What Is Overfitting?

Overfitting is when a model learns its training data too closely, memorizing noise and specific examples instead of the general patterns, so it performs well on that data but poorly on new, unseen data. An overfit model has essentially memorized the answers rather than learning to generalize.

How overfitting works

Overfitting happens when a model has enough capacity to memorize its training examples, including their quirks and noise, rather than learning the underlying pattern. It fits the training data almost perfectly, but the details it memorized do not carry over to new data, so its real-world accuracy drops.

The telltale sign is a large gap between performance on the training data and on data the model has not seen. Teams catch this by holding back a separate test set and checking the model against it. If it scores far worse there than on training data, it has overfit. The opposite problem, underfitting, is when a model is too simple to capture the pattern at all and does poorly everywhere.

Why overfitting matters for AI

Overfitting matters because a model is only useful if it works on new data, not the examples it was trained on. A model that looks excellent in training but overfit will disappoint in production, which is why guarding against it is a routine part of building any model. Common defenses include training on more and more varied data, keeping the model no more complex than it needs to be, and using techniques that discourage it from leaning too hard on any single feature. Checking against a held-out test set is how teams confirm a model generalizes rather than memorizes.