Overfitting is one of the central problems in machine learning, and it has a deeply human quality. A model that overfits has memorised the answers rather than learned the pattern. It has confused the specific examples it was shown for the general rule those examples were meant to illustrate.
The technical picture: during training, a model is shown a dataset and asked to minimise its errors on that data. If the model is powerful enough and trains long enough without constraint, it can reduce its training error to near zero — not by learning the underlying structure, but by memorising every example as a special case. A student who memorises every past exam question rather than understanding the subject will ace the practice tests and fail the actual exam. The model is doing exactly this.
The practical consequence is a model that generalises poorly. Generalisation — the ability to perform well on new, unseen data — is the entire point of training a model. A model that works perfectly on its training set and falls apart on everything else is not useful. Overfitting is the pathology that makes this happen.
There is a version of this in human terms. The person who has optimised relentlessly for one particular performance — one company's culture, one industry's norms, one definition of success — often discovers that their model doesn't transfer. Life, like the test set, contains things the training data didn't include.