RESEARCH DOSSIER
Automation-Related Skill Retention and Human Competence Continuity
Abstract. Automation-related skill retention and human competence continuity concern the maintenance of cognitive, perceptual, procedural, and supervisory capabilities as automated systems take over opportunities for practice. This research monograph evaluates historical models, longitudinal skill-retention evidence, human–automation studies, flight-deck simulations, clinical AI findings, learning theories, monitoring, metacognition, system recovery, and testable maintenance strategies. It distinguishes aided productivity from durable independent proficiency and identifies requirements for causal research, skill-transfer testing, and organizational training.
Keywords: automation-related skill retention; human competence continuity; skill decay; human factors; supervisory control; out-of-the-loop performance; deskilling; automation bias; situation awareness; cognitive offloading; procedural learning; training transfer; failure recovery; AI-assisted professional work; Artificial Era
Introduction: The Problem of Competence Continuity
Automation changes the conditions under which human competence is acquired, exercised, retained, evaluated, and renewed. In highly automated work, successful task completion can coexist with diminished opportunities to perform the cognitive and procedural operations on which independent task performance depends. A person may remain involved in a workflow, inspect its outputs, and assume responsibility for its consequences while the system performs the diagnostic, navigational, interpretive, or motor operations previously used to maintain expertise. The central question of this monograph is how human competence persists across such changes in the distribution of work.
Automation-related skill retention denotes the persistence of task-relevant knowledge, perception, judgment, and action capacities following a period in which automated systems have performed or supported operations that a human formerly executed. Human competence continuity denotes the maintained ability to meet justified performance standards across changing configurations of human and automated work, including ordinary assisted operation, independent execution, supervision, unfamiliar conditions, and recovery from system failure. These terms designate complementary research objects. Retention refers to changes within a repertoire of learned capabilities; continuity describes their operational availability at the times and in the circumstances when an organization or person depends on them.
The problem entered modern human-factors research through the tension between improved routine performance and the requirements of rare intervention. Bainbridge (1983)90046-8) showed why automation may redistribute difficult work toward human operators at the very moment that normal system operation reduces their practice. The phenomenon has since been investigated through experimental automation failures, pilot-simulator studies, work-domain analysis, training research, and quantitative syntheses of skill retention. Contemporary AI assistance extends the issue to reasoning, professional judgment, and perceptual expertise. The present work is a critical narrative scientific synthesis of published studies and methodological literature; it introduces no original experiment or pooled effect estimate.
The research questions organize the investigation at five levels: which constituents of human expertise are threatened by nonuse; which kinds of automation reduce or transform practice; how competence changes with time and assisted exposure; how changes are measured causally and psychometrically; and which evidence-based training and system-design approaches support continuing human capability. These questions require analysis of both the individual learner and the sociotechnical organization that assigns work, produces feedback, and controls access to practice.
Scientific Definitions and the Unit of Analysis
A skill is a learned capability that supports accurate, timely, and context-appropriate performance of a class of tasks. Complex competence is multidimensional: it includes declarative understanding, procedures, perceptual discrimination, situation assessment, selection of action, coordination of movement, monitoring of consequences, and adjustment when familiar routines cease to apply. Expert performance additionally depends on task-specific representations of meaningful cues, sufficient opportunities for corrective feedback, and experience with the conditions under which rules hold. The analysis of expertise by Kahneman, D., & Klein, G (2009) is relevant because it connects skilled intuition to environmental regularity and opportunities to learn those regularities.
Skill retention is observed through performance on a specified task after a retention interval, interpreted relative to a valid baseline and to measurement error. A measured decline in a delayed test is not automatically permanent forgetting: it may reflect temporary retrieval failure, altered motivation, changes in context, fatigue, response strategy, or adaptation to a new interface. Equally, preserved immediate performance may conceal difficulty recovering competence under time pressure. The distinction between acquisition performance and durable learning, developed by Schmidt, R. A., & Bjork, R. A (1992) and synthesized by Soderstrom, N. C., & Bjork, R. A (2015), requires delayed unassisted tests rather than judgments based on smooth assisted execution.
Deskilling denotes a reduction in task-relevant independent competence or in opportunities and institutional support for its development and maintenance. It is used in different ways across occupational scholarship. An individual can show a measurable decrement on a standardized task; a professional role can lose discretion or occasions for complex judgment even if a particular worker's latent skill is not directly assessed; and an entire organization can lose redundancy when few employees retain the capacity to intervene. These are distinct empirical claims requiring different evidence. The term should therefore be accompanied by an operational definition and the scale at which change is observed.
Competence continuity includes the ability to recognize when automation is inappropriate, obtain independent evidence, recover or reconfigure manual action, diagnose discrepancies, and reconstruct a safe task state. It does not imply that every person must retain every historically manual operation indefinitely. The relevant repertoire is determined by foreseeable failures, safety responsibilities, the availability of alternative support, and the expected future environment. A system may responsibly retire obsolete skills while deliberately maintaining capabilities that remain necessary for resilience, verification, and adaptation.
Automation as a Change in Practice Architecture
Automation varies in the functions it assumes. Parasuraman, R., Sheridan, T. B., & Wickens, C. D (2000) distinguish information acquisition, information analysis, decision and action selection, and action implementation. Their taxonomy directs attention to the location of human involvement rather than describing a system through the single label automated. Two workplaces may use similarly accurate software yet expose workers to different learning conditions when one presents evidence for human interpretation and another presents a recommended decision requiring approval. A third may implement the action without an opportunity for human rehearsal.
The learning consequences of automation therefore depend on function allocation. If a navigation display automatically calculates position, the operator may still infer location using instruments or may simply accept the displayed result. If a clinical assistant highlights a possible abnormality, the user may integrate the highlighting with independent perceptual search or may skip the search entirely. If an AI coding assistant completes software routines, the developer may review, test, and explain the proposed implementation or may accept a result without reconstructing its underlying logic. These alternatives create different exposure to error signals, search strategies, action sequences, and explanatory feedback.
A task's observable completion rate is a system-level outcome. It can reflect human expertise, automation capability, communication quality, or compensatory work by colleagues. The individual contribution to the outcome becomes identifiable only when the design includes conditions in which automation support is varied, withdrawn, degraded, or rendered misleading. This is particularly important in high-performing systems: uniformly successful output can mask a widening gap between what the combined human–machine arrangement can do and what its human component can independently accomplish.
Automation also changes the distribution of rare and routine events. Repetition ordinarily supplies the occasions on which a novice becomes reliable and an expert calibrates judgment. A reliably automated process may remove many low-complexity repetitions while preserving the expectation that humans handle complex exceptions. The events retained for human attention can be more time-sensitive, ambiguous, and emotionally demanding than those used during initial training. Bainbridge's (1983)90046-8) analysis remains influential because it locates the contradiction within work design: operators can be held accountable for interventions while the conditions required for intervention competence are progressively reduced.
Historical Foundations: From Control Engineering to Human Factors
Industrial automation research in the twentieth century examined the assignment of control between mechanical or computational systems and operators. Early accounts focused on the reliable regulation of physical processes and on the human operator's residual role in monitoring, diagnosis, and abnormal-event management. Bainbridge's formulation crystallized the observation that a successful transfer of routine control does not eliminate the need for understanding the process being controlled. It can make operator understanding more difficult to acquire and maintain when everyday work offers fewer occasions for observing consequential relationships.
The expansion of computational decision aids shifted research from manual workload alone to the quality of human decisions made with system recommendations. Skitka, L. J., Mosier, K. L., & Burdick, M (1999) used simulated aviation monitoring to distinguish errors of omission when an aid failed to alert users and errors of commission when users followed inaccurate recommendations despite available contradictory evidence. Their findings helped establish automation bias as a phenomenon requiring attention to the relationship between recommendation, independent evidence, and user response. A mistake arising during assistance is not itself proof of skill loss across time; it is evidence about reliance and control under a particular task configuration.
The out-of-the-loop program then connected automation with recovery performance. In an experimental navigation task, Endsley, M. R., & Kiris, E. O (1995) investigated how operator involvement affected situation awareness and intervention following expert-system failure. Their findings supported the importance of active processing and the quality of the operator's internal representation of the evolving task. A key historical development was the identification of several possible mechanisms behind poor takeovers: reduced access to relevant information, lower situational comprehension, disengagement from active operations, and degradation in component skills. These mechanisms need separate empirical measures.
