RESEARCH DOSSIER
Digital Coaching, Wearable Monitoring, and Algorithmic Agency in Sport
Abstract. Digital coaching, wearable monitoring, and algorithmic agency in sport form an integrated psychological research object at the intersection of skilled action, physiological sensing, motivational regulation, human–automation interaction, predictive modeling, and athlete governance. This fundamental critical monograph reconstructs the historical origins of athlete monitoring, defines the distinct tasks of measurement, prediction, advice, and organizational decision-making, and evaluates their evidentiary requirements. Original research on self-efficacy, automation reliance, and algorithmic advice is assessed alongside sport-specific sensor validation studies, training-load research, systematic reviews of injury prediction, and contemporary work on data rights and ethical AI. The scientific synthesis finds that sensor accuracy depends on the outcome and setting; predictive accuracy, when established, does not independently demonstrate benefit from acting on predictions; and athletes' responses to computational advice depend on motivation, trust, institutional power, and context. Recent reviews expose serious weaknesses in external validation and model calibration for injury prediction, whereas emerging experiments demonstrate that sport-related AI-framed advice can influence judgments. The research therefore specifies a multistage evaluation architecture covering physical signals, construct validity, model generalization, decision behavior, causal effectiveness, privacy, fairness, and lifecycle oversight. The principal conclusion is that the value of algorithmic coaching must be established through meaningful athlete outcomes and demonstrably accountable human–artificial decision processes. This work is an integrative critical monograph, not an original experiment, registered systematic review, or meta-analysis.
Keywords: sport psychology; digital coaching; wearable monitoring; athlete monitoring; algorithmic agency; human–AI interaction; sports artificial intelligence; training load; injury prediction; motor learning; trust in automation; athlete autonomy; self-determination; algorithmic fairness; causal inference; data governance; decision support; Artificial Era.
Scientific Overview
Digital coaching in sport is the organization of training-related feedback, recommendations, communication, and decisions through computational systems. Wearable monitoring is the repeated acquisition of physiological, movement, and behavioral signals using devices carried on or near the athlete. Algorithmic agency, as used in this monograph, describes the observable influence of algorithmic outputs on the selection, timing, justification, and revision of training actions. The phrase concerns decision functions and attributed authority; it does not assume that statistical software possesses consciousness, autonomous interests, or clinical competence. Its scientific object is a coupled athlete–coach–technology system in which measurements and recommendations can change behavior.
The field requires simultaneous attention to three questions. First, what can a sensor measure with acceptable error under the specific movement, environment, and population in which it is used? Second, what does a statistical model infer about a defined target, for which athletes and over what horizon? Third, does acting on that information improve a meaningful outcome compared with an available alternative? Success on the first question does not establish success on the second, and neither proves benefit on the third. These are separate links in a causal chain from sensing to decision and outcome. This distinction follows concerns raised in training-load theory by Passfield, L., Murias, J. M., Sacchetti, M., & Nicolò, A (2022) and in contemporary discussions of accountable sport AI by Kim, J.-H., Kim, J., Kang, H., & Youn, B.-Y (2025).
The empirical domain extends from elite football monitoring, endurance coaching, and team-sport position tracking to recreational training applications, rehabilitation-oriented support, and emerging multimodal predictive assistants. Relevant outcomes include motor learning, adherence, perceived competence, fatigue, psychological strain, injury occurrence, exercise exposure, sporting performance, autonomy, trust, privacy, and relationships with coaching staff. Each outcome must be separately operationalized. A device that increases the frequency of feedback, for example, may change adherence without improving competition performance; an accurate fatigue estimate may improve planning yet also generate pressure if used to enforce selection.
This work is a critical integrative research monograph. It examines identifiable experimental studies, validation investigations, consensus statements, systematic reviews, and human–automation research; it does not claim a registered systematic search, pooled effect-size meta-analysis, or original participant-level dataset. Sport-specific observations receive greater weight than general technology findings for sport-specific conclusions. Where evidence comes from physical-activity interventions outside competitive sport, it is treated as a hypothesis-generating analogy rather than direct confirmation of elite-coaching effectiveness.
Historical Foundations of Digital Athlete Monitoring
Coaching has always involved information acquisition: timing trials, recording training sessions, examining movement, interviewing athletes, and revising practice. The digital transition reorganized those activities around increasingly persistent records, higher-frequency measurements, and automated summaries. Session-level rating of perceived exertion, heart-rate observation, and performance tests established a longstanding practical distinction between what an athlete does and how the athlete responds. Contemporary devices attach standardized timestamps and machine-readable records to this distinction, making repeated intra-athlete comparison technically easier while introducing dependencies on vendor software and computational interpretation.
The history of sport-monitoring technology also illustrates that measurement techniques have evolved at different rates. Satellite-based tracking estimates movement in outdoor spaces; inertial systems register accelerations and angular motion; photoplethysmography estimates pulse-related optical changes; chest electrodes and other physiological sensors record different signals under different constraints. Their measurements become actionable only after a researcher or coach specifies the construct of interest. Bourdon, P. C., Cardinale, M., Murray, A., Gastin, P., Kellm (2017) treated athlete-load monitoring as a multidisciplinary task requiring clarification of the meaning, purpose, and limitations of measured loads.
A survey by Akenhead, R., & Nassis, G. P (2016) documented contemporary football practitioners’ approaches and perceptions. Such a descriptive study establishes what coaches reported doing at that time; it does not establish that each common monitoring practice improves performance. Sport science has since experienced a transition from descriptive dashboards toward machine-learning forecasts of injury, readiness, and performance. The added predictive layer creates new problems involving temporal leakage, representativeness, calibration, and responsibility for acting on outputs.
The human-factors history is equally important. Parasuraman, R., & Riley, V (1997) organized the human use of automation around appropriate use, misuse, disuse, and abuse. Lee, J. D., & See, K. A (2004) explained why trust influences reliance when system behavior is complex or partly opaque. These traditions predate current artificial-intelligence applications and provide testable psychological constructs for digital coaching: reliance, overriding, monitoring, perceived reliability, responsibility, and control.
The contemporary research history therefore contains interacting trajectories rather than a single invention. Sensor engineering improved the availability of data; sport physiology developed monitoring constructs; motor-learning research studied feedback; motivational psychology examined autonomy and competence; predictive modeling supplied classification and forecasting; and organizational ethics addressed the rights of the monitored athlete. The integrated object emerges when these trajectories determine one athlete’s practical training environment.
Conceptual Boundaries and the Unit of Analysis
Digital coaching refers to a broad intervention class. Its minimal form displays an observed quantity and lets a person interpret it. More complex systems provide targets, reminders, explanations, trend detection, adaptive prescriptions, injury alerts, conversational guidance, and automated changes to a training plan. A distinction should be retained between passive recording, descriptive visualization, diagnostic classification, prediction, recommendation, and direct actuation. Each imposes a different evidentiary burden. Recording heart rate is a measurement task. Estimating next-week injury probability is a prediction task. Reducing tomorrow’s training volume on that basis is a policy decision.
Wearable monitoring refers to the acquisition process and its supporting software. A wrist device, chest strap, satellite receiver, inertial unit, textile sensor, or video-assisted system may be studied independently of the advisory interface that displays its output. Measurement validity is defined relative to a criterion, construct, and application. Construct validity asks whether an indicator supports the intended psychological or physiological interpretation; criterion accuracy asks how well it agrees with an appropriate reference method. Repeatability and agreement under repeated comparable conditions form separate dimensions of reliability. These distinctions become especially important where a proprietary readiness score combines multiple imperfect inputs.
Algorithmic agency can be measured through behavior rather than attributed by rhetoric. A study can record recommendation acceptance, intervention timing, frequency of overrides, explanation requests, deviations from coach plans, changed goals, and subsequent outcomes. The amount of decision authority delegated to software may differ from its computational sophistication. A simple hard threshold can exert great organizational control when its alert triggers automatic non-selection. A sophisticated predictive model can exert very little control if its recommendation is treated as one fallible input to deliberation.
