Recent research has reported a counterintuitive finding. Polite prompts such as “could you please” or “thank you” can reduce large language model accuracy on tightly scored tasks with one correct answer, while blunt directives sometimes score higher. This article explains what those results do and do not imply, why they …