The design of higher automation levels became a systematic experimental question. Kaber, D. B., & Endsley, M. R (2004) compared levels and adaptive allocation within a dynamic control task, examining performance, awareness, and workload. Later, Onnasch, L., Wickens, C. D., Li, H., & Manzey, D (2014) integrated 18 experiments on stages and degrees of automation. The development from conceptual warnings to controlled experiments and evidence synthesis is essential: the consequences of automation are moderated by what is automated, how control is allocated, and which performance endpoint is evaluated.
Human Capability as a Multicomponent System
Complex work requires coordination among distinct skills. Perceptual skills include recognizing target features and discriminating signal from noise. Cognitive skills include maintaining task models, estimating future states, prioritizing possibilities, interpreting evidence, selecting responses, and recognizing anomalies. Procedural skills include executing action sequences and controlling timing. Metacognitive skills include judging whether a decision is sufficiently supported and whether assistance should be questioned. Social competencies include communicating uncertainty, transferring control, and maintaining shared understanding during incidents.
These components need not decay in parallel. Motor routines can remain familiar while planning or situation reconstruction becomes fragile, particularly when routine automation continues to present recognizable displays while eliminating occasions for independent inference. In Casner, S. M., Geven, R. W., Recker, M. P., & Schooler, J. W (2014) Boeing 747-400 simulator study, sixteen airline pilots largely retained instrument scanning and manual control, whereas some cognitive operations needed for manual flight were more problematic. The result provides a concrete caution against equating automation exposure with a uniform loss of all manual skills.
Competence has an additional relational structure. A human operator may know how to achieve a goal only when the software exposes intermediate states, supplies cues, or preserves certain controls. A different interface can remove those affordances without changing nominal job responsibilities. Reliable competence continuity therefore requires evaluating person, task, tools, team, and environment together. Apparent individual failure can arise from poor information access, an impossible takeover interval, ambiguous modes, or training that never represented the operational challenge.
In occupational contexts the unit of concern may be a team or an organization. Redundancy can be maintained when several people possess overlapping qualifications and practice critical operations; it can decline when few workers receive meaningful exposure to those operations. Institutional loss of expertise should be distinguished from the forgetting curve of any one worker. Personnel turnover, specialization, shifts in workflow, changes in certification, and unequal access to supervised practice can all alter the organization's future ability to recover from technological disruption.
Performance, Learning, and Retention
Learning is inferred from durable change, whereas performance is observable in a particular test situation. Soderstrom, N. C., & Bjork, R. A (2015) review evidence that conditions producing rapid improvement during practice may yield weaker long-term retention or transfer. This applies directly to assisted systems: an AI-generated hint can raise immediate accuracy while reducing the amount of retrieval, discrimination, planning, or error correction performed by the learner. The immediate benefit and the delayed effect require separate measurement.
The same principle explains why error-free practice is an ambiguous training target. Scaffolding, automation, and feedback can be valuable during difficult acquisition; however, their effectiveness depends on the intended future performance context. If the final requirement is independent judgment under degraded system conditions, training must test whether the person can independently reconstruct the task after guidance has been removed. If the future role is permanent tool-mediated supervision, training should still test whether the operator can identify failure modes and exercise justified control.
Retention intervals require explicit specification. A post-training assessment immediately after a session evaluates a different property from performance following weeks or months of nonuse. The comparison should record initial proficiency, the number of subsequent encounters, the proportion of independent task execution, exposure to similar tasks, and conditions of the delayed test. Changes in organizational standards or the software interface can alter measured performance even when underlying knowledge remains stable.
The distinction is especially consequential for interpreting the productivity benefits of AI. Productivity studies can establish changes in completion time or output quality under access to assistance, but they do not automatically identify effects on the underlying human repertoire. Research on independent competence needs pre-assistance assessment, repeated measures of assisted work, delayed unaided evaluation, and tests of transfer to novel problems. The scientific outcome should be the capability that remains available when a human must act, not merely the output achieved while technology works as intended.
Mechanisms of Skill Nonuse and Decay
Nonuse removes opportunities to retrieve task knowledge, rehearse action procedures, monitor errors, and update mental representations. The rate and form of retention depend on initial learning, task organization, the similarity between training and testing conditions, and subsequent practice. The quantitative review by Arthur, W., Jr., Bennett, W., Jr., Stanush, P. L., & McNelly (1998) demonstrated substantial variation in skill decay across tasks and testing conditions. Its moderator findings make it inappropriate to assume one universal decay curve for all human competencies.
A more recent meta-analysis by Tatel, C. E., & Ackerman, P. L (2025) synthesized procedural-skill retention across 457 reports and examined retention interval as a continuous predictor. It found that the magnitude of decline varied across accuracy, speed, and mixed performance outcomes and with task characteristics and intermittent practice. Its time-based estimates describe pooled research settings under nonuse; they are not calibrated deadlines for a particular profession. Automation research must additionally measure actual use of constituent skills rather than inferring nonuse solely from time spent with software.
Cognitive nonuse can be partially concealed by continued participation in the same broader task. An operator may appear experienced because the workstation, setting, and professional identity remain unchanged while crucial inferences are delegated to software. Such concealed changes create a measurement challenge: gross time-on-task, seniority, or service volume may become poor proxies for continuing independent expertise. Learning histories should record which subtasks were executed cognitively by the person, which were recommended or completed by the system, and what feedback the person received.
Skill loss is also compatible with skill reorganization. Repeated supervision can strengthen anomaly monitoring, verification, coordination, or prompt evaluation while weakening unaided search or calculation. A balanced study must detect newly acquired competencies and declines in older ones rather than treating them as a single dimension. Whether the new configuration is adequate depends on the reliability of the technology and on the real residual tasks assigned to humans.
Procedural and Declarative Knowledge Under Assisted Work
Procedural knowledge supports the execution of sequences in which perception, action, and corrective feedback become efficiently coordinated. Declarative knowledge supports explanations, facts, rules, and relationships that can be articulated and brought to bear when familiar routines fail. In skilled work these forms support one another: a pilot's instrument scan is organized by an understanding of flight dynamics, while repeated flying changes the speed with which meaningful instrument configurations are recognized. Automation may reduce opportunities to rehearse procedures without erasing underlying facts, or it may reduce the need to reason through causal relationships while preserving familiar physical movements.
The relevant retention measure consequently depends on the missing operation. An individual who can describe a procedure but cannot execute it accurately under time pressure has retained one portion of competence and lost operational fluency in another. Conversely, a person who can complete a familiar action sequence may struggle to diagnose a novel failure that requires explicit understanding. Casner, S. M., Geven, R. W., Recker, M. P., & Schooler, J. W (2014) provide a useful applied instance: preserved manual aircraft control in a simulator did not imply equivalent preservation of every cognitive aspect of flight conducted without automation.
Assistance can be designed to sustain complementary processes. A support system might expose a recommended action together with the environmental cues, constraints, and causal explanations used to derive it. Whether this supports future independent competence remains an empirical question. An explanation read after a recommendation may become a passive confirmation rather than an independently generated inference. Researchers should therefore compare conditions with and without initial independent answers, delayed explanatory prompts, and unaided transfer tests. The appropriate experimental outcome concerns what the learner can reconstruct, discriminate, and execute later.
Mental Models, Situation Awareness, and Mode Understanding
A mental model represents how a work system functions and how its state will change under action or disturbance. Situation awareness refers to an operator's perception of relevant elements, comprehension of their significance, and projection of likely future states. Endsley, M. R., & Kiris, E. O (1995) examined how levels of control changed operator awareness and performance after automation failure. Their work makes explicit the relation between participating in a task and maintaining the internal information required to take over that task.
Automated interfaces often make task outcomes highly visible while reducing access to intermediate causal processes. The operator can know that a controller has maintained a target without continuously knowing which constraints shaped the controller's action. A takeover becomes demanding when responsibility abruptly moves back to the person. Time must be spent locating the system's present mode, reconstructing its immediate history, identifying the cause of deviation, and selecting a recovery strategy. Each of these operations can fail independently of an individual's motor or diagnostic skill.