The appropriate unit of analysis is often the athlete–coach–organization–system relationship over time. An individual-level experimental contrast may estimate how a cue changes a participant’s immediate performance, while team-level deployment may alter trust, communication, fairness, and access to opportunities. Selection policies are typically controlled by institutions, not devices. Accordingly, sport-specific algorithmic governance needs organizational outcomes and actor-level outcomes in addition to sensor accuracy.
A full causal account must also distinguish human acceptance of a recommendation from benefit resulting from that recommendation. High acceptance can signify justified trust, a persuasive interface, compliance under authority, or lack of alternatives. Decision quality requires task-appropriate ground truth, counterfactual comparison, and transparent specification of whose interests the objective function serves. The psychology of sport enters at every step because beliefs, fatigue, aspirations, relationships, and incentives affect both the production of sensor data and the response to it.
Feedback, Practice, and the Psychology of Coaching
Digital coaching frequently borrows a familiar proposition from skill acquisition: feedback helps learners detect discrepancies between present execution and desired execution. The intervention literature makes this proposition conditional. The effect of feedback depends on its informational content, timing, frequency, relation to task goals, the learner’s expertise, and whether the feedback supports subsequent self-regulation. More frequent data streams do not automatically provide better motor instructions.
A feedback system can convey knowledge of results, such as distance, speed, successful actions, or task outcome, and knowledge of performance, such as movement characteristics. Some information may already be available through the learner’s sensory experience. Excessive augmented feedback may encourage learners to attend to artificial cues that are unavailable during actual competition. The distinction between performance during acquisition and retained learning is central: an immediate score improvement may dissipate when augmented feedback is withdrawn. Retention and transfer tests consequently belong in trials of digitally mediated skill training.
Research on attentional focus provides a testable example. Wulf, G (2013) synthesized evidence on directing attention toward movement effects versus bodily mechanics in motor-learning tasks. Its implications for digital interfaces are specific: feedback may redirect attention in ways that affect execution. Translating that evidence into a particular sport application requires evaluation of task, expertise, timing, and intervention design; the general review cannot validate every product using an external-focus prompt.
Coaches also exercise judgment that is not reducible to a one-dimensional target. A movement modification can interact with tactical purpose, confidence, fatigue, prior injury, and stage of learning. Computational feedback may be helpful when it makes a relevant discrepancy visible, but harmful when it suppresses contextual reasoning or penalizes a technique that is effective for a particular athlete. The quality of digital coaching therefore depends on opportunities to interpret, negotiate, and test cues, not merely on throughput of information.
The most informative experiments compare forms of feedback while keeping contact time and task demands comparable. They specify whether the outcome is learning, immediate performance, self-efficacy, or adherence. They also test withdrawal and transfer, because a system that produces improvements only while continuously supplying reminders may be training dependence on a tool rather than flexible skilled action. Digital coaching can be evaluated as an instructional intervention only when these outcomes are defined and measured separately.
Autonomy, Competence, and Motivational Regulation
Self-determination theory proposes that autonomy, competence, and relatedness are important psychological needs associated with motivational quality. Ryan, R. M., & Deci, E. L (2000) developed this framework for intrinsic motivation and well-being. In sport, an athlete may experience a training recommendation as useful information that supports self-directed mastery or as a controlling demand backed by selection consequences. The same device signal can therefore operate within distinct motivational climates.
Autonomy refers to experienced volition and endorsement of one’s actions. It does not require rejecting coaching expertise. Athletes can willingly follow a demanding program when they understand its rationale and participate meaningfully in decisions. Digitally delivered targets can strengthen competence when progress indicators are interpretable, sufficiently accurate, and matched to achievable goals. They can also alter competence judgments through normative comparisons or misleading daily volatility. Relatedness becomes salient when monitoring supports constructive conversations or replaces conversations with automated ratings.
Bandura, A (1977) provides a complementary framework for understanding how performance information changes expectations about capability. A verified improvement in a task may strengthen confidence. An unvalidated recovery score should not be treated as direct measurement of psychological self-efficacy. The construct is defined by capability beliefs concerning specified actions, whereas a readiness score may combine physiological proxies and vendor-designed weights.
Digital motivational interventions should be studied as packages of informational and social ingredients. Reminders, social comparison, individualized targets, coach messages, and reward contingencies often appear together. A positive outcome after deployment cannot isolate the contribution of the algorithm unless the study varies or otherwise identifies the relevant ingredient. Randomized factorial studies, micro-randomized interventions, and interrupted time-series designs can help distinguish feedback content from frequency, personalization, and coach involvement.
Research on non-athlete physical-activity promotion supplies useful but limited analogies. Digital devices can increase measured activity in some populations, while results differ by comparator and follow-up duration. Competitive performance requires sport-specific skill, physiological adaptation, exposure management, and other outcomes not captured by step counts. Inferences about elite coaching should therefore rest on athlete populations and competition-relevant endpoints rather than adoption rates or generalized fitness engagement.
The Athlete–Coach Relationship as a Technological Context
Sport coaching is a social relation involving instruction, mutual interpretation, status, accountability, and often considerable asymmetry of power. Digital monitoring changes which observations become shareable and how quickly they can be transformed into evaluations. A coach may use trends to open a conversation about fatigue. An organization may instead turn a single score into an eligibility threshold. These uses are psychologically and ethically different even when they depend on the same sensor.
Coaching systems can modify the temporal distribution of attention. Continuous reports create additional opportunities to adjust practice, but they also extend evaluation beyond sessions. An athlete may feel responsible for optimizing sleep, mobility, and recovery metrics at all hours. The potential motivational and well-being effects depend on transparency about purpose, boundaries for out-of-hours data use, coach communication, perceived control, and consequences of declining monitoring. These are research hypotheses requiring direct athlete reports and behavioral measures.
Relationships also shape data quality. Athletes who fear punishment for a poor wellness score may alter self-reports or conceal symptoms. When self-reported fatigue, soreness, or mood is a model input, such incentives can create systematic measurement error. A mathematically sophisticated classifier cannot recover information that participants have strategic reasons to misreport. Designs that examine trust, confidentiality, and data governance are therefore also designs about predictive validity.
The Reardon, C. L., Gouttebarge, V., Kroshus-Havril, E., Miller (2026) emphasizes the role of sporting environments, culture, safeguards, and appropriate professional care. Routine psychological assessment discussed in that document is not equivalent to unregulated continuous behavioral profiling. Professional clinical screening has specific instruments, confidentiality responsibilities, referral practices, and qualified personnel. Digital coaching products that infer mental-health status from ordinary performance telemetry require particularly careful validation and governance.
Longitudinal research should investigate whether digital coaching strengthens mutual understanding, displaces individualized conversation, or redistributes accountability after poor performance. It should obtain perspectives from athletes, coaches, clinicians, administrators, and sport scientists; each observes different facets of system operation. A reliable assessment requires attention to team hierarchy, contract incentives, gender, age, and level of competition rather than treating technology as a uniform intervention.
From Algorithmic Advice to Observable Agency
An algorithm becomes consequential in sport when its recommendations change training action. This effect can occur through direct automation, a coach’s deference to a dashboard, an athlete’s self-management, or institutional policy that assigns consequences to a score. A useful operational measure of algorithmic agency is the proportion and type of consequential decisions for which an algorithmic output is a material input, paired with measures of override opportunities and actual choice behavior. Such measures establish a researchable distribution of influence without attaching speculative psychological traits to code.
Decision authority must be distinguished from attribution. An athlete might describe a system as authoritative because its interface presents recommendations in precise numerical language, while a coach might retain final legal or organizational responsibility. If a model produces an apparently objective percentile and its derivation is opaque, the number may acquire more authority than its evidence warrants. Conversely, an accurately calibrated forecast may be dismissed after a visible error. Both reactions are compatible with the human-automation literature.
