When the prompts used to evaluate a model are influenced by the model's own responses, violating the assumption that evaluation data is independent.