Daily interactions with virtual assistants and language models like ChatGPT have sparked a debate about whether using polite formulas like ChatGPT really has any effect. “thank you” o “please” when communicating with artificial intelligence.
While these gestures are technically computationally expensive, research shows that a moderately polite tone can improve the quality of responses offered by chatbots. The of science is starting to show that being friends with AI isn’t as useless as many thought.
A study from Waseda University in Tokyo, published in 2024, analyzed how the level of politeness in messages—whether suggests— affects the quality of responses generated by language models such as ChatGPT or LLaMA.
The researchers tested instructions with varying degrees of politeness, from rude to overly formal messages, in English, Chinese, and Japanese.
The results showed a clear trend: Very dry or disrespectful prompts elicited less accurate and useful responses. On the other hand, a polite but not overbearing tone generally resulted in more reliable, complete, and better structured responses. That is, the best results of the conversation are achieved in a balance between sharpness and extreme formality.
Far from interpreting education as a a personlanguage models operate on the basis of linguistic patterns present in their training data. AI does not “feel” politeness or take offense at rudeness, but it does detect and reproduce conversational styles.
A clear, well-contextualized message with a collaborative tone can direct the system to offer more detailed and relevant responses, while ambiguous or authoritarian instructions tend to generate shorter and less detailed responses.
This interpretation was supported by Murray Shanahan, a Google DeepMind researcher who compared these systems to an intelligent intern: They tend to adapt to the style suggested by the user, responding with more cooperation and patience when they experience more polite treatment. cooperation.
The use of politeness formulas in communication with AI is not without debate. Sam AltmanCEO of OpenAI, warned that millions of goodbye and thank-you messages create a negligible resource and energy drain to keep the servers running.
However, research and surveys show that many users continue to use these expressions not because they expect an emotional response from AI, but because they are part of their daily communication habits.
Science suggests that though I.A does not recognize politeness in human terms, maintaining a polite, cooperative tone can be helpful in eliciting better responses.
A study conducted by the MIT Center for Constructive Communication (CCC) warns of biases in the most complex language models, a factor that negatively affects the quality of responses offered by users with low English proficiency, low educational attainment, or those who live outside the United States.
This situation, according to the consulted report, could exacerbate inequalities in global access to information.
The rise of tools like OpenAI’s GPT-4, Anthropic’s Claude 3 Opus, and Meta’s Llama 3 have raised expectations for greater equity in access to knowledge. Promoted as solutions for the democratization of information, these platforms promise universal and intuitive interfaces, regardless of origin or educational level of the user.
However, research presented by MIT’s CCC at the AAAI Artificial Intelligence Conference shows that AI systems tend to provide less accurate and reliable answers to those who could benefit the most from them.
Lead author Elinor Poole-Dayan emphasizes, “This vision cannot be realized without ensuring that model biases and harmful trends are safely mitigated for all users, regardless of language, nationality, or other demographics.”