Dietvorst, B. J., Simmons, J. P., & Massey, C (2015) demonstrated algorithm aversion in experiments involving forecasts: seeing an algorithm err reduced confidence in it, sometimes despite its superiority to a human forecaster. Logg, J. M., Minson, J. A., & Moore, D. A (2019) observed algorithm appreciation in several judgment tasks. These results support the proposition that reliance is context-dependent. Neither establishes a universal athlete response to a coaching algorithm; sport-specific studies must measure baseline expertise, identity, accountability, outcome stakes, and opportunities to intervene.
A newly published sport-specific primer by Wunderlich, F., & Reis, M (2026) identifies trust in AI in sport as an important underinvestigated field. The distinction between theoretical promise and direct evidence is decisive. Sport offers measurable performances and repeated decisions, making it a valuable setting for studying human–AI cooperation, but the existence of measurable outcomes does not establish that an advisory technology has improved them.
In the intellectual horizon of Artificial Era, digital coaching is a bounded empirical setting for examining cooperation between human participants and computational decision systems. The philosophical idea of Artificial as an independent historical order is a wider conceptual framework; empirical sport psychology investigates specific systems, behaviors, measurements, and decisions. Future work can examine adaptive, agentic coaching architectures with explicit behavioral criteria: persistence of plans, context-sensitive revision, memory of prior recommendations, appropriate uncertainty reporting, and documented effects on joint action. An empirical claim about an artificial coaching system remains tied to its observable design and experimental consequences.
The Measurement Architecture of Wearable Sport Systems
Wearable platforms operate across a sequence of representations: a physical signal is acquired, filtered, segmented, converted into features, summarized across a chosen time window, and presented as a construct relevant to training. Each transformation introduces assumptions. An accelerometer measures acceleration, not motivation; a pulse sensor estimates a cardiovascular signal, not readiness; a satellite receiver estimates movement, not tactical intelligence. System developers may eventually infer higher-order targets, but the validity of the inference must be established at each step.
The unit and sampling window affect scientific meaning. A high-frequency accelerometer stream can characterize movement events, while a weekly average may conceal intense but brief exposures. A device worn only during supervised practice misses privately completed workouts and some recovery contexts. A raw optical waveform and a proprietary heart-rate variability score are not interchangeable. Even where both are called biomarkers, the measurement procedures and known error structures can differ substantially.
The scientific program should begin with an explicit measurement model: source signal, preprocessing decisions, target quantity, reference criterion, conditions of use, clinically or practically meaningful error, and expected missingness. If the same label is applied to incomparable measurement pipelines, inter-device comparisons become unreliable. Device firmware and algorithm version belong in study provenance because updates can change the mapping from observations to reported values.
A systematic review by Čokorilo, N., Manolopoulos, N., Matijević, T., & Rajović, R (2026) examined validation studies of wearables in team sports and found that accuracy varied across outcomes and environments. Heart-rate measures were comparatively well supported under controlled conditions, whereas energy-expenditure and aerobic-capacity estimates were less consistently accurate. The review included eleven studies and 188 participants, limiting how broadly its conclusions can be transported across sports, devices, and competitive levels. Its practical contribution is to reject the notion of one global accuracy rating for a wearable.
This measurement-specific approach changes evaluation priorities. A coach deciding whether an athlete has completed a prescribed training distance requires different accuracy tolerances than a clinician evaluating an arrhythmia, and both differ from a selection committee predicting future performance. Every high-stakes use should state its own estimand, decision threshold, confidence limits, and safeguards against error.
Position Tracking, Inertial Sensing, and External Workload
Global navigation satellite systems, local positioning systems, and inertial measurement units are central to monitoring running, acceleration, velocity, displacement, and movement patterns. They are particularly attractive in team sports, where collective activity occurs across large spaces. A frequently reported distance estimate may appear straightforward, but accuracy depends on reception conditions, device placement, sampling frequency, algorithms, body motion, and the reference chosen for validation.
Luteberget, L. S., & Gilgien, M (2020) reviewed validation methods for global and local positioning systems in team sports. Their scoping review included 48 studies and showed that reference methods often used known distances, timing gates, or speed instruments rather than full dynamic ground-truth trajectories. This distinction matters because agreement on total distance does not guarantee accurate detection of abrupt changes of direction or brief high-speed events. External workload metrics can therefore be more reliable for one aspect of a task than another.
Inertial devices measure acceleration and rotation and can be placed on the trunk, limb, or equipment. Different locations record different mechanical events. An algorithm translating acceleration into jump counts, impacts, or a composite workload score requires validation against the underlying event definition. Device placement and movement artifacts also affect repeatability. A threshold for an acceleration event is a convention until linked to a meaningful sports outcome.
Comparability across athletes requires more than consistent units. Players may have different roles, body dimensions, tactical assignments, and exposures. Relative workload may be defined by distance per minute, high-speed efforts above an individual threshold, accumulated accelerations, or another task-specific quantity. Each measure answers a different question, and a team average may conceal individual exposure patterns. Longitudinal changes can be informative within an athlete even when comparisons between athletes are less reliable.
The monitoring literature contains a recurring inferential leap: a measurable external load is treated as an established dose that produces a known adaptation or injury risk. Passfield, L., Murias, J. M., Sacchetti, M., & Nicolò, A (2022) questioned the validity of common training-load metrics as measures of training dose. Their critique directs attention to what a measure actually predicts, whether duration confounds its value, and whether interventions differing in intensity or duration can be meaningfully compared. A device can accurately count an event while the composite interpretation remains scientifically weak.
The appropriate research design separately validates the recorded movement, the composite index, and the action rule built on that index. This separation prevents instrument precision from being mistaken for an established causal theory of physical adaptation.
Optical Heart Rate, Cardiovascular Response, and Artifact
Heart-rate monitoring is common because the signal is relatively familiar to coaches and can help characterize cardiovascular response during activity. Wrist-worn photoplethysmography estimates pulse-related changes in blood volume from optical signals. It is exposed to motion artifact, fit, contact pressure, sampling limitations, physiological variation, and optical characteristics of the measurement site. Chest-based electrical measurements record a different physical signal and may serve as a comparator in specified conditions.
The large consumer-wearable evidence base should be read at outcome and device level. Doherty, C., Baldwin, M., Keogh, A., Caulfield, B., & Argent (2024) conducted a living umbrella review of systematic reviews and found substantial heterogeneity across heart rate, sleep, activity, energy expenditure, and other measurements. A broad review of consumers is relevant for device-methodology questions, but its results do not eliminate the need to validate sport-specific intensity profiles, movements, and athlete characteristics.
The use of optical measurement requires investigation of differential error across populations. Koerber, D., Khan, S., Shamsheri, T., Kirubarajan, A., & Meh (2023) reviewed studies stratified by skin tone or race and found mixed results: some reported reduced accuracy in darker-skinned participants, while others reported no significant effect. The review’s small total evidence base limited certainty. A scientifically appropriate conclusion is that fairness assessment must include direct validation and missing-data analysis, not that every device performs identically or that any one group is always mismeasured.
Heart rate is also an imperfect indicator of training effectiveness. It varies with temperature, hydration, emotional arousal, illness, sleep, and medication, among other influences. A high heart rate during a session can be meaningful but does not uniquely identify exertion, psychological strain, or maladaptation. An algorithm that combines heart rate with perceived exertion or session duration must articulate how the composite is interpreted and test whether its recommendations improve outcomes.
Heart-rate variability metrics add methodological complexity. An index computed from interbeat intervals depends on recording quality, artifact correction, posture, breathing, measurement period, and analysis conventions. A nightly proprietary recovery score may use HRV alongside temperature, sleep estimates, and historical averages. Without component-level documentation and validation, the score should be interpreted as a product-specific model output rather than a direct physiological diagnosis.
Sleep, Recovery, and the Limits of Readiness Scores
Sleep and recovery occupy a privileged place in digital coaching because devices permit repeated home monitoring. Yet wearables estimate sleep using combinations of movement and physiological signals, and their algorithms classify sleep-related states against training data and device-specific assumptions. Polysomnography, actigraphy, questionnaires, and vendor-derived sleep stages have different measurement targets and validation requirements.
