Overfitting: the difference between learning the rule and memorising

Two colleagues are preparing for a certification.

• Anna understands how the system works.
• Bo has memorised the exact answers from the manual: "Customer type A requires answer X", "Customer type B requires answer Y".

Both pass the exam on the examples included in the manual. But in practice, a customer type C appears — one that was not in the material. Anna can handle it. Bo is stuck.

Bo has overfitted: he knows the examples, not the rule. AI models do exactly the same thing.

Why we test on new examples. If the model only answers questions it has been trained on, you are measuring its memory, not its competence.

	Training examples	New examples	What it means
Good model	92 %	89 %	has learned the pattern
Overfitted	100 %	61 %	has memorised
Underfitted	58 %	56 %	has learned nothing at all

The difference between the columns is the warning light.

What helps? More and more varied examples, or a simpler model that does not have room to memorise. It is roughly like Bo needing to see more different customer cases — or someone removing his cheat sheet.