Artificial Intelligence: ChatGPT gives better answers when you are friendly

The next time you ask artificial intelligence a question, you should phrase it nicely: ChatGPT and Co. like it when you say “please” and “thank you”.

According to a preprint study by Waseda University in Tokyo, so-called large language models (LLM) – the technology that all AI-Chatbots drive higher quality responses when treated politely

Schleimer doesn’t like ChatGPT

But don't overdo it! Excessive flattery can actually worsen performance. The study authors recommend a middle ground, i.e. using requests that are similar to the norm in most human social interactions. Because: “LLMs reflect the human desire to be respected to a certain extent,” the Japanese scientists are sure.

Nathan Bos, a research associate at Johns Hopkins University in Baltimore (USA), who studies relationship patterns between humans and AI, is also convinced of this. He told the magazine “Scientific American” he says: “Polite prompts could cause the system to seek information from more polite and therefore more credible corners of the Internet.”

A rude question could therefore have the opposite effect and direct the system to questionable forums.

This is how we can become nicer too

Support, such as that shown by a patient teacher to a student with learning difficulties, can also influence a chatbot's performance. For example, a team of researchers at Google DeepMind found in a study that prompts such as “Take a deep breath and work through the problem step by step” improved an LLM's ability to solve math problems.

Polite interaction with AI can also train our general way of communicating: If we learn to be nice to technology, we will, in the best case, also apply this to our interactions with other people.