The umbrella review by Doherty, C., Baldwin, M., Keogh, A., Caulfield, B., & Argent (2024) identified substantial variability in sleep-estimation accuracy. A device may approximate total sleep duration with some utility while performing less well at distinguishing wakefulness from sleep or classifying individual sleep stages. At the level of an individual athlete, a nightly difference between reported sleep scores is informative only relative to the known error, within-person variability, and the importance of the proposed change.
A recovery label can merge several mechanisms. Physiological adaptation, subjective fatigue, psychological distress, tissue injury, sleep opportunity, and perceived preparedness are correlated in some contexts but represent distinct constructs. A global readiness score may conceal the source of a change and its uncertainty. If a coach changes a session because a black-box score drops, the causal justification may be unclear even when the score has predictive association with a subsequent measurement.
Repeated monitoring can itself influence sleep behavior. An athlete may use the information to protect sufficient sleep opportunity or become preoccupied with meeting a numerical goal. These effects should be studied through validated measures of sleep-related beliefs, anxiety, and behavior. They should not be inferred from a wearable’s sleep score, which is an estimate of observed physiological and movement patterns rather than a direct measure of psychological wellness.
The Reardon, C. L., Gouttebarge, V., Kroshus-Havril, E., Miller (2026) situates sleep and mental-health concerns within the broader athlete-care system. Clinical decisions require appropriately qualified assessment. A consumer sleep score, however polished the interface, should not substitute for diagnosis of a sleep disorder or for individualized medical decision-making. Coaches and athletes can use monitoring to identify questions for discussion while maintaining the boundary between training support and clinical care.
Training Load, Internal Response, and Adaptation
A foundational distinction in sport monitoring separates external workload from internal response. External workload characterizes the performed task: distance, duration, speed, repetitions, intensity categories, or mechanical events. Internal response can include heart rate, ratings of perceived exertion, and other physiological or subjective reactions to the task. The relationship between the two depends on the athlete and context.
Bourdon, P. C., Cardinale, M., Murray, A., Gastin, P., Kellm (2017) offered a consensus-based framework for monitoring training load while emphasizing the need for purposeful measurement and interpretation. The framework organizes practice, but individual metrics still require construct validation. Passfield, L., Murias, J. M., Sacchetti, M., & Nicolò, A (2022) raised a direct conceptual challenge: commonly used load indices may not behave like valid measures of a training dose. The two contributions should be read together as a useful progression from widespread practice to scrutiny of its assumptions.
Training adaptation takes place across time. A signal that indicates a large response during a single session does not by itself identify whether the session will improve performance, contribute to accumulated fatigue, or provoke a later problem. Relationships may vary with recent history, session structure, sleep, nutrition, environmental stressors, previous injury, and sport-specific demands. Some variables can be causes, others effects, and still others markers of common unobserved factors.
Algorithmic monitoring often uses rolling averages, exponentially weighted summaries, or ratios of recent and longer-term workload. Any such model must state the time windows, thresholds, feature definitions, and intended outcome. Ratio-based measures may be unstable when denominators are small, and a warning can appear highly precise despite having limited validated predictive value. Prediction should be evaluated against alternatives including simple baselines, clinician judgment, and models using prior injury or training history alone.
A methodologically strong study asks whether a monitoring-derived decision rule improves a defined outcome when used prospectively. It records adherence to recommendations, reasons for nonadherence, competing interventions, and exposure to training. If an algorithm recommends rest and the athlete becomes less exposed to sport, a reduction in injury events may reflect reduced exposure rather than safer performance at equivalent exposure. This issue demonstrates why observational correlation and prospective intervention effectiveness must be treated separately.
Data Integrity, Missingness, and Measurement Inequality
Sport-monitoring datasets are rarely complete. Devices can be forgotten, incorrectly positioned, removed during competition, disconnected, or affected by synchronization failures. Participants may opt out of monitoring, and particular sport actions may be poorly captured by a selected sensor. Missingness can correlate with the outcome: an injured athlete may cease wearing a device, while an exhausted athlete may avoid completing a daily wellness survey. Treating those observations as missing at random can distort prediction.
Missingness has a psychological dimension. Athletes may change reporting behavior in anticipation of team selection, public comparison, or disciplinary consequences. A favorable self-report in a high-pressure environment may reflect truthful well-being, social desirability, strategic self-presentation, or confusion about confidentiality. Combining subjective data with wearables does not automatically correct these mechanisms; it may encode them into a model.
Population fairness requires measurement validation across relevant groups and contexts. Important strata may include sex, age, skin pigmentation, disability status, sport discipline, competition level, equipment configuration, and body morphology. Small subgroup sample sizes can yield imprecise error estimates. Data-quality audits should therefore report participant characteristics, device versions, completeness, error distributions, and uncertainty rather than assuming that one aggregate accuracy figure guarantees equity.
Koerber, D., Khan, S., Shamsheri, T., Kirubarajan, A., & Meh (2023) provide a concrete example of unresolved subgroup measurement concerns in optical heart-rate sensing. Kim, J.-H., Kim, J., Kang, H., & Youn, B.-Y (2025) connect sports AI to broader fairness and privacy questions. The empirical challenge lies in observing where unequal measurement, unequal model performance, and unequal institutional consequences occur, because these mechanisms require different remedies.
The integrity of longitudinal evidence also depends on provenance. Researchers should record sensor hardware, firmware, preprocessing, thresholds, timestamps, code versions, missing-data rules, site, sport, and exposure definitions. A model trained on data collected by one instrument may fail when deployed with another even when both display the same variable label. Dataset shifts can arise from a change of coaching strategy or competition rules as readily as from a change in hardware.
Psychometrics, Validation, and Scientific Interpretability
Digital coaching increasingly presents scores with names such as readiness, stress, resilience, fatigue, or mental sharpness. Such labels are psychological or physiological claims. Scientific validation requires a clear construct definition, a theoretical explanation for how signals relate to that construct, appropriate comparison measures, and evidence that the score supports the proposed interpretation. A prediction of tomorrow’s training completion is not automatically a measure of intrinsic motivation.
A validation program should combine analytical accuracy, reliability, construct coherence, consequential validity, and prospective utility. Analytical accuracy concerns agreement with a stated criterion. Reliability concerns consistency under comparable conditions. Construct coherence involves convergent and discriminant evidence. Consequential validity concerns what happens when the score is used in a particular decision context. Prospective utility concerns improvements relative to a suitable comparator. These properties can diverge.
For longitudinal sport applications, within-person and between-person variation must be modeled distinctly. An athlete’s deviation from their usual physiological range may be useful for understanding a current session even when comparing absolute scores across athletes is misleading. Aggregating repeated observations as if they were independent individuals inflates the apparent sample size. Model validation must also prevent features recorded after a target event from leaking into predictions made before it.
Interpretability includes both computational and practical dimensions. A transparent equation may still rely on a poorly justified construct. A sophisticated model can provide useful, calibrated probabilities even where its internal operations are difficult to summarize. Scientific explanations should specify input provenance, outcome definition, uncertainty, major predictors, intended population, and the limits of the recommendation. A readable interface is useful only when these foundations are defensible.
The weight of the evidence supports treating wearable-generated outputs as task-specific measurements and model-based inferences. Scientific credibility is earned through validation of each inferential step and ongoing assessment after deployment. The next problem is whether predictions based on those measurements can guide safer or more effective decisions.
Predictive Modeling and the Problem of Transportability
Algorithmic coaching can use linear statistical models, classification algorithms, tree ensembles, neural networks, time-series models, or combinations of these approaches. The scientifically relevant question begins with a specified prediction target and decision horizon, such as the probability of a defined injury during a subsequent exposure period or the expected response to a particular training plan. A generic claim that a system predicts performance has little evidentiary content until outcome, horizon, population, and comparison model are stated.