Mode awareness therefore belongs in a competence-continuity assessment. A system may offer several operating modes, each with different constraints and transitions. Users who have practiced only normal automation may learn the surface mapping between buttons and outcomes without acquiring a reliable model of what the system does after partial failures. Sarter, N. B., & Woods, D. D (1997) described automation surprises in the operational experience of advanced flight-deck systems. Such observations underscore that understanding automation is itself a learned professional competence that must be trained and measured.
Longitudinal assessment can separate memory for system modes from the skill needed to recover the underlying task. Tests can ask operators to predict system behavior under changed conditions, identify control authority, and explain discrepancies in instrument readings before presenting the option to intervene. Delayed scenario tests should then determine whether earlier assisted experience preserves or erodes these capabilities. A single subjective awareness rating provides limited evidence because participants may feel informed while overlooking essential state changes.
Attention, Monitoring, and the Out-of-the-Loop Problem
Routine active performance continuously binds attention to consequences. Monitoring automation changes the temporal distribution of attention: the user must remain prepared for uncommon errors during long intervals in which intervention is unnecessary. Monitoring itself is demanding, but it supplies different information from active operation. The evidence reviewed by Parasuraman, R., & Manzey, D. H (2010) links automation-related complacency and bias with failures to allocate attention to critical information when the automated system appears reliable.
The out-of-the-loop problem describes difficulty reestablishing control after prolonged automation involvement. It includes at least three separable consequences. The person may have an incomplete picture of the current system state, may have become less efficient in component task operations, or may lack an effective plan for managing the transition from monitoring to action. Recovery demands can therefore increase even when underlying skill scores remain stable. An experiment that measures only time to intervene cannot identify which mechanism caused the delay.
Research on levels and stages of automation identifies trade-offs in this architecture. In their integrated meta-analysis, Onnasch, L., Wickens, C. D., Li, H., & Manzey, D (2014) found that higher automation could improve routine performance while imposing costs in failures of the automated function, especially when humans needed to regain task control. The effects depend on stages and levels, so statements that automation inevitably destroys human skills exceed the evidence. The more precise conclusion is that high routine performance and resilient human intervention are different outcomes requiring separate design attention.
For continuity, monitoring tasks should make critical anomalies interpretable and permit periodic active engagement. Interventions must be studied under realistic alarm rates, workload, time constraints, and failure modes. If a laboratory system signals failure clearly in advance, it may test the operator's response to a known signal rather than the ability to notice and explain an unexpected failure. Ecological validity is particularly important in rare-event domains, where the everyday absence of error can make apparent monitoring success difficult to distinguish from passive trust.
Automation Bias and Verification Behavior
Automation bias is a tendency to accept, follow, or fail to challenge automated information in circumstances where independent evidence warrants further examination. Skitka, L. J., Mosier, K. L., & Burdick, M (1999) distinguished omission errors, in which users failed to act when an aid omitted a necessary warning, from commission errors, in which users acted on an incorrect recommendation. These are errors in the human–automation decision process; they can arise without any preceding decline in the user's long-term knowledge.
Verification skill requires the ability to identify what must be checked, obtain independent evidence, interpret discrepancies, and revise an initial belief. These operations may receive limited practice if most automated recommendations are accurate. A system that displays a clear recommendation first can create anchoring and can reduce the perceived value of fresh inspection. Some tasks also impose verification costs greater than the immediate cost of acting on an automated suggestion. The operator's behavior is shaped by the system's incentive structure as well as by personal diligence.
The systematic review by Goddard, K., Roudsari, A., & Wyatt, J. C (2012) examined automation bias in computerized decision support and the factors affecting its frequency. The literature includes studies with different designs, definitions, and outcomes. Because errors of reliance are task- and interface-dependent, an observed automation-bias effect should be interpreted alongside the accuracy of the aid, availability of contradictory cues, base rates of error, and the user's previous experience with the system.
Continuity research can measure verification as an independent capability. A useful protocol assesses unassisted detection of anomalies, aided judgments with both accurate and misleading outputs, frequency and quality of independent checks, and delayed transfer to unfamiliar discrepancies. Merely increasing the number of confirmations may produce ritualized clicking rather than substantive examination. Good verification requires diagnostic action that is sensitive to a plausible failure, with a testable relation between the check and the decision it is supposed to support.
Trust Calibration, Reliance, and Error Histories
Trust affects whether an operator delegates work, monitors the aid, and intervenes when uncertainty becomes salient. Lee, J. D., & See, K. A (2004) established a widely used framework for appropriate reliance, linking the user's beliefs about system performance, processes, and purpose with the decision context. Calibration is achieved when reliance tracks system capabilities and limitations across relevant circumstances. Appropriate reliance is task-dependent because the costs of trusting and distrusting a system vary with its role.
Trust and skill can influence each other over time. Greater reliance can decrease direct task practice; reduced practice can make independent completion more difficult; that difficulty can subsequently encourage further reliance. This feedback loop is a plausible model of increasing dependency, but each arrow requires empirical evidence. A longitudinal design should record actual delegation, unassisted competence, opportunities for practice, error exposure, and changes in system performance. Cross-sectional association between trust and low self-rated skill cannot determine the direction of the mechanism.
Error histories are especially informative. Users may encounter a long run of correct recommendations followed by a rare but consequential fault. If they have repeatedly seen errors and received specific feedback about their causes, they may have learned when to challenge the system; if they have never encountered a failure, their reliance might reflect reasonable experience rather than a stable dispositional trait. Designing authentic failure encounters without creating unacceptable operational risk is a central methodological challenge for high-stakes domains.
The resulting measurement should distinguish attitude from behavior. Trust scales can capture beliefs and expectations, but the actual frequency of independent checks, choices to override, response latency, and error recovery provide behavioral evidence about reliance. A person may report low trust yet habitually accept suggestions under time pressure. Another may report strong trust and still verify exceptional cases. Competence continuity concerns the behavioral repertoire available when the situation warrants independent action, and thus requires both kinds of measurement.
Cognitive Offloading and Metacognitive Self-Assessment
Cognitive offloading occurs when people use actions or external systems to reduce the demands placed on internal cognitive processes. Risko, E. F., & Gilbert, S. J (2016) synthesized experimental research on external reminders, environmental restructuring, and technology-mediated cognition. Offloading can increase reliability and free resources for more complex work. It can also change the frequency with which memory, calculation, planning, or problem-solving procedures are rehearsed. The presence of external support alone does not establish long-term skill deterioration.
Metacognitive calibration refers to the correspondence between a person's confidence in competence and actual performance under specified conditions. Assistance can create ambiguous feedback about personal ability. When a task succeeds with software support, the user may attribute part of the success to their own understanding, especially when the system's contribution is not visible or easily decomposed. A study of apparent learning should therefore collect predictions of unaided performance and compare those predictions with delayed independent tests.
The conceptual analysis by Macnamara, B. N., Berber, I., Çavuşoğlu, M. C., Krupinski, E (2024) raises the possibility that AI-supported performance can alter skill development and retention without the user recognizing the change. Its status as a conceptual and theoretical argument is important. It motivates direct experiments; it does not by itself establish a population-level causal effect of generative AI on professional expertise. Separating hypotheses from demonstrated effects preserves the explanatory value of the argument.
Researchers can operationalize metacognitive monitoring through confidence judgments attached to individual responses, global predictions of later independent performance, and forecasts of performance after aid withdrawal. Confidence should be compared with observed accuracy and with the ability to explain failures. Such measures can reveal misplaced confidence, excessive uncertainty, and well-calibrated knowledge. Their interpretation must consider whether the test itself provides cues that were absent during routine assisted work.
Failure Recovery as a Distinct Competence
Failure recovery is an organized sequence rather than a single button press. It may require detecting a discrepancy, classifying the fault, establishing which information remains reliable, prioritizing hazards, selecting an alternate procedure, acting under time pressure, and evaluating whether control has been restored. Each component can be taught and measured separately. A person who performs a normal manual task accurately may still struggle to recover from a complex automation fault because normal execution and recovery have different cue structures and decision demands.
Recovery also depends on the amount of information retained during automation. If the system's intermediate reasoning is hidden, operators may need time to reconstruct a history that active participation would have supplied. If the design permits reversible decisions, clear control transfer, and interpretable status displays, the same residual human skill can support better recovery. This shows why competence continuity is jointly determined by personal learning and engineered affordances, not simply by the number of hours spent practicing a manual procedure.
