When we talk to Chatgpt, The famous artificial intelligence model often comes naturally to us thank him For the answers it gives us. Indeed, there are those who ask him how they are before asking a question. We do a little by habit, a little because some consider it almost “human”. There are also those who hopefully (jokingly) that, in a future dominated by the machines, the Ai remember who has been Kind. But it really serves something be educated with chatgpt? The answer is yes, but not because this will save us in a possible “apocalypse ai”.
A recent study has shown that be educatedbut not too much, with the AI, can lead to Best answers. There are two main reasons: the first is that the Ai tends a reflect the communicative style of the user, therefore if we use a clear and respectful tone, it is more likely that the response is also well structured and appropriate. The second is more intuitive: when we try to be courteous, we take more time for formulate the request wellwe are more precise, we give more context and therefore we get better results.
In this article we see how the way we talk to Chatgpt influences the answers and what are the Economic and energy costs of saying “please” to the AI.
If we insult Chatgpt, we get worse answers
A group of Japanese researchers from Waseda University has recently conducted a study on how kindness and education in writing the instructions to AI (called “prompt”) influence the answers. In particular, they analyzed 8 different levels of education in prompt: from the lowest, in which they insulted and threatened the AI, at the highest, in which formulas were used excessively respectful and redundant. The study tested the behavior of the AI on three different tasks: summarize an article, answer a question and analyze a sentence. Each task was carried out in three languages: English, Chinese And Japanese.
They discovered that it is Extremely rude prompt than those too kind the answers worsen. The former can generate content imprecisefull of prejudices or even bring the ai a block For security reasons. The latter, on the other hand, risk confusing the model, generating more cars, less centered and with a greater risk of hallucinations (i.e. incorrect information shown as real).
In most cases, a moderate level of education It is the best option. Of course, this standard of “moderation” is also relative: the concept of kindness changes from culture to culture. A phrase deemed courteous in English could be too direct, or even rude, for a Japanese. Exactly as in the conversations between human beings, even in dialogue with the AI the cultural context counts.
When we are educated we write the prompts better
But why be kind Should I improve the AI responses? If you think chatgpt is offended … well, no. Chatgpt is a model of language trained on huge quantities of text: does not feel emotions, Nor does it worry about the tone we use. It doesn’t change anything if we write “can you summarize this email?” Or “I summarize this email now”. The point is another: the Ai tends a Copy the user’s tone. So if we use too many swear words or insults in the prompt, the model will choose similar words to write the answer. On the other hand, excessive courtesy risks only “confusing” it and making him more difficult to interpret the request. For example, phrases like “if you are not too disturbed, you could kindly summarize this email?” They are not very effective to get good answers.
Beyond the results of this study, an interpretation shared by many experts is that When we try to be educatedOften We organize our thoughts better. Let’s think about what we mean, add context, form the request in a more precise way. And all this helps the Ai to answer better. For example, saying simply “explain me climate change” is very generic. If, however, we stop for a moment and try to write the prompt in an educated way we could get: “You could explain to me what climate change is clearly, please? I am interested in understanding what the main causes are and what consequences it has”. Adding even this little even contextwe increase the chances of obtaining a useful and accurate response. The most precise answer will not depend so much on “please”, but on having reflected a moment more on how to structure the prompt.
Being educated with chatgpt costs millions of dollars
Being educated with chatgpt improves the responses generated, but also has a huge impact energy and economic.
A user of X (ex Twitter) recently asked how much money he spent Openii (the company that develops chatgpt) to “cause of the people who say ‘please’ and ‘thanks to their models”. The answer came directly from Sam Altman, the CEO of Openai: according to him, these courtesy formulas cost the company tens of millions of dollars. Altman, however, commented jokingly: “Millions of dollars well spent. You never know”. It looks really too much, but Chatgpt is a huge model and every request, even the simplest, requires calculation power. The longer or complicated the sentence, the more energy is needed to process it.
