Maturation, Bias and the Training Set

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.