Griffiths, N., Bowden, V., Wee, S., Strickland, L., & Loft (2024) examined whether manual aircraft-conflict discrimination skill predicted intervention when automation failed in a simulated air-traffic-control task. Better baseline manual skill predicted faster and more accurate failure intervention. A substantial amount of variance remained unexplained, however, indicating that manual task skill is only one component of successful supervision. The study supports including manual proficiency in operator assessment while motivating research on attention, state understanding, control-transition planning, and experience with failure scenarios.
A comprehensive recovery evaluation should vary failure timing, diagnostic transparency, rate of misleading alerts, task load, and the opportunity for prior rehearsal. Outcomes should include correct detection, diagnostic accuracy, intervention appropriateness, time to safe state, unnecessary overrides, and downstream harms. Training should then be evaluated on previously unseen failures, because improvement on rehearsed examples can overestimate the capacity to generalize. Retention is demonstrated when these capabilities remain accessible after meaningful intervals of ordinary automation exposure.
Quantitative Evidence on Skill Retention
The general skill-retention literature provides the most rigorous broad evidence for what happens when practice ceases, although it does not isolate automation as the cause of nonuse. In the quantitative review of Arthur, W., Jr., Bennett, W., Jr., Stanush, P. L., & McNelly (1998), retention varied with the nature of the skill and with conditions of acquisition and later testing. The review established the importance of skill type, proficiency, and retention interval and helped replace a uniform account of forgetting with moderator-sensitive research. Its source studies were heterogeneous and were not designed to estimate one causal effect of algorithmic assistance.
Tatel, C. E., & Ackerman, P. L (2025) conducted a more extensive meta-analysis of procedural skill retention, including 1,344 effect sizes from 457 reports. Their random-effects meta-regressions examined retention intervals continuously and found larger standardized declines with longer intervals. The authors reported mean declines that increased by approximately 0.08 standard-deviation units per month for accuracy-based outcomes and 0.06 for speed-based and mixed measures. This is evidence about the average relation between nonuse interval and measured procedural performance, not a uniform biological or organizational law governing all kinds of human judgment.
A precise inference must take account of the varied definitions of proficiency, different task families, and the fact that a skill may be practiced intermittently while appearing unused in an organizational record. A person could retain relevant component skills through related work even when a particular procedure has not been performed directly. Conversely, routine software interaction could count as nominal experience while providing little retrieval or independent problem solving. The independent competence of automation users therefore needs measurement of task-specific practice exposure and delayed transfer, not extrapolation from aggregate curves alone.
The two quantitative syntheses support a cautious general proposition: opportunities for practice, skill characteristics, and the passage of time are important predictors of retention. They do not prove that every form of automation causes deskilling or identify a single safe refresher interval. Practical schedules must be calibrated against local tasks, safety requirements, individual variation, and longitudinal observations. Where a neglected operation is rare but critical, its significance may depend more on the consequences of failure than on a small average change in a skill test.
Flight-Deck Automation and Manual Flying
Commercial aviation provides a mature empirical setting because automation can be varied in simulators while preserving many flight-deck demands and measuring task performance. The concern is specific: pilots are required to supervise and occasionally recover from automated control in conditions that differ sharply from normal routine flight. Casner, S. M., Geven, R. W., Recker, M. P., & Schooler, J. W (2014) tested sixteen airline pilots in a Boeing 747-400 simulator on manual flight tasks and separated routine physical control from cognitive operations accompanying flight without automation.
The study's central finding was differentiated preservation. Basic instrument scanning and manual control skills remained relatively intact in its tested circumstances, while some cognitive activities, including those needed to manage manually flown conditions, were less robust. Reports of task-unrelated thought during portions of the task offered a possible explanation for disengagement, although they cannot independently identify the full causal chain. The small sample and simulator context constrain generalization to different fleets, pilot populations, retention periods, and abnormal-event conditions.
A related study by Casner, S. M., Geven, R. W., & Williams, K. T (2013) examined the effectiveness of airline pilot training for abnormal events. Abnormal-event training can produce task-specific responses, but the ability to recognize and respond to genuinely novel combinations of cues remains a different standard. This distinction has direct implications for competence continuity. Rehearsed procedures may remain accessible while scenario construction, prioritization, or planning under unanticipated constraints remains weakly tested.
Flight-deck evidence therefore suggests a research design built around skill decomposition. Independent hand-flying proficiency should be evaluated alongside state estimation, navigation planning, energy management, mode awareness, and unexpected failure recovery. Performance should be tested with the automation functioning, degraded, and absent. It is also necessary to distinguish a forced manual task from an authentic takeover following prolonged monitoring; the latter adds a transitional cognitive burden. A training program that preserves one component cannot be assumed to preserve the whole operational capability.
Air Traffic Control and Automation Failure Intervention
Air-traffic-control automation illustrates the relation between manual task proficiency and the ability to supervise an automated function. Griffiths, N., Bowden, V., Wee, S., Strickland, L., & Loft (2024) first assessed participants' manual discrimination of potential aircraft conflicts and subsequently measured their performance supervising automated conflict detection and resolution in simulation. When automation failed, stronger manual discrimination skill was associated with quicker and more accurate intervention.
This evidence is especially relevant because the study examined independent manual skill as a predictor rather than simply assuming that automation experience had changed it. The findings make a case for preserving component knowledge as one contributor to resilient automated operations. At the same time, substantial variance in intervention remained unexplained, and the simulated task was narrower than operational air-traffic-control work. Baseline skill can predict response without demonstrating that an automation system later damaged or preserved that skill.
To demonstrate skill decay caused by automation, a study would need to assign or otherwise credibly compare different automation-exposure histories, measure equivalent baseline proficiency, document task practice, and retest independent skill after a defined interval. A failure intervention score obtained only at the end of the experiment is a useful applied outcome but does not isolate whether a decrement occurred. The distinction between predictive validity and causal identification is one of the most important in this research field.
Air-traffic-control environments also show why competence is distributed. The controller's action is shaped by traffic visualization, procedures, team communication, and authoritative responsibility. A future deployment study should examine these system-level supports alongside individual cognition. Reliable failure response might depend on an expert's internal conflict model, an effective display of automated uncertainty, rapid clarification among team members, or their interaction. Reducing competence continuity to one psychometric ability would omit these dependencies.
Clinical AI and the Measurement of Diagnostic Competence
Clinical AI introduces a related question with direct implications for patient safety: does repeated exposure to automated detection or recommendation alter a clinician's later performance without assistance? The strongest study design would randomize clinicians or clinical units to meaningful exposure conditions, obtain baseline independent competence, and retest on blinded, comparable cases after prolonged practice. It would also evaluate case mix, disease prevalence, software updates, and independent patient outcomes. Most available studies do not meet all of these conditions.
A notable observational finding comes from Budzyń and colleagues (2025)00133-5), who examined adenoma detection during standard non-AI-assisted colonoscopies before and after AI tools had entered routine use at four Polish centers. The analysis involved 1,443 non-AI-assisted examinations, with 795 before and 648 after implementation. Adenoma detection fell from 28.4% to 22.4%, an absolute difference of 6.0 percentage points; the reported 95% confidence interval for the difference was −10.5 to −1.6 percentage points. This constitutes a directly relevant signal about the performance of clinicians after AI exposure.
The interpretation requires care. The comparison was observational and assessed procedures across time rather than assigning clinicians randomly to independently measured retention trajectories. Changes in case mix, practice, institutional routines, and exposure to concurrent training or organizational changes can influence the observed outcome. Adenoma detection is a clinically meaningful behavioral endpoint but is not a pure test of latent visual skill. The authors' result supports a concern that repeated AI exposure might affect unassisted detection; it does not establish irreversible loss of expertise or prove that the entire difference was caused by the AI system.
Clinical fields also differ sharply in the types of competence that algorithms support. A second reader for images, a triage model that rearranges worklists, a transcription system, and a decision aid proposing differential diagnoses have different consequences for clinicians' opportunities to perceive, reason, act, and learn. They may preserve exposure while changing feedback, or remove independent reasoning while increasing the number of cases handled. Generalization requires detailed specification of each function allocation and outcome rather than invoking clinical AI as a single intervention.
