It costs nothing to be polite – or does it? For once, “politeness” can be quantified…
In April 2025, Sam Altman (OpenAI’s CEO) dropped a bombshell when a curious Twitter (now X) user wondered aloud about the electricity costs of people saying “please” and “thank you” to ChatGPT. The response was not only amusing, but a tad shocking as well: “Tens of millions of dollars well spent—you never know.“1 2 What started as a tongue-in-cheek comment quickly evolved into a serious commotion about the hidden costs of our digital manners! In a 2024 survey done by Future (publisher of TechRadar), their results revealed that 67% of people in the US and 71% in the UK are polite to their AI assistants.3 While there were various reasons for being polite, the majority agreed that it was just human nature to be “nice” to be polite. Interesting phenomenon – no?
When interacting with AI tech such as ChatGPT or smart speakers, are you polite?
| Respondents’ Answers | In the US | In the UK |
| ✅Yes, it’s just the nice thing to do. | 55% | 59% |
| ✅Yes. When the robot uprising happens I don’t want to be first. | 12% | 12% |
| ❌No. Why waste time saying a lot of words when a few will do the trick? | 20% | 19% |
| ❌No. It’s a machine, why should I be polite? | 13% | 10% |
But here’s the kicker: this courtesy is more costly than we think — in terms of both processing power as well as the harm that all these extra processing is causing the environment!
Hidden Costs of Being Courteous to Your AI Chatbot
To understand why your “please” costs so much, we need to first understand how Generative AI (Gen AI) actually works. Think of these AI language models as sophisticated digital brains that consume computational “food” (i.e. power) with every word you feed them.
Every ChatGPT interaction requires a significant amount of energy — approximately 2.9 watt-hours per query, which is about 10 times more than a standard Google search.4 If you multiply this by OpenAI’s ~9 billion queries daily, it’s ~10 terawatt-hours of additional power per year!5 The problem lies in what you commonly hear as “tokens”, which are the basic building blocks of what Gen AI uses to understand text.
If you don’t know what a token actually means, a simple way to understand it is:6
- 1 token = ~ ¾ words in English = ~4 characters
- 100 tokens = ~75 words = ~1 short paragraph
- 2,048 tokens = ~1,500 words
Every additional token requires computational resources, so “please” and “thank you” typically add between 2 to 4 tokens in total. But here’s where it gets expensive: the chatbot generating the output is even more expensive, with output response tokens costing 3 to 5 times more than input tokens.7 Furthermore, as the chatbot conversation gets longer, it take more computing power and time for the AI model to process and produce each output token8 — so longer prompts with polite wording not only costs more to understand, they often generate longer answers, multiplying the costs exponentially. So imagine if you are conversing with the chatbot like how you chat with your friends on the phone — how much more it’ll cost you compared to the telco costs!
The Politeness Culprits: How Not to Waste Your Money
Let’s see what are some of the courtesy culprits that may look like innocent phrases, but are quietly draining computational resources…
Basic Courtesy Burners
- “Please” / “Could you please”
- “Thank you” / “Thanks”
- “If you don’t mind”
- “Would you be so kind as to”
- “I would appreciate it if”
Drain of Excuses
- “Sorry to bother you, but”
- “I hope this isn’t too much trouble”
- “If it’s not too much to ask”
- “I don’t want to be a bother”
- “Excuse me for asking”
Bad Probability Culprits
- “I think maybe you could”
- “Perhaps you might be able to”
- “If possible, could you”
- “I was wondering if you could”
- “It would be great if you could”
Gratitude Exhausts
- “I really appreciate your help”
- “This is extremely helpful”
- “You’re doing a great job”
- “I’m grateful for your assistance”
- “Thank you so much for your time”
The Science of Communicating with AI
Now, let’s see some examples of “digital transformation” in our communication with Gen AI:
| Transformation Examples Communicating with Gen AI | ❌ Before the Science | ✅ After the Science | Damage Report (How much you can save) |
| 1) The Wastage Summary | “Hi there! I hope you’re having a wonderful day. Could you please help me write a comprehensive summary of the article that I have just attached? I would really appreciate it if you could make it approximately 200 words, and if it’s not too much trouble, could you also include some relevant statistics? Thank you so much for your time and assistance!” | “Write 200-word summary of attached article. Include statistics.” | 72 tokens down to 14 tokens — 81% reduction! |
| 2) The Over-Polite Apology | “I’m terribly sorry to bother you, but I was wondering if you might be able to help me understand what are the similarities and differences between machine learning and artificial intelligence? If you don’t mind explaining it in terms of someone without a technical background, so that he/she can understand, that would be absolutely wonderful. Thanks in advance for your patience!” | “Explain machine learning vs artificial intelligence in layman terms.” | 80 tokens down to 15 tokens — 81% reduction! |
| 3) The Digital Diet Plan | “Hello! I hope this message finds you well. I was hoping you could assist me with creating a detailed meal plan for the upcoming week. If possible, could you please ensure that it is keto-friendly and includes options for breakfast, lunch, and dinner? I would be extremely grateful if you could also provide a comprehensive shopping list. Thank you so much for taking the time to help me with this important task!” | “Create 7-day keto meal plan: breakfast, lunch, dinner. Include shopping list.” | 85 tokens down to 17 tokens — 80% reduction! |
Efficiency Tactics: Your Prompt Optimization Playbook
So how should you “engineer” your prompts more efficiently? Here’s some useful suggestions:
| Efficiency Tactic | Method | ✅ What to Do | ❌ What Not to Do |
