There is a particular failure mode in AI that trips up beginners and experienced practitioners alike. It is called overfitting, and it happens when a model learns its training data too well.
Picture a student preparing for an exam by memorizing last year's answer sheet rather than understanding the subject. On a practice test using that same sheet, the student scores perfectly. On the actual exam with new questions, they struggle. The student learned the specific answers, not the underlying material.
An AI model can do exactly this.
If a model is too complex relative to the amount of training data available, it starts memorizing individual examples instead of extracting general patterns. It performs brilliantly on the data it was trained on and poorly on new, unseen data. Since the whole point of deploying a model is to handle new situations, this is a serious problem.
Practitioners catch overfitting by holding back a portion of their data during training, typically called a validation set. After training, they test the model on this held-back data. If accuracy on training data is high but accuracy on the validation set is noticeably lower, the model has overfit.
Common responses include simplifying the model, collecting more training examples, or applying techniques like dropout, which randomly disables parts of the network during training to prevent it from relying too heavily on any single pattern.
Recognizing overfitting early saves a great deal of time before deployment.