A serious clinical research program should include unassisted performance on cases sampled across disease rarity, severity, and ambiguity. It should record patients' actual outcomes when ethical and feasible and prevent misleading model recommendations from affecting care outside approved protocols. Training interventions should be evaluated through both competence tests and clinical system outcomes. Any suggestion about clinical practice must be distinguished from educational design hypotheses until validated by appropriately controlled evidence and professional review.
High-Risk Professions and Refresher Training Evidence
Work in aviation, nuclear operations, emergency response, medicine, and other high-risk domains creates a particular retention problem: many critical responses are used infrequently in ordinary operation, yet failure to execute them correctly carries substantial consequences. The scoping review by Klostermann, M., Conein, S., Felkl, T., & Kluge, A (2022) examined factors potentially reducing skill decay in high-risk industries. Its emphasis on training conditions, task specificity, and refresher strategies illustrates how competence maintenance can be treated as a design responsibility.
The systematic literature review by Vlasblom, J. I. D., Pennings, H. J. M., van der Pal, J., & O (2020) examined competence retention in safety-critical professions and identified variation in the methods and domains studied. The resulting literature is useful for selecting retention measures but provides limited support for one universal schedule applicable across professions. Procedural performance, recognition, diagnostic reasoning, and the integration of teamwork may respond differently to rehearsal frequency and testing.
The critical difference between a training intervention and a validated maintenance program is the existence of delayed, realistic, independent evaluation. A refresher session can temporarily raise confidence or immediate checklist performance without preserving subsequent ability to handle an unannounced abnormal event. Such a program should therefore include spaced assessments, repeated opportunities for corrective feedback, and tasks that require reconstruction rather than recognition of a previously practiced scenario.
In safety-critical settings, the consequences of testing must also be considered. Organizations cannot routinely create dangerous failures merely to study learning. High-fidelity simulation, controlled scenario training, and protected operational exercises make it possible to introduce rare faults without exposing real users or the public to unacceptable risk. Yet simulation fidelity is multidimensional: visual similarity alone does not reproduce workload, consequences, social pressure, organizational communication, and the distribution of rare-event uncertainty. Transfer should be tested against operationally relevant criteria.
Generative AI and Professional Expertise
Generative AI extends automation from well-specified physical or computational routines to open-ended writing, coding, research, analysis, and professional decision support. A single system can propose a plan, generate explanations, perform calculations, draft documents, and critique its own output. This breadth changes the acquisition environment for novices and the maintenance environment for experienced workers. The central scientific questions remain familiar: what operations did the person personally practice, what feedback was available, and what can the person do without the system after repeated reliance?
The conceptual article by Macnamara, B. N., Berber, I., Çavuşoğlu, M. C., Krupinski, E (2024) identifies a potential risk that users may confuse aided performance with independently acquired skill. Their argument is a research hypothesis about skill development, decay, and self-awareness. Demonstrating it in software engineering or scientific writing requires independent skill tests and prospective exposure measures rather than survey evidence alone. Claims that routine access to a generative assistant universally weakens cognition would exceed the present evidence.
A new form of competence may emerge through learning how to formulate requests, evaluate generated hypotheses, test solutions, and manage the system's uncertainty. These skills can be valuable and measurable. Their acquisition does not establish that the previous manual repertoire has been preserved. The correct research strategy evaluates both sets of abilities and specifies which are required for the worker's future responsibilities. A programming professional who can verify code efficiently but cannot reconstruct a fault without the assistant may have gained an important verification skill while retaining a different vulnerability.
Workplace studies should therefore report assisted productivity and independent learning as separate endpoints. Automated output quality, speed, and user satisfaction are not adequate proxies for retained professional judgment. Where a team uses AI in a sustained fashion, training comparisons can include guidance that requires the person to make an independent initial attempt, guidance that reveals intermediate reasoning, and immediate finished answers. Delayed transfer to new tasks can then establish whether each interaction structure supports durable capability.
Conflicting Evidence, Bias, and Limits of Generalization
The research record contains different kinds of claims that should not be collapsed into a single narrative. Quantitative forgetting research demonstrates associations between elapsed nonuse and retention across task families. Human-automation experiments demonstrate effects of varying control allocation on performance and intervention under controlled conditions. Longitudinal observational studies can identify concerning changes after deployment. Conceptual essays articulate potential mechanisms. Each design answers different questions and supports different levels of causal inference.
The flight-deck study by Casner and colleagues, for example, found important domains of manual proficiency remained accessible. The colonoscopy study observed a substantial decline in an independent clinical outcome after AI introduction. Neither finding invalidates the other: populations, tasks, intervention histories, measurement criteria, and organizational contexts differ. The scientifically useful synthesis is to identify which attributes make particular skills vulnerable or robust rather than average superficially comparable outcomes into a claim about all automation.
Potential biases include selective recruitment of experts, reliance on convenient student simulations, incomplete pre-exposure baselines, short retention intervals, unrepresentative failure frequencies, and assessments that resemble the exact training scenarios. In organizational data, adoption of automation may coincide with changes in staffing, incentives, case mix, or reporting. Test conditions can advantage previously trained groups when the interface or manual procedure has changed. The availability of the full research protocol, preregistration where suitable, and sufficient outcome reporting matters to interpretation.
Publication bias is an additional consideration. Studies reporting striking deskilling effects may attract greater attention than careful studies reporting stability or task-specific improvement. Conversely, large industrial deployments with safety or commercial sensitivities may generate little publicly accessible independent evaluation. The literature's effective evidence base is therefore conditioned by what researchers can observe and organizations are willing to report. Researchers should avoid asserting a consensus about long-term professional deskilling without evidence from representative, independently replicated longitudinal designs.
The strongest present conclusion is conditional: reduced opportunities for relevant practice can threaten future independent performance, especially when automation also changes attention, feedback, and recovery demands. The available evidence supports explicit monitoring of those mechanisms and evaluation of carefully designed maintenance programs. It leaves open the magnitude, reversibility, occupational distribution, and causal prevalence of long-term skill loss attributable specifically to modern AI systems.
Operationalizing Independent Human Competence
A scientific account of competence continuity begins by specifying the performance capacity that is expected to remain available. The construct cannot be inferred from job title or from overall output under automation. A defensible assessment defines the task domain, the skills composing it, the operating conditions, the standard of acceptable accuracy, the cost of errors, the dependence on external cues, and the maximum interval permitted for action. Competence becomes measurable through performance on samples of tasks representing these requirements.
For research purposes, a competence-continuity profile can distinguish routine independent execution, assisted execution, monitoring of automation, detection of system error, recovery after failure, and transfer to novel conditions. These are separable test conditions rather than interchangeable indicators. A participant's performance in one condition may predict performance in another but should not be treated as equivalent without empirical validation. The comparison between standard manual performance and unexpected takeover is particularly informative because the second condition adds state reconstruction and control-transfer demands.
The primary outcomes might include sensitivity to task-relevant events, diagnostic accuracy, procedural errors, response latency, recovery time, appropriate intervention, and lasting success in a novel task. Secondary measures can examine mental workload, confidence calibration, eye movements, and understanding of the system. These variables should be chosen because they represent theoretically specified mechanisms, not because an interface makes them easy to log. An overall score can conceal compensating changes in different subskills, so disaggregated outcomes are crucial.
A study should articulate whether competence continuity refers to meeting an absolute standard, preserving a baseline individual level, or maintaining performance within an equivalence margin. These standards answer different questions. A participant may decline from exceptional to adequate proficiency while remaining above a safety threshold; another may show only a small decline but cross a critical operational boundary. The interpretation of change therefore requires predeclared decision criteria and uncertainty estimates as well as a model of task consequences.
Longitudinal Designs for Measuring Skill Change
The ideal longitudinal design measures the same person's unassisted proficiency before meaningful automation exposure and at repeated later intervals. It records the frequency, quality, and difficulty of intervening practice and the features of the automation encountered during each interval. Parallel tasks or calibrated item banks can reduce simple test-retest learning while preserving measurement comparability. A relevant control group performs a similar quantity and difficulty of work with a different function-allocation arrangement.
Random assignment of eligible participants to forms of assistance can improve causal identification, particularly in simulated or educational settings. A factorial design might vary the extent of delegated decision-making and the frequency of required unassisted retrieval. A further factor can manipulate whether the interface explains its intermediate steps. The design can then test interactions among assistance, practice, and time rather than equating automation exposure with a single binary treatment.
