Overfitting
Also called: memorization
When a model learns noise and idiosyncrasies specific to its training data rather than the underlying pattern, performing well in training and poorly elsewhere.
Overfitting is the classic failure of machine learning. A sufficiently flexible model can fit its training data nearly perfectly, including its random noise, and that fit does not transfer.
Medical datasets are especially vulnerable because they are small relative to model capacity and contain strong site-specific signals a model can exploit.
Where This Gets Misread
Models have been shown to learn shortcuts nobody intended: identifying the hospital from image artifacts and using its base rate, or keying on a laterality marker rather than the pathology. These score well on internal validation and collapse externally, and the model gives no indication it learned the wrong thing.