From being overly polite to pretending to be on “Star Trek,” the tips for talking to chatbots often seem strange — and they don’t always work. A group of researchers tested whether “positive thinking” makes artificial intelligence more accurate. They praised the AI as “smart,” encouraged it to think carefully, and even ended the question with “This will be fun!” None of it yielded consistent results. But one technique stood out: when they asked the AI to pretend to be on “Star Trek,” it improved in basic math.
People are using some of the strangest strategies to get better answers from large language models (LLMs), the technology behind tools like ChatGPT. Some say you have to threaten them, others say you have to be overly polite, and still others ask the AI to play the role of an expert. These practices are known as “prompt engineering.” But experts say many of the popular “rules” don’t work — and can sometimes be dangerous.
“People think there’s a magic formula for words,” says Jules White, a professor of computer science at Vanderbilt University. “But it’s not about the choice of words, it’s about how you express what you’re looking for.”
Should we be polite?
Each word you type is broken down into “tokens” and statistically analyzed by the model. Even an extra comma can affect the response — but it’s very difficult to predict how. A 2024 study suggested that chatbots give more accurate answers when you address them politely. Another test showed that an earlier version of ChatGPT was more accurate when you insulted it. The results are inconsistent, and the models are constantly being updated, making any conclusions tentative.
Experts agree that today's techniques like flattery, threats, or insults are a waste of time if you're aiming for accuracy. New models are much better able to understand the essence of the request.
How to get better answers
– Ask for several options. Don’t ask for one answer, but three or five variants.
– Provide examples. If you want an email with your own style, send some previous examples.
– Let the AI interview you. Ask it to ask questions until it has enough information.
– Be careful with the role of the expert. For creative tasks it works, but for questions with a correct answer it can increase the risk of errors.
– Stay neutral. Don’t direct the answer with your preferences.
Does politeness have value?
Politeness doesn't make the model more accurate, but it can make you more comfortable using it. According to some surveys, most users say "please" and "thank you" when talking to AI. Not because of fear of a robot rebellion, but because it's human habit.
The bottom line? Artificial intelligence has no feelings – it's a tool that mimics human behavior. If you want better results, treat it as a tool, not a person.
