Feb 26, 2024 - Technology

Being nice to chatbots pays off

Illustration of a human hand and a computer cursor hand engaged in a handshake

Illustration: Annelise Capossela/Axios

It pays to be nice to your AI: Large language models (LLMs) tend to give better answers when prompted respectfully and failure to do that can "significantly affect LLM performance," per a new cross-cultural research paper.

Why it matters: Your prompt may affect the answer you get today — and also how well an AI model can answer for everyone else tomorrow.

  • "Impolite prompts may lead to a deterioration in model performance, including generations containing mistakes, stronger biases, and omission of information," the researchers found.

What they did: The researchers tested a half-dozen chatbots against dozens of tasks, using up to 150 prompts per task.

What they found: LLMs mirror certain human communication traits, which means politeness toward chatbots tends to generate better responses, just as politeness does in human conversation.

  • The thesis proved true across English, Chinese and Japanese prompts and with each chatbot tested.
  • "Impolite prompts often result in poor performance, but excessive flattery is not necessarily welcome," the researchers found.
  • One effect of excessively rude or flattering prompts was that they generated longer answers in English and Chinese.

Reality check: Chatbots are not sentient, and your politeness isn't making them feel good.

Go deeper