42(3), 462–492. https://doi.org/10.1080/001401399185595
- Endsley, M. R., & Kiris, E. O. (1995). The out-of-the-loop performance problem and level of control in automation. Human Factors, 37(2), 381–394. https://doi.org/10.1518/001872095779064555
- Goulas, S., Megalokonomou, R., & Sotirakopoulos, P. (2026). Why do experts miss AI's errors? Evidence from a randomized labeling experiment. PNAS Nexus, 5(6), pgag146. https://doi.org/10.1093/pnasnexus/pgag146
- Hart, S. G., & Staveland, L. E. (1988). Development of NASA-TLX (Task Load Index): Results of empirical and theoretical research. In P. A. Hancock & N. Meshkati (Eds.), Human mental workload (pp. 139–183). North-Holland. https://doi.org/10.1016/S0166-4115(08)62386-9
- 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
- International Organization for Standardization. (2019). Ergonomics of human-system interaction—Part 210: Human-centred design for interactive systems (ISO 9241-210:2019). https://www.iso.org/standard/77520.html
- Jin, L., Liu, X., Guo, B., Han, Z., Wang, Y., Cao, Y., Yang, X., & Shi, J. (2025). Impact of non-driving related task types, request modalities, and automation on driver takeover: A meta-analysis. Safety Science, 181, 106704. https://doi.org/10.1016/j.ssci.2024.106704
- Kaber, D. B., & Endsley, M. R. (2004). The effects of level of automation and adaptive automation on human performance, situation awareness and workload in a dynamic control task. Theoretical Issues in Ergonomics Science, 5(2), 113–153. https://doi.org/10.1080/1463922021000054335
- 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
- Merat, N., Jamson, A. H., Lai, F. C. H., Daly, M., & Carsten, O. M. J. (2014). Transition to manual: Driver behaviour when resuming control from a highly automated vehicle. Transportation Research Part F: Traffic Psychology and Behaviour, 27, 274–282. https://doi.org/10.1016/j.trf.2014.09.005
- 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
- Muir, B. M., & Moray, N. (1996). Trust in automation. Part II. Experimental studies of trust and human intervention in a process control simulation. Ergonomics, 39(3), 429–460. https://doi.org/10.1080/00140139608964474
- National Institute of Standards and Technology. (2023). Artificial intelligence risk management framework (AI RMF 1.0) (NIST AI 100-1). https://doi.org/10.6028/NIST.AI.100-1
- Norman, D. A. (1990). The 'problem' with automation: Inappropriate feedback and interaction, not 'over-automation'. Philosophical Transactions of the Royal Society of London. Series B, Biological Sciences, 327(1241), 585–593. https://doi.org/10.1098/rstb.1990.0101
- Onnasch, L., Wickens, C. D., Li, H., & Manzey, D. (2014). Human performance consequences of stages and levels of automation: An integrated meta-analysis. Human Factors, 56(3), 476–488. https://doi.org/10.1177/0018720813501549
- Parasuraman, R., & Manzey, D. H. (2010). Complacency and bias in human use of automation: An attentional integration. Human Factors, 52(3), 381–410. https://doi.org/10.1177/0018720810376055
- Parasuraman, R., Molloy, R., & Singh, I. L. (1993). Performance consequences of automation-induced 'complacency'. The International Journal of Aviation Psychology, 3(1), 1–23. https://doi.org/10.1207/s15327108ijap0301_1
- 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
- SAE International. (2026). Taxonomy and definitions for terms related to driving automation systems for on-road motor vehicles (J3016_202609). https://saemobilus.sae.org/topics/electrical-electronics-and-avionics/automation/driving-automation/level-3-conditional-driving-automation
- Sarter, N. B., & Woods, D. D. (1997). Team play with a powerful and independent agent: Operational experiences and automation surprises on the Airbus A-320. Human Factors, 39(4), 553–569. https://doi.org/10.1518/001872097778667997
- Sheridan, T. B. (1992). Telerobotics, automation, and human supervisory control. MIT Press. https://mitpress.mit.edu/9780262515474/telerobotics-automation-and-human-supervisory-control/
- Sheridan, T. B., & Verplank, W. L. (1978). Human and computer control of undersea teleoperators (Technical Report ADA057655). Massachusetts Institute of Technology. https://ntrl.ntis.gov/NTRL/dashboard/searchResults/titleDetail/ADA057655.xhtml
- 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
- 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, 2293–2303. https://doi.org/10.1038/s41562-024-02024-1
- Weaver, B. W., & DeLucia, P. R. (2022). A systematic review and meta-analysis of takeover performance during conditionally automated driving. Human Factors, 64(7), 1227–1260. https://doi.org/10.1177/0018720820976476
- Wickens, C. D. (2008). Multiple resources and mental workload. Human Factors, 50(3), 449–455. https://doi.org/10.1518/001872008X288394