Paper one of six in the Athleet.AI Working Papers on AI in coaching and athlete development, written eight weeks after conversational AI reached the public and long before most pathways had considered what it would mean for selection.
What does a talent model actually learn?
The pathway that produced its training data. In youth sport that pathway is already bent by relative age, by early maturation and by early selection, so a model trained on who was picked learns who tends to get picked. Selection history is a record of maturity as much as of talent.
The paper followed an invented composite it called Sam. At 13, Sam trains in a regional pathway, born in late summer, among the smallest in the squad, and two years into the sport after several seasons of swimming and athletics. The pathway has trialled a dashboard that blends sprint times, a jump test, match minutes and a coach rating into a single potential index. Sam sits in the bottom quartile. At the end-of-season review the index is one of four items on the page and the only one expressed as a number. Sam is released. Two years later, after a growth spurt and a season at a local club, Sam is faster than most of the retained group, and nobody in the pathway is looking. The dashboard holds no record that Sam ever improved.
What does the development evidence say about linear progress?
That it is the exception. A survey of 256 elite athletes across 27 sports found 83.6 per cent had non-linear pathways, with fewer than seven per cent showing pure junior-to-senior linearity. A model that assumes smooth progression is therefore assuming the exception.
Worse for prediction, junior and senior predictors run in opposite directions. A meta-analysis covering 71 study reports and 9,241 athletes found that adult world-class athletes had done more coach-led practice in other sports, started their main sport later, accumulated less practice in it and progressed more slowly early on. Juniors reversed on every count. A second review across 6,096 athletes, including 772 of the world’s top performers, associated varied early practice with slower initial progress and more sustainable long-term development.
The research base under talent identification is also thinner than its use suggests. One review found only 20 studies meeting inclusion criteria, 60 per cent concerned with physical profiles and 65 per cent drawn from male samples. Another concluded that talent is complex and largely misunderstood, and recommended against talent identification and selection at younger ages.
Where do the modelling assumptions break?
Supervised learning assumes the world holds still, that the training sample represents the population, and that the label means what it claims to mean. Youth sport strains all three.
Errors are inherited three ways. Drift, because the pathway itself changes over time. Sampling, because selected athletes dominate the record and those who leave are never followed. Labels, because selection is used as a proxy for talent, which bakes the bias into the target the model is trained to hit. The health care parallel is instructive: an algorithm that used costs as a proxy for health need produced racially biased results, and nobody had to intend it.
There is a further problem peculiar to prediction that gets acted upon. A score that shapes a deselection has become part of the treatment. The athlete marked low-potential receives less coaching, fewer minutes and lower expectations, and duly becomes lower-potential. Parents, coaches and athletes each carry that forward. A model retrained on the results will measure its own influence and call it accuracy, growing more confident over time without growing more accurate.
What is the honest role for AI in a talent pathway?
Description of where an athlete is now, with uncertainty shown and maturation in view, and no verdict on where they end up. The paper set out five reporting rules for any tool used with young athletes.
- Current status. Measured capacities today, with the test date and the conditions recorded.
- Uncertainty. Measurement error and maturity-estimate error shown beside each result rather than in a footnote.
- Maturation context. Estimated maturity status, with comparison within maturity bands as well as age groups.
- Change over time. Each athlete’s own trend across several seasons, which is the only comparison that respects a non-linear pathway.
- No verdict. No ranking, no potential score and no selection recommendation below mid-adolescence.
Bio-banding supplies the design lessons. Make maturation explicit, compare within maturity bands, and treat any assessment taken during a growth spurt as provisional. A model that cannot see age can still see size, which is why maturity has to be measured and reported rather than assumed away.
Who can actually change this?
Procurement, which means governing bodies rather than clubs. An individual club has no leverage over a vendor. A governing body setting conditions of pathway funding or accreditation has a great deal.
Before buying, require the vendor to state the training population, the label the model predicts, the evaluation design, and the handling of relative age, maturation and survivorship. A builder that cannot answer those questions has not yet built a tool fit for selection.
For a coach, the practical version is narrower. Treat any potential index as one input with no privileged status. Measure and plot height regularly so the growth spurt is visible. Judge late developers against biological rather than chronological age. Keep simple follow-up records on released athletes, so that the system can eventually measure its own misses, which at present no pathway can.
Reading the paper now, the development evidence has not moved against it, and bio-banding has continued to spread. What it did not foresee is how ordinary the potential index would become, as cheap video analysis and bundled AI features arrived inside mainstream academy platforms. Its four recommendations remain, as far as can be established, standard practice nowhere.
Get the full paper
Read Predicting the Unpredictable in full, with every claim carrying an evidence grade and the full reference list. The full paper is open to read from the research page.
Frequently asked questions
Can AI predict which children will become elite athletes?
There is no published, prospectively validated evidence that machine learning can forecast adult elite status from measurements taken before or around puberty. Tools can describe current performance accurately. Predicting the adult from the 12-year-old is a different claim, and it is not supported.
What is the relative age effect and why does it matter to a model?
Children born early in the selection year are over-represented in squads, an effect confirmed across 14 sports and 16 countries and most pronounced in adolescent boys at representative level. A model trained on selection history absorbs that pattern and reproduces it, with the appearance of objectivity.
Why is a growth spurt a problem for assessment?
Because it changes the thing being measured while it is being measured. Maturity offset can be estimated without invasive testing, and the estimate carries an error band of roughly half a year. Any score taken mid-spurt should be treated as provisional and rescored afterwards.
What should a pathway do about athletes it releases?
Keep a simple follow-up record. Without it, the pathway sees only the athletes it retained, which is precisely the sampling problem that makes its data unsuitable for training a predictive model in the first place.
Is there an age below which predictive scoring should not be used at all?
The paper argues for no ranking or potential score below mid-adolescence and leaves the exact threshold open, since it may reasonably differ by sport. What it does not leave open is the principle that a number which closes a door should not be produced from a body that has not finished changing.
Legal and regulatory notice. This article describes research findings and governance practice. Nothing here constitutes legal advice, and the provision of legal advice sits outside the terms of any engagement with the author or with Athleet.AI. The material is presented to support discussion and further review by qualified advisers.