When randomization is infeasible, prospective cohorts can record deployment dates and repeated outcome measures across settings. Comparison units must be evaluated for differences in training, workload, prior expertise, case mix, and organizational change. Interrupted time-series approaches can be informative if sufficient observations precede and follow automation adoption, and if changes at the transition are clearly distinguished from ongoing trends. A simple before-and-after average can motivate further study but often cannot exclude secular changes and shifting populations.
Attrition is a serious problem: workers who struggle may change roles or leave, while those who adapt may be overrepresented among those retested. The research record should report missing outcomes and the reasons participants could not complete follow-up. Differential testing exposure can also distort interpretation. The group that repeatedly performs unassisted assessments may receive more practice than workers who only use automation in the workplace, turning the measurement procedure into an active intervention.
Causal Inference and the Automation Exposure Problem
The causal effect of automation on independent competence can be formulated as the difference between a person's future unaided performance after one automation-exposure regime and the performance that same person would have shown after an alternative regime. Only one of those histories can be observed directly for the same person over the same period. Experimental randomization or credible quasi-experimental comparisons provide ways to estimate the missing counterfactual. The relevant treatment is a detailed exposure regimen rather than merely installation of software.
A useful causal diagram would separate automation availability, actual reliance, opportunities for independent practice, corrective feedback, baseline expertise, task difficulty, and subsequent independent competence. Some variables are mediators, some are confounders, and some can occupy different roles depending on the research question. For instance, reliance may mediate the effect of available automation on practice, while initial skill may influence both reliance and later performance. Conditioning on variables without a causal rationale can hide a genuine effect or create a misleading association.
Interference among workers matters as well. When members of a team share automated tools, one person's exposure can change the cases, responsibilities, and instruction available to colleagues. A conventional individual-randomization design may underestimate such spillovers. Cluster randomization by team or work unit can help, but requires enough units, clear exposure records, and adjustment for group-level outcomes. The operational ability of a team also depends on how expertise is distributed across members, so an intervention can change organizational competence without producing the same change in every individual.
Longitudinal outcomes should be modeled with uncertainty appropriate to the design. Baseline adjustment, multilevel models for repeated measures, preregistered primary outcomes, and sensitivity analyses for missing data can strengthen inference. Researchers should report effect sizes and confidence intervals, and distinguish whether the target is an average difference, probability of failure, or proportion falling below a competence threshold. For rare safety events, an effect on incident counts is difficult to estimate from small studies; validated intermediate endpoints may be necessary but should not be presented as observed incident prevention.
Psychometric Validity and Task Sampling
Independent competence cannot be measured by arbitrary collections of quiz items or by one convenient simulation. Construct validity depends on a defensible relation between the test and the psychological operations under investigation. The item set must represent the performance domain, and scores must capture the intended skill rather than familiarity with one interface. Experts and novices can achieve similar top-level accuracy through different strategies, making process-sensitive outcomes useful when they are theoretically justified.
Reliability must also be established. A low score after automation might reflect difficult test cases, inconsistent judging, fatigue, or measurement instability. Parallel forms should be shown to represent comparable levels of difficulty, and thresholds should account for expected error in individual estimates. When a study compares workers with different occupational or educational backgrounds, measurement invariance and differential item functioning warrant investigation. A single universal proficiency scale is unlikely to work across all manual, cognitive, and supervisory domains.
The choice of outcome can influence the apparent direction of skill change. Tatel, C. E., & Ackerman, P. L (2025) found that retention relationships differed across speed, accuracy, and mixed measures. A participant might respond more slowly because they are more careful after learning to distrust automation; such a result should not automatically be classified as deskilling. Conversely, faster decisions achieved by uncritical acceptance of suggestions may conceal a decline in independent verification. Both accuracy and decision strategy are needed for interpretation.
Validated assessment should also include transfer. If a person successfully repeats a practiced manual task, this supports retention of that task; it does not prove flexible performance across the work domain. Researchers can use novel configurations of familiar components, changed operating constraints, and unfamiliar failure modes. Such assessments should avoid tasks so remote that they test unrelated expertise. The essential requirement is an explicit theory of what a competent worker should be able to generalize and why.
Simulations and Rare-Event Stress Testing
Simulation permits ethically controlled tests of situations that organizations hope never to encounter in production. Dynamic control tasks, flight-deck simulators, clinical decision vignettes, and synthetic software failures can introduce rare errors while recording intervention behavior. The studies of Endsley, M. R., & Kiris, E. O (1995) and Kaber, D. B., & Endsley, M. R (2004) demonstrate the usefulness of deliberately manipulating levels of automated control and observing situation awareness, workload, and failure response.
The validity challenge is that the operator knows the environment is an experiment. Anticipation of errors can increase monitoring relative to ordinary work, particularly when a task contains an unusually high proportion of failures. Conversely, a low-fidelity interface can create confusion unrelated to real system behavior. To improve generalizability, investigators should sample scenarios based on documented operational constraints, involve domain experts in assessment construction, and evaluate whether process measures resemble those observed in actual work.
Rare-event stress tests also require distinguishable scenarios. A failure that is announced by a clear warning measures response to the warning; a failure hidden within ambiguous data measures fault detection and inference. A takeover following short automation exposure tests different readiness from one following a sustained period of passive monitoring. Experiments can manipulate both dimensions and document the interaction between retention and transition difficulty. These distinctions make it possible to identify the cognitive functions that interventions must preserve.
A practical testing program should combine frequent low-cost probes of component knowledge with occasional high-fidelity integrated exercises. Short probes permit estimation of retention trajectories without monopolizing operational time. Integrated simulations test coordination among skills, interfaces, and team members. Their scores should be reviewed for both false reassurance and unnecessary intervention, since successful automation supervision includes knowing when the system is functioning correctly as well as recognizing failure.
Retrieval, Spacing, and Deliberate Maintenance
Memory research gives a principled basis for designing retention activities. Roediger, H. L., III, & Karpicke, J. D (2006) showed that retrieval practice can improve later retention even when additional study appears more fluent during acquisition. Cepeda, N. J., Pashler, H., Vul, E., Wixted, J. T., & Rohrer (2006) synthesized evidence that distributed practice benefits later recall across many verbal-learning settings. These results justify testing retrieval and spacing principles in automation-affected skills, while recognizing that complex professional performance requires domain-specific validation.
The retention program should rehearse the operations the human must later execute rather than merely recall textual instructions. For manual control, this includes timed perceptual-motor coordination and diagnosis under constraints. For interpretive expertise, it includes independent examination of raw evidence before exposure to a recommendation. For supervisory control, it includes prediction of system modes, recognition of misleading outputs, and recovery after simulated failure. Each practice event should offer diagnostic feedback so the worker can correct errors and update task representations.
Spacing has to be defined in terms of the required skill and available opportunities. A maintenance exercise conducted too frequently may consume resources without adding meaningful retention; an exercise performed too infrequently may allow critical procedures to become unreliable. The general evidence about distributed retrieval provides a research rationale, not a universal calendar for airline operations, medicine, or AI-enabled professional work. Training intervals should be adapted from delayed performance data, risk thresholds, and the profile of errors observed in the local work system.
A promising design uses periodic automation-free practice in protected settings. The goal is to maintain independent access to the task representation and action sequence while preserving the productivity benefits of automation in routine work. Intervals can include unassisted analysis followed by comparison with the system, unfamiliar failure problems, and brief explanations of why a result is correct. Whether such programs improve real transfer should be tested against active alternatives with matched practice time and feedback, not only against no training.
Training Transfer, Fading, and Feedback
Guidance can help novices acquire competent performance in tasks initially beyond their unaided capacity. The key design question is when and how that guidance should be reduced. Schmidt, R. A., & Bjork, R. A (1992) emphasized the distinction between conditions that optimize acquisition performance and those that support retention and transfer. Gradual withdrawal of assistance can serve as a testable training intervention: the learner is asked to take over components of the process as competence develops.
Fading should be linked to assessed proficiency rather than simply time spent using the software. A learner who can produce correct responses with hints may still need independent retrieval practice before advancing to more demanding cases. An experienced worker may need targeted training on a newly automated subtask that has received little recent manual practice. Adaptive guidance should therefore depend on item-level evidence of performance and the generalization of skill across conditions.
