Sport has an
interpretation
problem.

Artificial intelligence will democratise and stratify coaching at the same time. It puts something close to expert planning support in a volunteer's pocket, and it hands the best-resourced programmes a compounding advantage through proprietary longitudinal data, the ability to commission models against it, specialist staff who can reject a bad output, and equipment that generates the next dataset. Both forces operate at once, and the gap can widen while everyone improves. This paper defines both terms in what reaches the athlete rather than what is purchased, locates the question within the established treatment of system strength in sporting fairness, and takes the availability provision of the athletics shoe rules as the transferable precedent, since a rule about open availability is enforceable where a rule about conferred advantage is not. Six forms of fairness are set out, of which the epistemic question of whose coaching knowledge gets encoded receives least attention and may matter most. Five institutional choices follow for governing bodies. The author records that he found no evidence AI improves coaching outcomes at grassroots level, and nobody claiming to have it.

In long-term endurance development the main danger from AI load tools is apparent precision rather than missing data. Load monitoring has defensible foundations in the distinction between internal and external load, in intensity distribution and in session rating of perceived exertion, and around them has grown a layer of derived metrics whose predictive claims were criticised in the same journals that popularised them. This paper reviews that criticism, including the mathematical coupling that produces spurious correlation in a widely used workload ratio, the finding that screening is unlikely ever to predict injury, and the base-rate arithmetic showing that a young athlete seeing a red warning is more likely to be well than not. It traces a six-step chain from contested metric to a colour that a 15-year-old acts on before speaking to her coach, and notes that an honest model shown through a misleading interface still misleads. Five design principles follow: describe before predicting, label classifications as descriptive and non-clinical, make uncertainty visible, leave the decision with the coach, and make no injury-risk claim without validated evidence for that population.

Fairness for AI tools used on youth pathways has four dimensions, and only three are covered by general AI ethics. The fourth, developmental fairness, asks whether an output prematurely narrows the options of someone who has not finished growing, and it is distinctive because it concerns time. This paper shows that a tool can be statistically accurate on average, drawn from representative data and fairly allocated, and still be unfair to the late maturer, the relatively young and the girl whose physiology the model was never trained on, because models trained on selection history learn the relative age and maturation biases that history contains. It draws on the evidence that average accuracy describes whoever dominates the test set, that proxy targets import bias directly, and that competing fairness conditions cannot hold together where base rates are themselves shaped by maturity. Four dimensions, each with a failure mode and a pre-deployment test, are followed by a six-step procurement test and seven recommendations for governing bodies. Its stated position on error asymmetry is that wrongful exclusion costs more, because it removes the athlete from the evidence altogether.

A child cannot give informed consent to uses of their data that nobody present can describe, and in a selective pathway refusal is never free, because declining measurement carries a cost to selection. This paper argues that AI changes the stakes by converting measurements into inferences, and that inferences such as maturity status, injury flags and potential ratings are more consequential, less visible and less protected than the data they were drawn from. It separates measured, derived and inferred data, showing that inferred data is rarely shown to the athlete and travels furthest between organisations, and it removes the usual fallback by noting how readily small squads are re-identified from attributes that are already public. The remedy proposed is structural rather than a better consent form: collect only what a named coaching or safeguarding purpose requires, hold inferred data apart under restricted access, set retention periods that actually expire, and give every athlete measured before 18 a formal review at adulthood. Consent becomes a continuing process. The framework is untested in any pathway, and the practical obstacles to a review at 18 are conceded.

Coaching expertise is formed in the gap between a plan and an athlete, through deciding, observing the consequence and adjusting. An AI tool that makes the planning decision closes that gap, and it does so most completely for the coaches who have least judgement to spare. This paper distinguishes chat-based tools from measurement tools, since the former draft the coach's own product in fluent prose that invites acceptance, and it maps them onto the functions of an automated system to show that the choice of plan, the function most tied to judgement, is the one most exposed. Its central framework sets out four roles AI can occupy in that decision, being reference, adviser, co-planner and substitute, with the role set by the user rather than the software and a documented tendency to drift from adviser to substitute on busy weeks. Four design principles and a staged permission scheme follow, tying permitted roles to stages of coach development and excluding the substitute role from work with young athletes. The transfer from aviation and process control is analogy, and the paper asks that it be tested.

A predictive model learns the pathway that produced its training data, and in youth sport that pathway is already bent by relative age, early maturation and early selection. Pointed at 12-year-olds, such a model reproduces those distortions with more confidence and less visibility than a human selector would. This paper sets the evidence on athlete development, where the overwhelming majority of elite pathways are non-linear and junior and senior predictors run in opposite directions, against the assumptions of supervised learning, which require a stable world, a representative sample and a meaningful label. It identifies three routes of inherited error in drift, sampling and labels, noting that using selection as a proxy for talent writes the bias into the target itself. It then addresses performativity, since a score that shapes a deselection has become part of the treatment, and a model retrained on the result measures its own influence and calls it accuracy. Five reporting rules follow, covering current status, uncertainty, maturation context, individual change over time, and no verdict below mid-adolescence. Procurement, not the individual club, is identified as the practical lever.