Retrospective prediction studies are particularly exposed to temporal leakage. A random split of individual training sessions may place observations from the same athlete in both training and test sets. A model can then exploit stable athlete characteristics or near-duplicate sequences rather than generalizable patterns. Data collected after the prediction time, including eventual treatment choices or labels assigned after an event, may also leak information into model inputs. A credible test separates earlier from later observations, accounts for team and athlete clustering, and ideally validates in a new institution or season.
Calibration and discrimination answer different questions. Discrimination concerns ranking of higher- and lower-risk cases. Calibration concerns correspondence between predicted probabilities and observed event frequencies. A model with high discrimination may still systematically exaggerate risk; a model with acceptable calibration may not separate individuals well enough for useful decisions. Neither metric measures whether adopting the model improves welfare or performance. For rare injuries, absolute risk, false-alarm burden, and positive predictive value depend on the event’s base rate.
Transportability poses special difficulties in sport. Training environments vary in calendar structure, competition density, support resources, playing style, and medical practices. Models built for one professional male football squad may not generalize to female athletes, adolescents, para athletes, or recreational runners. Shifts in coaching policy can change the data-generating process itself. A model that affects training also changes the distribution of future exposures on which the model was trained.
The clearest methodological principle is to validate at the level of the intended decision. If the system will recommend individual rest days in one league, evidence from an unrelated sport or from purely descriptive correlations is insufficient. The relevant comparison may be a simple rule based on recent workload, an experienced clinician’s judgment, or a conventional individualized training plan. Model complexity has no privileged status over demonstrated generalization and practical benefit.
Injury Prediction, Prevention, and Clinical Boundaries
Injury prediction is a major motivation for athlete-monitoring technology because injury has severe personal and institutional consequences. Yet injury events may be heterogeneous: acute contact injuries, noncontact musculoskeletal injuries, overuse syndromes, and symptom-based absences have different causal structures. Studies must operationalize the injury outcome, athlete-time at risk, exposure denominator, and appropriate follow-up. A prediction about future injury cannot by itself identify a safe preventive intervention.
A systematic review published in 2022, Just How Confident Can We Be in Predicting Sports Injuries?, evaluated thirty studies and 204 injury prediction models. It found that no included study externally validated a prediction model and that most models had high or unclear risk of bias. The authors concluded that none of the reviewed models could be recommended for practice at that time. This is an important baseline for judging claims that injury prediction had already become a settled clinical technology.
More recent work has expanded the literature without dissolving the central concerns. Lee, C.-H., Chang, M.-H., & Li, Z.-H (2026) systematically reviewed eighteen empirical AI injury-prediction studies and reported that only one was judged low risk of bias under PROBAST; internal validation and analysis-domain problems were common. The review also noted limited demographic representation. These findings are about study quality, not evidence that artificial intelligence is incapable of useful forecasting.
The causal bridge from a forecast to prevention requires intervention evidence. A risk score is a description of estimated risk under a set of existing conditions. Reducing training might change the exposure distribution, but it might also interfere with desired adaptation. A modified technique, clinical assessment, or recovery action may or may not address the mechanism that produced the elevated prediction. Prospective trials must record what recommendation was delivered, who acted on it, what alternative action was available, and whether the outcome improved.
Clinical return-to-sport decisions require additional safeguards. A digital score should be interpreted alongside examination, injury history, symptoms, physical performance, and appropriate professional judgment. Automated alerts must have clear escalation pathways and delineated authority. Systems should disclose whether they are intended for wellness assistance, training programming, or regulated medical decision support, because the evidentiary and safety requirements differ.
A scientific synthesis therefore treats AI injury prediction as an active research program with important methodological weaknesses. The field needs independent external validation, calibrated probabilities, explicit exposure denominators, adequate representation of different athlete groups, prospectively evaluated action rules, and transparent tracking of unintended consequences.
Adaptive Recommendations and Personalization
Personalization is often presented as the defining promise of digital coaching. In operational terms, it means that the recommendation function changes with athlete-specific information: training history, observed response, preferences, constraints, injury context, goals, or current state. A system may personalize fixed thresholds, use a learned predictive model, or adapt recommendations sequentially. These approaches differ in what they learn and how they should be evaluated.
Individualized recommendations require data sufficient to distinguish stable athlete attributes from short-term fluctuations. A model that learns from many athletes may estimate average effects reliably but remain poorly calibrated for a person whose physiology or schedule differs from the training population. Conversely, a purely individual model may have too few observations to identify rare outcomes. Hierarchical methods and partial pooling offer one conceptual response, but the benefits must be established with genuine out-of-sample evaluation.
Time-varying personalization creates feedback loops. The algorithm chooses a training action, that action changes physiological state and future measurements, and later recommendations are conditioned on the resulting observations. If a system only learns from its own previous choices, it may have little evidence about unchosen alternatives. Observational logs cannot automatically distinguish response to the recommendation from selection into that recommendation. A carefully designed experiment needs variation in actions, sufficient safety oversight, and predefined outcomes.
A digital coach also needs to represent multiple objectives. Short-term speed, adherence, recovery, enjoyment, development, injury prevention, and mental well-being are not the same optimization target. Athletes and coaching teams may legitimately place different weights on them, and those weights may change during rehabilitation or a major competition. A scoring function that silently collapses all goals into one proprietary index embeds institutional values as technical parameters.
The best research questions are thus not simply whether personalization works, but for whom, in what sport, under which constraints, through what mechanism, and against which comparator. Transparent preference setting and athlete participation in goals are psychologically relevant components of personalization. A system that adapts numerical thresholds while ignoring the person’s endorsed objectives remains only partially individualized.
Digital Interventions and Causal Identification
Intervention studies must distinguish technology exposure from technology effectiveness. Buying a wearable, opening an application, reading a recommendation, and changing a training plan are successive events. Attrition and nonadherence can select a highly motivated subgroup, making outcomes among continuing users look favorable even when the initial deployment has little average benefit. Evaluations should follow participants assigned to intervention conditions and document engagement as a process measure rather than as a substitute for results.
Randomized controlled trials can identify average effects of access to a digital coaching package when randomization, measurement, adherence, and follow-up are adequate. However, a package may combine tracking, coaching attention, social support, reminders, and goal setting. A trial showing a benefit for the package does not isolate algorithmic personalization. Factorial trials can separate components; micro-randomized trials can examine proximal responses to repeated recommendations; cluster trials can reduce contamination when an entire coaching staff changes practice.
Nonrandomized designs have a role when randomization is infeasible, but they require stronger assumptions. A team adopting wearables during a period of broader organizational change may improve for reasons unrelated to monitoring. Before-and-after performance comparisons are vulnerable to seasonal trends, coaching turnover, regression to the mean, and changes in competition. Difference-in-differences, interrupted time series, and carefully specified causal models may improve inference when their identifying assumptions are plausible and tested.
In studies of competitive outcomes, the endpoint itself needs care. Win–loss results are influenced by opponent quality and randomness; training performance is affected by fatigue and context; long-term skill development requires repeated assessment. An adaptive coaching system may optimize a laboratory exercise without improving competitive transfer. Outcomes should therefore include behavior, learning, performance, health, and potentially adverse psychological consequences.
The wider digital-activity literature indicates that wearables can support physical-activity change in some populations, but this cannot be equated with proof of elite sporting benefit. A systematic review of older adult tracker interventions, for example, studies exercise time, step counts, and function rather than tactical or competition outcomes. Extrapolation from these studies should be explicit and modest. Sport-specific intervention trials remain a priority.
Trust Calibration, Algorithm Aversion, and Overreliance
The central psychological problem of algorithmic advice is calibrated reliance: using the recommendation when it is sufficiently informative and resisting it when it is irrelevant, unreliable, or outside its validated domain. Trust should be examined in relation to measurable system performance and context. High subjective trust can coexist with poor decisions; low subjective trust can lead to unnecessary rejection of useful information.
