Overfitting
Definition
When a model learns the noise in its training data well enough to perform worse on anything new.
Diagnosed by the gap between training and held-out performance, not by training performance alone. A model that scores perfectly on its training set and poorly on a test set has memorised rather than learned.