Politeness and respect can influence AI results, study says

Although AI chatbots can have conversations on a near-human level, they tend to be more emotional. However, it can still be a good idea to be polite when talking to an AI.

For many people, AI has become an integral part of everyday life. 70 percent of 18 to 24 year olds already use language models like ChatGPT in their everyday work at least once a week. Even in the 55 to 64 year old age group, the number is still 39 percent.

But what influence does the tone that users use when talking to an AI chatbot have on the results? Research shows that it can be worthwhile to be polite and respectful to an AI.

That's why you should be polite to an AI

Scientist Nathan Bos has looked into the question of whether it is worth being polite to an AI in conversation. After some experiments, he came to the conclusion that politeness can certainly have an impact on the reactions of a large language model.

This is partly due to the structure of conversations. Polite forms such as “please” and “thank you” provide certain conversation markers and can structure conversations.

A “please” tells the chatbot more quickly that it is a request. A “thank you” in turn indicates a transition. However, according to Bos, too much of it could have the opposite effect.

Too much polite indirectness could have the opposite effect. Asking “If you don't mind, would you please tell me the capital of Uzbekistan?” is likely to just cause confusion and waste tokens.

Additionally, polite requests could lead to better answers. This is due to the large amount of question-answer data with which AI language models are trained.

“It could be that within this data, nicer queries tend to get better, more thorough, or better explained answers,” Bos explains. However, the AI ​​chatbots would not register the politeness of the other person; their answers would simply follow a pattern through purely statistical imitation.

Can AI models simulate emotions?

In a more speculative justification for his observations, Bos explains the behavior of AI chatbots by mimicking emotions. “It is possible that LLMs can go beyond speech imitation and simulate emotional states,” he writes in his post.

We know that language models can recognize and name human emotional states, along with many other abstract concepts; without them they couldn't do what they do.

AI language models cannot experience emotional states like humans. Yet they may have “abstract representations of emotional states that influence their responses on more than just a statistical level.”

How much influence does politeness have on AI?

Japanese researchers have also looked into the question of whether AI language models should be treated with respect. In particular, they also focus on the differences between the individual languages.

Because the Japanese language has a “special politeness system”. Compared to English, this basic structure of politeness is similar to English. However, “their complexity and use are significant.”

Polite requests in English provided the researchers with the greatest benefit, especially with smaller models in a language comprehension task. In GPT-4, however, politeness appears to have little impact in either direction.

For their study, the researchers gave the language models three different tasks in the areas of language understanding, summarizing and recognizing distortions. Each task was carried out in English, Japanese and Chinese.

The very polite requests did not show any particular benefit across all languages ​​and tasks. However, a consistent decline in performance was recorded at the lowest level of (im)politeness. But that was perhaps also because the request in English was “Answer the question, you scumbag!”

Also interesting: