Google warns that including too many examples in a prompt can cause a model to overfit to them. What does that look like in practice?

The principle — A model generalises from what your examples share. The more of them there are, the stronger every shared feature becomes — including the ones you never meant to demonstrate.

Why the key is correct — Google notes that models can often pick up on patterns using a few examples, but adds that if you include too many, the model may start to overfit the response to the examples. In practice this shows up as new answers that carry incidental properties of the sample set: the same sentence count, the same hedging phrase, the same category chosen when the input is ambiguous. Anthropic describes the same risk from the other direction, advising that examples be varied enough that the model does not pick up unintended patterns.

Why the others are wrong — Models attempt unfamiliar inputs rather than refusing them. Exhausting the context window is a different and far more visible failure. And examples demonstrate a task rather than supplying facts to cite.

Remember this — If every answer looks like your examples, the examples have stopped being a guide and become a mould.

Sources — Google, Prompt design strategies.