In plain English
Overfitting is when a model performs well on its training data but poorly on new, unseen data, because it has memorised specific examples rather than learning the general patterns. It is the classic failure mode of machine learning. More and more varied data, and careful evaluation on held-out examples, are the usual defences.
Why it matters
An overfit model looks great in testing on familiar data and then disappoints in the real world, which is why a separate, honest evaluation set matters so much.
A worked example
A student who memorises last year's exam answers scores perfectly on a re-run but fails a fresh paper. An overfit model behaves the same way.
Common confusion
High accuracy is not automatically good. If it was measured on the same data the model trained on, it can be hiding overfitting.