Should You Be Polite to A.I.?

Written By Jack Spiegel

In April of this year, a post was made on X saying, “I wonder how much money OpenAI has lost in electricity costs from people saying ‘please’ and ‘thank you’ to their models” (@tomieinlove, 2025). Billionaire CEO of OpenAI Sam Altman responded, “tens of millions of dollars well spent--you never know” (Altman, 2025). OpenAI developed ChatGPT, DALL-E and SORA (OpenAI, 2025). So what does Altman mean? Are AI companies really losing millions of dollars because of common courtesy? And what about the “you never know”? Does being courteous really affect A.I. (Artificial Intelligence) models? 

A.I. companies can end up losing substantial amounts of money due to polite prompts. A recent study found that generating a 100-word email requires 0.14 kilowatt-hours of electricity, equivalent to having an LED bulb on for 14 hours (Verma, 2024). The politeness of a prompt significantly affects the amount of electricity used when generating a response. Another academic study found that only 42% of the content of all A.I.-generated responses provides the key information necessary to answer questions. While some of the extra detail was useful, a large portion was not. (Poddar et al., 2025). “Trimming the fat from AI outputs can yield huge energy savings. By prompting the model to give only a minimal answer (no fluff), researchers cut the response length by up to 88%, which translated to about a 58% reduction in energy consumption for that query,” Murat Köse said in a blogged exploration of the Poddar paper (Köse 2025).

The significant presence of unnecessary content in A.I. answers makes a big difference, considering the large number of prompts LLMs (Large language models) are answering daily. For example, ChatGPT uses roughly 0.5 gigawatt-hours of energy per day to answer more than a billion prompts. At a cost of upwards of $75,000, that much energy could power 40 thousands homes per day. Because of bloated A.I. responses, OpenAI loses $11 million annually (Köse, 2025). Because A.I. has learned polite inquiries are usually returned, it adds unnecessary bulk to its answers in the form of positive language. Every word of positivity it adds is electricity used by OpenAI and money wasted (Yin et al., 2024). 

Not all evidence suggests that instructing A.I.s to be more succinct is ideal, as Altman’s second point remains: “you never know”. Being polite to A.I.s can alter behavior in more ways than just adding insubstantial content (Yin et al., 2024). LLMs have learned from human patterns that when someone is polite, they should respond equally politely, lengthening their answers. However, they have also learned that when asked polite questions, humans tend to provide better answers. A.I.s have mimicked this human behaviour, and give better answers when the prompts given to them are more polite (Yin et al., 2024). 

A recent study evaluated the performance of three A.I. models using three benchmarks for three different languages. While contrasting combinations had varying degrees of effect on models, they found that “Using impolite prompts can result in the low performance of LLMs, which may lead to increased bias, incorrect answers, or refusal of answers” (Yin et al., 2024).

While testing English-speaking models, they found that the most effective form of questions was overly polite, formatted as: “Could you please (Task Description)? Please feel free to (Answer Format). You don’t need to (Answer Restriction)”. In the data the A.I. had seen from previous conversations, it observed that people gave better answers when respected and asked nicely. It then followed this trend, which was made clear through the variation based on politeness (Yin et al., 2024).

The study to determine whether politeness affects LLMs' answers had three types of tests: A 5,700-question English test, a 5,200-question Chinese test, and a 5,591-question Japanese test. Of the eight levels of prompts used, ranging in politeness, level 8 was the highest. ChatGPT 4 scored its best at a level 4 in every language. LLama 2 reached its peak at English level 7, and Chinese and Japanese level 6. ChatGPT 3.5 demonstrated the widest difference in best-scoring politeness levels between languages, scoring its best English score at level 8, its best Chinese at 6, and its best Japanese at only 2. The researchers found that often prompt variations had a minimal effect on the answer, concluding that, “Highly respectful prompts do not always lead to better results. In most conditions, moderate politeness is better, but the standard of moderation varies by languages and LLMs” (Yin et al., 2024).

In conclusion, considering both energy cost and effectiveness leads to the necessity of compromise: be direct but not rude. A.I. is a complex tool that requires consideration of energy and efficiency when in use. As a tool, it is meant to be used, but finding the best way to do so is no easy task, considering the moral evaluation of the trade-offs LLMs require to be used. Being polite uses both money and energy, negatively impacting the environment and driving up electricity prices, yet it will likely improve the quality of response. Finding what you are most comfortable with when using A.I. is up to you, but remember that your word choice is important. It seems that, despite the turn of phrase, it costs to be kind.

References

Altman, S. [@sama]. (2025, April 16). tens of millions of dollars well spent--you never know [Post]. X. https://x.com/sama/status/1912646035979239430

Köse, M. (2025, June 12). The Cost of Courtesy: Does Polite Language in AI Prompts Waste Energy? Medium. Retrieved November 13, 2025, from https://medium.com/@ul.tu/the-cost-of-courtesy-does-polite-language-in-ai-prompts-waste-energy-949d735b8f05

OpenAI [Computer software]. (2025). https://openai.com/

Poddar, S., Koley, P., Misra, J., Ganguly, N., & Ghosh, S. (2025). Brevity is the soul of sustainability: Characterizing LLM response lengths. 21848–21864. https://doi.org/10.18653/v1/2025.findings-acl.1125

@tomieinlove. (2025, April 15). I wonder how much money OpenAI has lost in electricity costs from people saying "please" and "thank you" to their models. [Post]. X. https://x.com/tomieinlove/status/1912287012058722659

Verma, Tan. (2024, September 18). A bottle of water per email: the hidden environmental costs of using AI chatbots. The Washington Post. https://www.washingtonpost.com/technology/2024/09/18/energy-ai-use-electricity-water-data-centers/

Yin, Z., Wang, H., Horio, K., Kawahara, D., & Sekine, S. (2024, October 14). Should We Respect LLMs? A Cross-Lingual Study on the Influence of Prompt Politeness on LLM Performance [Working paper submitted for publication]. ArXiv. https://doi.org/10.48550/arXiv.2402.14531

Jack Spiegel
Next

The Sustainability of the AI Revolution