When a text-to-image model generates an object mentioned in a prompt even when the prompt's linguistic context (negation, imagination, etc.) indicates it shouldn't appear.