Overfitting

Overfitting is when an AI has studied its training data so intensely that it can't handle new, unfamiliar problems. Think of a student who memorized the answers to past exam questions word for word — the moment the test is phrased even slightly differently, they get it wrong, because they memorized the answers instead of understanding the underlying concept.

This happens because the model ends up learning "coincidental features" in the data instead of the "true underlying patterns." If all 100 cat photos in a training set happen to have grass backgrounds, the model might mistakenly learn "grass means cat."

Preventing overfitting is a basic, never-ending challenge in AI development. Developers gather more diverse data, periodically test the model on problems it has never seen during training, and impose limits (regularization) so the model can't memorize things too precisely. When you see "benchmark contamination" mentioned in the news, it's a close cousin of overfitting — the suspicion that a model was trained on the very test questions used to evaluate it, inflating its scores.

"The model scored high on benchmarks but underperformed in real-world use — raising suspicions of overfitting."