Lee, J. D., & See, K. A (2004) developed a framework for appropriate reliance on automation, identifying the roles of system performance, process, purpose, and task context. Applied to sport, this suggests that an athlete or coach should know what the system is predicting, how stable its predictions are, and under which circumstances it should be overridden. A probability without information on uncertainty, calibration, and target population is difficult to use appropriately.
The experimental literature challenges any simple thesis that people always mistrust or always accept algorithms. Dietvorst, B. J., Simmons, J. P., & Massey, C (2015) demonstrated reluctance after visible algorithmic mistakes in forecasting tasks. Logg, J. M., Minson, J. A., & Moore, D. A (2019) found greater reliance on algorithmic than human advice in several other tasks. The differences show why task design, expertise, error exposure, source framing, and accountability must be measured. They also warn against importing laboratory conclusions wholesale into high-stakes athlete management.
Overreliance can occur when recommendations carry an aura of numerical objectivity. Underreliance may arise when a system makes salient errors, conflicts with expertise, or offers opaque explanations. Both forms of miscalibration can be studied experimentally using advice-taking paradigms in which participants decide whether to follow model forecasts with varying known levels of reliability. Actual decisions, not merely attitude questionnaires, should be collected.
Wunderlich, F., & Reis, M (2026) argue that sport provides valuable conditions for studying human–AI trust because repeated choices and observable performances are available. Their article identifies a research opportunity; it does not furnish broad causal evidence that athletes already use AI advice appropriately. Direct sport experiments should consider professional identity, responsibility for losses, pressures from team management, and social consequences of overriding a model.
Automation Bias, Deskilling, and Human Oversight
Automation bias concerns decision errors associated with excessive reliance on automated cues, including failures to detect errors and omissions in a system’s output. Parasuraman, R., & Riley, V (1997) emphasized that misuse and disuse emerge from system reliability, workload, design, and organizational choices. Sport-specific analogues include taking a fatigue warning as dispositive without considering the athlete’s current explanation, or ignoring a valid warning because previous alerts were frequent and unhelpful.
Human oversight is often invoked as if the presence of a coach automatically guarantees accountability. Effective oversight requires access to relevant information, authority to revise decisions, time to deliberate, knowledge of model limits, and an organizational climate that allows justified overrides. If a coach is formally responsible but institutionally penalized for disagreeing with a system, the nominal human-in-the-loop arrangement may supply little actual control.
Deskilling is an empirically investigable possibility. Coaches who continually outsource interpretation might practice some diagnostic judgments less often; alternatively, computational summaries might free attention for more complex interpersonal coaching. Both trajectories are plausible. Longitudinal comparisons should measure retained professional judgment, adaptation to system failure, and transfer to unsupervised settings. Short-term user satisfaction cannot resolve the issue.
Explanations should be evaluated by their effect on decisions, not only by whether participants report liking them. A model may provide feature rankings that appear persuasive while failing to convey causal meaning. For instance, training load may correlate with injury through several pathways, some of which involve response to earlier symptoms. Explanation quality therefore requires accurate disclosure of what the model did and a clear distinction between predictive association and causal mechanism.
Safety-critical decision processes should include escalation thresholds, documentation of overrides, uncertainty ranges, audit logs, and procedures for device or model failure. The purpose of these measures is to support justified human judgment and traceable responsibility. Their effectiveness should be evaluated with concrete error scenarios rather than assumed from a governance checklist.
Athlete Data Rights, Consent, and Surveillance
Wearable data acquire institutional significance when they are accumulated, linked across contexts, and used to evaluate a person. Training information may reveal location, routines, physical limitations, sleep patterns, or sensitive health-related characteristics. A sport organization can have legitimate reasons to examine some records while also possessing power over selection, contracts, and access to opportunities. Research on adoption therefore requires data governance as well as technical reliability.
Consent in competitive sport must be evaluated in its social context. An athlete may formally agree to tracking while believing refusal would jeopardize selection. Researchers should distinguish willingness, comprehension, perceived voluntariness, and available alternatives. Consent practices should specify purpose, recipients, retention time, secondary analyses, withdrawal procedures, and whether a coach can infer sensitive information from a derivative score. Longitudinal location data can be reidentifiable when combined with competition records.
The same signal may circulate through several institutional pathways. A daily wellness report could inform a confidential discussion, a team-wide performance dashboard, a scouting decision, or a commercial product-development dataset. Each pathway changes the interests and consequences involved. Purpose limitation and role-based access control reduce risks of using data collected for care as hidden instruments of employment or selection discipline.
Kim, J.-H., Kim, J., Kang, H., & Youn, B.-Y (2025) reviewed ethical questions arising from AI in sport, including privacy, fairness, and accountable decision-making. Their scoping synthesis maps the topics, while the effects of particular governance practices require empirical evaluation. Athletes can report understanding and perceived control, while organizations can document access, incidents, and restrictions on secondary use.
Data-rights analysis extends to technology suppliers. Proprietary platforms may control storage location, model updates, data export, and continuation of access when a team changes provider. Researchers should document vendor dependence and auditability as conditions that shape the longevity and reproducibility of observations. A longitudinal scientific study must remain interpretable when a commercial interface changes or disappears.
Fairness, Age, Disability, and Competitive Context
Algorithmic fairness begins with adequate identification of the groups and decision contexts in which a model will be used. Sporting datasets can be concentrated in particular leagues, teams, and competition levels. Findings from a single professional roster should not be generalized across sexes, adolescents, para athletes, or international competition settings without validation. Training history and access may also differ systematically across groups, producing dataset imbalances independent of athletic potential.
A system can produce inequitable consequences through several mechanisms: a sensor may measure some people less accurately; a model may be less calibrated for certain populations; an arbitrary threshold may be applied across different baseline distributions; or an institution may use a prediction to justify unequal opportunity. Measurement error, predictive error, and distributional harm require separate tests.
Koerber, D., Khan, S., Shamsheri, T., Kirubarajan, A., & Meh (2023) found mixed evidence about heart-rate measurement differences across skin tones in a limited literature. That warrants device-specific subgroup validation and transparent uncertainty. It does not establish that all wrist sensors share a universal bias magnitude.
For young athletes, additional issues arise from developmental change and guardianship. Models trained on adults may misinterpret maturation or recovery. A youth athlete may have limited freedom to refuse monitoring. Appropriate safeguards include age-specific validation, proportionality, protection of sensitive information, and restrictions on automated decisions with serious developmental consequences.
For para athletes, conventional activity measures may require adaptation to mobility pattern, equipment, impairment, and sporting rules. A step-count goal, for instance, cannot serve as a universal measure of activity. Researchers should define meaningful outcomes with participants and sport specialists rather than forcing incomparable physical patterns into a common template.
Scientific fairness assessment needs sufficient subgroup sample sizes, uncertainty intervals, and analysis of downstream actions. Equal predictive error does not guarantee equitable results when exposure to selection decisions differs. The most consequential evaluation follows the path from sensing to recommendation and then to opportunity, health, performance, and welfare.
Safety, Accountability, and Lifecycle Governance
Responsible digital coaching is a lifecycle problem. A system can perform adequately in development and become unreliable after changes in sensors, competition conditions, populations, or training strategies. Once deployment begins, organizations need procedures for detecting model drift, reviewing errors, reporting incidents, evaluating overrides, and suspending unsafe applications.
The Tabassi, E (2023) provides a cross-sector approach for governing, mapping, measuring, and managing AI risk. It is a general framework, not a sport-specific certificate. In sport, it helps structure responsibility for evaluation, athlete consultation, data security, and responses to system errors.
Alecu, S., & Onea, G. A (2026) proposed six conceptual layers of responsible sports AI decision support: data and context, prediction, decision translation, individualization, human oversight, and deployment monitoring. Their framework is conceptual rather than a validated certification instrument. It identifies points where scientific evidence must be documented when a prediction becomes an action.
Accountability requires naming actors responsible for choices. Coaches may control training adjustments; clinicians hold defined clinical responsibilities; sporting institutions set selection and data-use policies; technology providers have responsibilities for documented product performance and maintenance. Oversight requires relevant information, actual authority to revise decisions, time to deliberate, and a culture allowing justified overrides.