Feedback quality is central. A system that announces the correct answer without explaining discrepancies can sustain accurate output while failing to teach the reason for an error. A system that supplies diagnostic contrasts between competing hypotheses may create richer learning opportunities. However, explanations can also encourage superficial agreement. Comparative experiments should measure delayed independent reasoning, not merely user preferences for explanatory displays. In generative AI, the capacity to produce fluent rationales makes independent verification especially important.
Training should also include purposeful failure exposure. Operators need experience with credible false alarms, omissions, mode confusion, and incomplete information, with errors introduced at a manageable level of risk in simulation. Successful maintenance includes the ability to recover from mistakes and recognize the limits of personal certainty. This type of training should be accompanied by evaluation of transfer: the worker must handle a failure whose superficial appearance differs from the examples used in instruction.
Interface Design and Allocation of Human Control
Preserving human competence is partly an interface and work-design problem. Parasuraman, R., Sheridan, T. B., & Wickens, C. D (2000) provide an architecture for considering which stages of information processing and action should be automated. Lee, J. D., & See, K. A (2004) show why calibrated reliance requires information about the system's capabilities and limits. Together, these approaches motivate designs that make human judgment possible and support proportionate oversight.
Interfaces can preserve access to evidence, intermediate state, and alternative actions. A user may be better positioned to maintain expertise when the system indicates how it reached a suggestion, exposes uncertainty or conflicting inputs, and permits independent inspection. These features need evaluation because transparency can add workload or produce false confidence when explanations are plausible but unfaithful to the underlying process. Good design must be judged by accuracy, intervention behavior, retention, and safety outcomes.
Human control also requires viable timing. An operator assigned formal authority to override automation may be unable to act effectively if the warning arrives after the critical intervention window. Control-transfer procedures should therefore specify what information is presented, what functions remain available, and how much time the person has to reconstruct the situation. Maintaining manual skill is valuable only if the work system gives that skill a practical opportunity to influence outcomes.
The allocation of ordinary practice is another design variable. Organizations can reserve opportunities for independent execution in training environments or selected low-risk tasks while leaving high-performing automation in place for routine operations. They can rotate workers through diagnosis and recovery exercises and monitor competence at a level suitable for professional responsibility. Such arrangements reflect a system-level approach to continuity: capability is maintained through scheduled exposure, feedback, assessment, and usable control pathways rather than individual willpower alone.
Organizational Responsibility for Competence Continuity
Individual learning is embedded in a work system that determines the distribution of tasks, access to practice, performance feedback, staffing, and responsibility for intervention. A workforce can have access to advanced automation while losing the institutional routines through which experts previously trained successors. The risk is especially clear when a function remains formally assigned to human staff for abnormal situations but has disappeared from ordinary practice. Organizations must therefore decide explicitly which independent capabilities they require and how those capabilities will be sustained.
An organizational competence register can map critical functions to qualified personnel, specify realistic failure scenarios, and document the frequency and quality of independent exercises. Such records should track demonstrated proficiency rather than simply time in a role or completion of an online course. An organization may need overlapping skills among multiple workers so that resilience survives absence, turnover, or simultaneous failures. The appropriate degree of redundancy depends on exposure to hazards, the consequences of delayed response, and availability of alternate procedures.
Work redesign can either increase or decrease opportunities for learning. When automation removes routine work, personnel may gain time for complex reasoning, multidisciplinary communication, or simulation-based practice. They may also experience fragmented roles in which their remaining tasks are limited to checking outputs without access to causal information. Measuring those consequences requires longitudinal organizational evidence rather than an assumption that all automation produces the same career trajectory.
The distribution of responsibility must remain coherent. A worker cannot be expected to recover from a system failure for which they have received neither training nor sufficient information or authority. The organization selecting the automation and configuring the workflow shapes the operator's opportunities for competence maintenance. This perspective makes continued human capability a documented system outcome involving procurement, interface design, professional education, supervision, and evaluation.
Workforce Development, Equity, and the Experience of Expertise
Automation can redistribute valuable learning opportunities unevenly. Workers who already possess strong independent skill may use assistance selectively and retain access to complex cases. Novices may receive finished results that support task completion while bypassing essential stages of acquisition. If early-career staff are assigned increasingly narrow review roles, their later ability to assume senior responsibility can be affected. The size and direction of this possibility require direct study, but the mechanism is important for the design of professional training.
Access to independent practice is not merely a personal preference. It can depend on organizational permission, seniority, workload, scheduling, resources, and the ability to practice without jeopardizing service quality. A maintenance program may unintentionally favor workers with protected training time or high-quality simulations. Researchers should examine whether training opportunities are distributed fairly and whether performance thresholds are interpreted consistently across roles and populations.
The experience of competence includes confidence, professional identity, and perceived autonomy. Some workers may welcome automated assistance because it reduces burdensome tasks and creates space for difficult work. Others may perceive reduced control over their craft or fear that their expertise is becoming less visible. These experiences can influence willingness to practice, report errors, or challenge a system. However, subjective loss of identity must be measured separately from actual performance decline, just as subjective confidence must be distinguished from skill proficiency.
A humane and scientifically grounded competence policy should make assessments developmental when possible. Workers need feedback about which capabilities remain strong, which require maintenance, and how changes in work design affect performance. Assessment systems that focus exclusively on individual blame can discourage the reporting of uncertainty or automation failure. A culture of learning is relevant because accurate reporting supplies the evidence necessary to improve both automated tools and the human capabilities supporting them.
The Artificial Era and Changing Forms of Human Expertise
The Artificial Era names the historical setting in which increasingly autonomous artificial systems participate in activities previously organized around human cognition and action. In that setting, research on competence continuity addresses more than the preservation of familiar manual routines. It examines how capability is distributed across people, artificial systems, interfaces, and institutions, and how the distribution changes through long-term use. The scientific questions remain operational: what the human can do unaided, what the combined system can do, and what happens when the system becomes unavailable or misleading.
The concept of distributed or externalized cognition can help describe a task performed across human and technical resources. Risko, E. F., & Gilbert, S. J (2016) provide an experimental basis for studying external support; automation research supplies methods for analyzing reliance and functional allocation. Such perspectives do not settle whether every traditional component of expertise must be retained. They make it possible to ask which internal abilities and which external resources are essential to a particular continuity requirement.
AI assistance has distinctive implications for epistemic work. A professional may continue to produce sophisticated documents or decisions while delegating the formulation of intermediate arguments, interpretation of evidence, or generation of options. This can change the evidence available for judging expertise. An output may be correct because the person understood the problem, because the system generated an appropriate solution, or because the person successfully checked a solution they could not have generated independently. All three achievements can be valuable, but they are different capacities.
A competence-continuity framework must therefore describe the function being evaluated and the conditions of its exercise. Independent human reasoning should be assessed wherever professional responsibility, foreseeable loss of assistance, or the need for genuinely novel solutions demands it. Collaborative capability should be evaluated wherever human and artificial contributions jointly determine task outcomes. Successful human–artificial cooperation can be accompanied by maintenance, transformation, or selective decline of particular human abilities; empirical measurement is needed to establish which occurs.
Integrative Scientific Synthesis
The evidence supports a multilevel account of automation-related competence continuity. At the individual level, durable skill depends on initial acquisition, the kind and frequency of practice, retrieval, feedback, and transfer. At the interaction level, automation determines which cognitive operations are performed by the person and which are delegated, how system errors become visible, and how control is returned. At the work-system level, training access, role allocation, procedures, and institutional redundancy shape the opportunities to preserve and exercise expertise.
Three research distinctions organize the synthesis. First, current assisted performance must be separated from later independent competence. Second, routine manual task performance must be separated from diagnosis and recovery after automation failure. Third, individual skill change must be separated from the organizational allocation of expertise and responsibility. These distinctions make apparently contradictory findings interpretable and permit research designs to specify relevant outcomes.
