A metaphor for finding patterns that aren't there. People see faces in toast; flexible algorithms find patterns in noise. Only fresh data exposes them.

ExampleA model trained on coin-flip labels scores 100% on what it studied and 49% on new data.