Incident monitoring should document actual harms and near misses, including inappropriate rest recommendations, unauthorized disclosure of sensitive information, misleading sensor readings, and overreliance on automated assessments. Reports should record system version, input context, decision sequence, participants, and consequences. Product updates that affect decisions require evaluation of the affected functions.
Operational indicators include the fraction of consequential decisions with recorded rationale, calibration checks across groups and seasons, documented override rates, review of significant incidents, athlete access requests, and prospective evaluations of high-stakes use cases. Their presence indicates monitoring capacity, while benefits must still be tested empirically.
Methods for Evaluating Human–Algorithm Collaboration
Studying digital coaching requires an evaluation design matched to the level at which technology acts. At the signal level, investigators compare a device with an appropriate reference instrument under controlled and sport-realistic conditions. At the prediction level, they examine temporal, external, and subgroup validation with prespecified metrics. At the decision level, they compare choices made with and without algorithmic input. At the organizational level, they evaluate athlete outcomes, confidentiality, authority, and sustainability.
A single experiment rarely identifies all these processes. A staged research program first establishes measurement properties, then tests prediction out of sample, studies advice-taking behavior, and finally examines decisions in prospective field settings. This structure prevents the launch of expensive trials before the underlying measurements or predictions are demonstrated to be reliable.
For binary injury prediction, research should define outcome, time horizon, sensitivity, specificity, probability calibration, and predictive value at relevant base rates, with uncertainty and credible baselines. For continuous forecasts, mean absolute error, calibration, and meaningful performance tolerances may be informative. For collaboration, outcomes include decisions taken, time to action, error detection, overrides, workload, calibration of trust, and ability to perform when automated recommendations are unavailable.
Prospective designs must state precisely what counts as an algorithmic intervention. Displaying a dashboard does not ensure that its information influences action. Investigators can record when recommendations were viewed, whether they were accepted, altered, or rejected, and what happened afterward. Repeated decisions require multilevel analyses accounting for clustering within athletes, coaches, and teams.
Mixed-methods research is especially important when system effects depend on organizational interpretation. Interviews can document perceived pressures, confusion about data ownership, and changes in trust; observations can reveal how coaches negotiate recommendations; logs can establish timing and actions. Contradictions between reported attitudes and observed behavior may themselves be informative findings.
Studies should disclose software and sensor versions, missing-data rules, conflicts of interest, data-sharing constraints, and limits on generalization. A model validated through random partitioning of sessions from the same roster should be described as internally validated. A technology demonstration should be labeled accordingly and should not substitute for a prospective test of effectiveness.
Critical Scientific Discussion and Rival Explanations
Digital-coaching studies must address mechanisms that can imitate genuine technological benefit. Athletes may train more diligently when observed, coaches may increase attention during a new program, and monitoring may be introduced alongside better facilities or staffing. People who adopt wearables voluntarily may already be more motivated. Improved outcomes after introduction therefore cannot automatically be attributed to algorithmic personalization.
Measurement sophistication can also be mistaken for causal explanation. A high-dimensional dataset may describe an athlete in extraordinary detail while leaving the determinants of success uncertain. Physiological signals may correlate with later injury through training exposure, previous symptoms, tactical assignment, or common unobserved factors. A larger computational model cannot by itself solve these identification problems.
The promise of individualized support must be tested against possible overstandardization. A platform may adapt numerical targets while limiting how athletes express the meaning of soreness, confidence, or readiness. Effective support may depend on preserving contextual interpretation and a real opportunity to disagree. Sport psychology offers instruments and qualitative methods for studying whether this opportunity is genuinely available.
The motivation literature makes a further prediction: feedback can be experienced as informational or controlling depending on context. Continuous scores could help athletes regulate practice while also generating persistent evaluative pressure. Randomized comparisons of informational framing, privacy policies, and athlete choice can distinguish effects of the measurement from those of its social use.
The trust literature illustrates how rapidly intuitive generalizations can fail. Wunderlich, F., Reis, M., Kunde, W., & Memmert, D (2026) conducted a sport-related judgment experiment about Olympic Games group-match outcomes. Participants received numerical estimates presented as coming from either AI or a human expert. Their forecasts shifted toward these supplemental estimates; the anchoring effect did not significantly differ by supposed source, while confidence was higher when the supplemental estimate was labeled AI. This shows task-specific acceptance of AI-framed information. It does not establish that AI coaches improve training, injury prevention, or athlete welfare.
Newer frameworks also require careful status labels. Alecu, S., & Onea, G. A (2026) provide a conceptual six-layer account of responsible sports AI. It can guide studies of implementation but is not yet a validated scoring scale. Bullock, G. S., Mylott, J., Hughes, T., Nicholson, K. F., Ri (2022) and Lee, C.-H., Chang, M.-H., & Li, Z.-H (2026) identify important weaknesses in the validation of injury prediction. Evidence of a promising technique must remain separate from evidence that a complete decision system reliably improves outcomes.
These rival explanations support a unified scientific criterion. A digital-coaching intervention should show that the information is accurately measured, that derived predictions generalize, that people interpret and act on advice appropriately, and that the resulting decisions improve meaningful outcomes without imposing disproportionate costs. Strong evidence at one stage is a reason to evaluate the next, rather than a substitute for that evaluation.
Artificial Era and Human–Artificial Sporting Relations
Within the intellectual horizon of the Artificial Era, digital coaching is a concrete setting in which human practice is reorganized around persistent computational representations and adaptive recommendations. Its historical significance concerns the redistribution of memory, comparison, information processing, and decision support. Sport makes these changes unusually visible because practice decisions recur and many consequences can be measured.
Scientific constructs must remain operational. Artificial agency in this monograph denotes observable system influence on action: maintaining a training plan, updating predictions, requesting information, proposing modifications, or performing authorized adjustments. Human autonomy denotes actual and experienced opportunities to endorse, negotiate, and revise actions. These constructs can coexist in a collaborative sporting environment and can be separately evaluated.
Psychology of Artificial can examine systems that retain histories and update policies. Psychology of Human–Artificial Relations can investigate whether repeated interactions change trust, perceived competence, expectations, dependence, or patterns of cooperation. Human Factors and Engineering Psychology can study workload, uncertainty displays, shared control, and error recovery. Sport and Performance Psychology supplies outcomes of skilled execution, training, development, and athlete welfare.
The philosophical concept of Artificial as a new historical order belongs to a wider theoretical framework. The present empirical inquiry concerns observable system functions and relations. Assertions about computational selfhood or phenomenal experience would require independent arguments and evidence. A prospective longitudinal research design can investigate whether a persistent coaching agent reliably adjusts to athlete goals, learns from correction, and maintains interpretable records of joint decisions.
A longitudinal human–artificial partnership might improve continuity of advice across a season or amplify dependence on automated evaluation. Both possibilities are substantive research hypotheses. Meaningful outcomes include sustained self-regulation, accuracy of joint decisions, the ability to correct erroneous recommendations, relational trust, and safe transitions when a system is unavailable.
Open Questions and Research Priorities
The first unresolved question is whether continuous monitoring adds benefit beyond high-quality intermittent assessment and attentive coaching. High-frequency signals may improve the timing of selected decisions, but they can also increase false alerts and demands for attention. Trials should compare monitoring intensity while holding coaching support comparable.
The second is whether personalized recommendations outperform simpler and transparent decision rules. Athlete-specific learning may capture meaningful variation, yet sparse records and limited variation in prior choices can obscure what different prescriptions would have achieved. Studies should separate personalized prediction from personalized action and report incremental value against credible baselines.
The third concerns performance under system failure. Digital coaching may support learning when available but create vulnerabilities if athletes become dependent on cues. Retention and transfer tests can determine whether athletes preserve relevant skills without assistance. Error scenarios can reveal whether coaches recognize recommendations made outside a model's validated conditions.