The empirical foundation has unequal strengths across questions. General procedural retention has substantial quantitative evidence, including the meta-analysis of Tatel, C. E., & Ackerman, P. L (2025). Experimental human-factors studies identify conditions under which automation changes situation awareness and failure intervention, including the integrated analysis by Onnasch, L., Wickens, C. D., Li, H., & Manzey, D (2014). Applied simulator studies, such as Casner, S. M., Geven, R. W., Recker, M. P., & Schooler, J. W (2014), demonstrate that skill components can differ in persistence. Clinical observational evidence, notably Budzyń and colleagues (2025)00133-5), provides a contemporary signal relevant to AI exposure. Longitudinal causal estimates across diverse professional AI uses are less developed.
The resulting explanatory picture is conditional rather than universal. Delegation affects future human capability through its impact on meaningful practice, monitoring, feedback, and mental representation. Its consequences depend on the task, person, exposure schedule, interface, organizational environment, and criterion used to define competence. A system that reduces repetitive manual activity can also create new learning opportunities if it allocates time and feedback toward independent diagnosis and resilient intervention. The effects must be established for the actual work arrangement.
Unresolved Questions and a Research Agenda
A central open question is the causal contribution of automation exposure relative to the effect of ordinary time without practice. Studies need to compare matched exposure regimes while measuring the actual component operations performed by users. Without such designs, the phrase automation-induced deskilling may refer to unrelated pathways, such as loss of task access, ordinary forgetting, changes in task difficulty, or reliance decisions under time pressure.
Another question concerns the boundaries of transfer. Does periodic independent practice on a narrow subset of cases preserve broad diagnostic or motor capability? Which kinds of variation in training scenarios support robust response to novel failures? Can simulation predict real-world safety outcomes sufficiently well for training programs to use it as a qualification criterion? Research should examine generalization beyond rehearsed examples, stability over meaningful intervals, and the role of foundational explanatory knowledge.
Metacognitive awareness warrants explicit longitudinal investigation. Researchers should measure whether operators accurately anticipate their own unaided performance after extensive assisted work. If confidence remains high while performance falls, that discrepancy may influence later reliance decisions and training uptake. Yet calibration can also improve when interfaces expose uncertainty and give meaningful feedback. The direction and magnitude of these effects remain questions for experiments rather than assumptions about the psychology of AI use.
Comparative research across occupations is needed. Visual detection, software development, language work, aviation control, and clinical decision-making differ in the degree to which performance is procedural, perceptual, interpretive, or socially coordinated. An intervention effective for one class of skill may transfer poorly to another. Research should develop domain-specific measurement instruments within a shared conceptual framework and report occupational contexts with sufficient detail for later synthesis.
Finally, competent governance requires evaluation of whole systems. Organizations need to know whether human recovery capacity survives software updates, task redistribution, personnel changes, and simultaneous disruptions. Studies should measure the consequences of designated practice time, training access, staffing redundancy, and transparent error reporting. The most consequential effects may be organizational and delayed, emerging only after enough time has passed for normal learning opportunities to have been displaced.
Conclusion
Automation-related skill retention and human competence continuity concern the persistence of human capacities when technology changes what people practice, what they observe, what they decide, and when they must intervene. The phenomenon is scientifically tractable through the analysis of learned skills, function allocation, delayed unassisted performance, awareness of system state, verification, failure recovery, and organizational arrangements for training.
Research on forgetting and procedural retention establishes the significance of task type, elapsed time, and opportunities for practice. Human-factors experiments show that high levels of automation can improve routine performance while imposing costs when intervention is required. Applied evidence reveals important differentiation among skill components; preserved manual fluency can coexist with weakness in cognitive operations needed for autonomous work. Emerging clinical studies and theoretical work on AI-assisted expertise expand the research program while leaving long-term causal claims open.
Competence continuity is achieved when the capabilities required for reliable human responsibility remain demonstrably accessible under the conditions in which they will be needed. Its evaluation therefore requires a combination of individual skill tests, realistic failure simulations, longitudinal measurement, and assessment of the work system. The enduring scientific issue is how human and artificial contributions can be organized so that immediate operational performance, durable human learning, and resilience under technological change are measured and sustained together.
References
- Arthur, W., Jr., Bennett, W., Jr., Stanush, P. L., & McNelly, T. L. (1998). Factors that influence skill decay and retention: A quantitative review and analysis. Human Performance, 11(1), 57–101. https://doi.org/10.1207/s15327043hup1101_3
- Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775–779. https://doi.org/10.1016/0005-1098(83)90046-8
- Budzyń, K., Romańczyk, M., Kitala, D., Kołodziej, P., Bugajski, M., Adami, H. O., Blom, J., Buszkiewicz, M., Halvorsen, N., Hassan, C., Romańczyk, T., Holme, Ø., Jarus, K., Fielding, S., Kunar, M., Pellisé, M., Pilonis, N., Kamiński, M. F., Kalager, M.,... Mori, Y. (2025). Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: A multicentre, observational study. The Lancet Gastroenterology & Hepatology, 10(10), 896–903. https://doi.org/10.1016/S2468-1253(25)00133-5
- Casner, S. M., Geven, R. W., & Williams, K. T. (2013). The effectiveness of airline pilot training for abnormal events. Human Factors, 55(3), 477–485. https://doi.org/10.1177/0018720812466893
- Casner, S. M., Geven, R. W., Recker, M. P., & Schooler, J. W. (2014). The retention of manual flying skills in the automated cockpit. Human Factors, 56(8), 1506–1516. https://doi.org/10.1177/0018720814535628
- Cepeda, N. J., Pashler, H., Vul, E., Wixted, J. T., & Rohrer, D. (2006). Distributed practice in verbal recall tasks: A review and quantitative synthesis. Psychological Bulletin, 132(3), 354–380. https://doi.org/10.1037/0033-2909.132.3.354
- 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
- Goddard, K., Roudsari, A., & Wyatt, J. C. (2012). Automation bias: A systematic review of frequency, effect mediators, and mitigators. Journal of the American Medical Informatics Association, 19(1), 121–127. https://doi.org/10.1136/amiajnl-2011-000089
- Griffiths, N., Bowden, V., Wee, S., Strickland, L., & Loft, S. (2024). Operator selection for human-automation teaming: The role of manual task skill in predicting automation failure intervention. Applied Ergonomics, 118, Article 104288. https://doi.org/10.1016/j.apergo.2024.104288
- 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
- Kahneman, D., & Klein, G. (2009). Conditions for intuitive expertise: A failure to disagree. American Psychologist, 64(6), 515–526. https://doi.org/10.1037/a0016755
- Klostermann, M., Conein, S., Felkl, T., & Kluge, A. (2022). Factors influencing attenuating skill decay in high-risk industries: A scoping review. Safety, 8(2), Article 22. https://doi.org/10.3390/safety8020022
- 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
- Macnamara, B. N., Berber, I., Çavuşoğlu, M. C., Krupinski, E. A., Nallapareddy, N., Nelson, N. E., Smith, P. J., Wilson-Delfosse, A. L., & Ray, S. (2024). Does using artificial intelligence assistance accelerate skill decay and hinder skill development without performers’ awareness? Cognitive Research: Principles and Implications, 9, Article 46. https://doi.org/10.1186/s41235-024-00572-8
- 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., 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
- Risko, E. F., & Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences, 20(9), 676–688. https://doi.org/10.1016/j.tics.2016.07.002
- Roediger, H. L., III, & Karpicke, J. D. (2006). Test-enhanced learning: Taking memory tests improves long-term retention. Psychological Science, 17(3), 249–255. https://doi.org/10.1111/j.1467-9280.2006.01693.x
- 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
- Schmidt, R. A., & Bjork, R. A. (1992). New conceptualizations of practice: Common principles in three paradigms suggest new concepts for training. Psychological Science, 3(4), 207–218. https://doi.org/10.1111/j.1467-9280.1992.tb00029.x
- 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
- Soderstrom, N. C., & Bjork, R. A. (2015). Learning versus performance: An integrative review. Perspectives on Psychological Science, 10(2), 176–199. https://doi.org/10.1177/1745691615569000
- Tatel, C. E., & Ackerman, P. L. (2025). Procedural skill retention and decay: A meta-analytic review. Psychological Bulletin, 151(6), 696–736. https://doi.org/10.1037/bul0000481
- Vlasblom, J. I. D., Pennings, H. J. M., van der Pal, J., & Oprins, E. A. P. B. (2020). Competence retention in safety-critical professions: A systematic literature review. Educational Research Review, 30, Article 100330. https://doi.org/10.1016/j.edurev.2020.100330