Correction of AI systems by linear discriminants: Probabilistic foundations