The fourth concerns meaningful consent in institutions with uneven power. Will athletes report symptoms accurately when performance data affect selection? Do opt-out policies protect opportunity? Can privacy protections be audited across teams and vendors? These questions need longitudinal investigation of behavior, perceived autonomy, and actual governance practices.
The fifth concerns equitable performance across populations. Measurement validation, subgroup calibration, and assessment of real downstream actions should be integrated. The consequences of false alerts and overlooked risks can differ with age, disability, competition level, and availability of medical care.
Finally, the field needs cumulative replication across independent teams, seasons, and technologies. Robust research combines external model tests, prospectively evaluated decision rules, qualitative evidence about institutional use, and post-deployment monitoring of failures or drift. These linked methods offer a route from attractive demonstrations to defensible scientific knowledge.
Conclusion
Digital coaching, wearable monitoring, and algorithmic agency in sport form an integrated psychological and socio-technical research object. The sequence begins with physical and behavioral measurement, proceeds through psychological interpretation and statistical prediction, enters human or automated decisions, and produces observable outcomes for athletes and institutions. Each link has its own evidentiary requirements.
Foundational literatures in motor learning, motivation, training science, psychometrics, and human–automation interaction provide important explanations of feedback, self-regulation, reliance, and expertise. Contemporary wearable reviews show conditional usefulness and variation in measurement quality. Reviews of injury prediction expose problems in external validation, reporting, and causal translation. New experiments demonstrate that people can respond to AI-framed sporting advice, but evidence concerning sustained effects on coaching, health, and performance remains limited.
The scientifically defensible digital coach is a distributed decision process with purposeful measurement, calibrated uncertainty, appropriate human authority, meaningful athlete participation, and accountable lifecycle governance. Its value must be established through outcomes that matter to sport and the people who practice it. The central agenda is to determine when computational guidance improves sporting decisions while sustaining competence, safety, autonomy, and athlete rights.
References
- Akenhead, R., & Nassis, G. P. (2016). Training load and player monitoring in high-level football: Current practice and perceptions. International Journal of Sports Physiology and Performance, 11(5), 587–593. https://doi.org/10.1123/ijspp.2015-0331
- Alecu, S., & Onea, G. A. (2026). From prediction to decision support: A critical review and six-layer framework for responsible artificial intelligence in sport science. AI, 7(9), 343. https://doi.org/10.3390/ai7090343
- Bandura, A. (1977). Self-efficacy: Toward a unifying theory of behavioral change. Psychological Review, 84(2), 191–215. https://doi.org/10.1037/0033-295X.84.2.191
- Bourdon, P. C., Cardinale, M., Murray, A., Gastin, P., Kellmann, M., Varley, M. C., Gabbett, T. J., Coutts, A. J., Burgess, D. J., Gregson, W., & Cable, N. T. (2017). Monitoring athlete training loads: Consensus statement. International Journal of Sports Physiology and Performance, 12(Suppl. 2), S2-161–S2-170. https://doi.org/10.1123/IJSPP.2017-0208
- Bullock, G. S., Mylott, J., Hughes, T., Nicholson, K. F., Riley, R. D., & Collins, G. S. (2022). Just how confident can we be in predicting sports injuries? A systematic review of the methodological conduct and performance of existing musculoskeletal injury prediction models in sport. Sports Medicine, 52(10), 2469–2482. https://doi.org/10.1007/s40279-022-01698-9
- Čokorilo, N., Manolopoulos, N., Matijević, T., & Rajović, R. (2026). How valid are wearable devices in team sports? A systematic review. Sports, 14(7), 264. https://doi.org/10.3390/sports14070264
- Dietvorst, B. J., Simmons, J. P., & Massey, C. (2015). Algorithm aversion: People erroneously avoid algorithms after seeing them err. Journal of Experimental Psychology: General, 144(1), 114–126. https://doi.org/10.1037/xge0000033
- Doherty, C., Baldwin, M., Keogh, A., Caulfield, B., & Argent, R. (2024). Keeping pace with wearables: A living umbrella review of systematic reviews evaluating the accuracy of consumer wearable technologies in health measurement. Sports Medicine, 54, 2907–2926. https://doi.org/10.1007/s40279-024-02077-2
- Kim, J.-H., Kim, J., Kang, H., & Youn, B.-Y. (2025). Ethical implications of artificial intelligence in sport: A systematic scoping review. Journal of Sport and Health Science, 14, 101047. https://doi.org/10.1016/j.jshs.2025.101047
- Koerber, D., Khan, S., Shamsheri, T., Kirubarajan, A., & Mehta, S. (2023). Accuracy of heart rate measurement with wrist-worn wearable devices in various skin tones: A systematic review. Journal of Racial and Ethnic Health Disparities, 10(6), 2676–2684. https://doi.org/10.1007/s40615-022-01446-9
- Lee, C.-H., Chang, M.-H., & Li, Z.-H. (2026). From data to prevention: A systematic review of artificial intelligence applications in sports injury prediction. Technology and Health Care, 34(6), 666–684. https://doi.org/10.1177/09287329261455797
- Lee, J. D., & See, K. A. (2004). Trust in automation: Designing for appropriate reliance. Human Factors, 46(1), 50–80. https://doi.org/10.1518/hfes.46.1.50_30392
- Logg, J. M., Minson, J. A., & Moore, D. A. (2019). Algorithm appreciation: People prefer algorithmic to human judgment. Organizational Behavior and Human Decision Processes, 151, 90–103. https://doi.org/10.1016/j.obhdp.2018.12.005
- Luteberget, L. S., & Gilgien, M. (2020). Validation methods for global and local positioning-based athlete monitoring systems in team sports: A scoping review. BMJ Open Sport & Exercise Medicine, 6(1), e000794. https://doi.org/10.1136/bmjsem-2020-000794
- Parasuraman, R., & Riley, V. (1997). Humans and automation: Use, misuse, disuse, abuse. Human Factors, 39(2), 230–253. https://doi.org/10.1518/001872097778543886
- Passfield, L., Murias, J. M., Sacchetti, M., & Nicolò, A. (2022). Validity of the training-load concept. International Journal of Sports Physiology and Performance, 17(4), 507–514. https://doi.org/10.1123/ijspp.2021-0536
- Reardon, C. L., Gouttebarge, V., Kroshus-Havril, E., Miller Aron, C., Bahr, R., Blauwet, C., Castaldelli-Maia, J. M., Cheng, C., Currie, A., Derevensky, J. L., Edwards, C., Fussek, S., Gorczynski, P., Grandner, M., Han, D. H., Hitchcock, M. E., Lu, F., Massey, A., McDuff, D.,. Hainline, B. (2026). Mental health in elite athletes: International Olympic Committee consensus statement (2026). British Journal of Sports Medicine, 60(15), 1083–1130. https://doi.org/10.1136/bjsports-2025-111514
- Ryan, R. M., & Deci, E. L. (2000). Self-determination theory and the facilitation of intrinsic motivation, social development, and well-being. American Psychologist, 55(1), 68–78. https://doi.org/10.1037/0003-066X.55.1.68
- Tabassi, E. (2023). Artificial intelligence risk management framework (AI RMF 1.0) (NIST AI 100-1). National Institute of Standards and Technology. https://doi.org/10.6028/NIST.AI.100-1
- Wulf, G. (2013). Attentional focus and motor learning: A review of 15 years. International Review of Sport and Exercise Psychology, 6(1), 77–104. https://doi.org/10.1080/1750984X.2012.723728
- Wunderlich, F., Reis, M., Kunde, W., & Memmert, D. (2026). Trusting the AI rather than the expert? Evidence from sports-related assessments in relation to the Olympic Games 2024. Psychological Research, 90, Article 195. https://doi.org/10.1007/s00426-026-02409-z
- Wunderlich, F., & Reis, M. (2026). Studying trust in AI in sports: An interdisciplinary primer. Psychology of Sport and Exercise, 88, 103274. https://doi.org/10.1016/j.psychsport.2026.103274