The tendency of users to change their prompting strategy and task difficulty based on observing a model's responses.