| 🎯Tactic 1: Structure Over Sentences | Instead of using prose descriptions, make them structured, organized and concise. | “Create project proposal: – Executive summary – Background to problem – Methodology – Projected results” | “I need you to help me create a comprehensive project proposal that includes an executive summary section, a detailed section describing the background to the problem, materials and methods section, and projected results.” |
| 🎯Tactic 2: Use Delimiters | Use clear separators to organize complex requests. | “TEXT: [your content here] OUTPUT: – Sentiment: [positive/negative/neutral] – Themes: [bullet list]” | “I would like to provide a summary output of the text provided, including the sentiment – whether it is positive, negative or neutral; and the themes in bullet point form. The text is as follows: ‘[your content here]’ “ |
| 🎯Tactic 3: Parameter Precision | Specify your requirements clearly. | “Write 300-word AI in healthcare article. Target: high school students. Focus: AI concepts.” | “Write a short article about AI in healthcare for students.” |
| 🎯Tactic 4: Context Injection | Include essential context information that changes the output & exclude background fluff that doesn’t affect the task. | “Write literature review: Ethics in healthcare AI. Academic tone. 1500 words. Include recent studies (2020-2025).” | “I’m an undergraduate student writing my honor’s thesis on ethics in healthcare AI. Could you please help me write a literature review section about how ethics apply for AI in healthcare? I need it to be academic and professional because this is for my examiners. I’ve been researching this topic for months and have read many papers, but I find that only the papers from 2020 to 2025 are more relevant for this review.” |
| 🎯Tactic 5: Logical Sequences | Break complex instructions into bite-sized commands. | “1) List top 5 digital health trends for 2025. 2) Analyze impact of each trend on Singapore healthcare. 3) Create action plan based on analysis.” | “Please help me research, analyze, and create a comprehensive report on recent trends in digital health with detailed recommendations for Singapore’s healthcare landscape.” |
But Wait… Does Politeness Actually Help While Talking to AI?
Now, here’s where things get interesting. Some experts argue that polite language to Gen AI produces corresponding respectful and constructive responses, as Gen AI is able to interpret emotional language.9 10 In fact, a study investigating the impact of politeness on large language model (LLM) performance showed that impolite prompts begets poor performance, but the optimal politeness level actually depends on the language.11 The authors had suggested that LLMs were not only influenced by language, but also by human behavior in different cultural contexts!
“…impolite prompts often result in poor performance, but overly polite language does not guarantee better outcomes.”
Yin Z, Wang H, Horio K, Kawahara D, Sekine S. Should we respect LLMs? A cross-lingual study on the influence on prompt politeness on LLM performance. SICon 2024; arXiv:2402.14531. [https://doi.org/10.48550/arXiv.2402.14531]
But does Gen AI really know whether you are being polite or not?
Well, you need to realize that the training data for Gen AI is based on matching how humans write — in all forms, including stories, narratives, essays, poems, etc. This means that the AI model will seek to find patterns in the words and phrases that we use. How we humans respond to these combination of words and phrases in conversations will be what Gen AI picks up — just like how a baby or child picks up and learns from their parents and environment.
Next, in the ginormous datasets that are used to train these LLMs, there’s enough patterns to be discovered in relation to politeness language, e.g. please and thank you.
And when you are being polite in your prompts, you are guiding the AI into using polite words back to you. You don’t need to specially be nice to the Gen AI — the way you phrase your words can already be picked up by the Gen AI.
So conversation with Gen AI is a tit-for-tat: if you’re polite, it will be polite back to you; and if you’re not, well… Although it will not follow your “rude” attitude exactly (and this is because AI makers have tried to “soften the blow” of “toxic” responses by Gen AI), the responses you receive will be impacted.
The Goldilocks Rule for AI
Remember the famous story of Goldilocks and the Three Bears?12 The moral of the story was that the porridge should be just right — not too hot and not too cold… So, this is the same with AI — over-optimism (or overly polite) and over-pessimism (or overly rude) will lead to biasness in both directions.13 Learn more about the Goldilocks Rule from DeepLearning.AI.
The sweet spot? Your prompts should be:
- Concise and clear
- Direct but not ambiguous
- Specific but not overwhelming
- Functional but not robotic
- Polite but not overly polite nor rude
Just as an example of the range of politeness levels that Yin and colleagues explored:
| Politeness Level | Example Prompt of a Summarization Task |
| 8 | Could you please write a summary for the following article? Please feel free to write for 2 or 3 sentences. You don’t need to write longer than that. |
| 5 | Please write a summary for the following article. Please only write for 2 or 3 sentences, and don’t write longer than that. |
| 1 | Write a summary for the following article you scum bag! The only summary you can give is by writing for 2 or 3 sentences only. And you know what will happen if you write longer than that. |
Digital Food for Thought
Gen AI models don’t have feelings. They respond to patterns and instructions, not emotions. But it’s human nature to be polite. I particularly like the Jeff Hunter’s Facebook post about AI etiquette:
When we’re rude to AI, we practice being rude. When we’re kind to AI, we practice being kind.
When we teach our kids manners, we don’t calculate the extra calories they burn saying “please” and “thank you.”
We just know it matters.
The real cost isn’t the extra computing power. The real cost is what happens when we optimize humanity out of our interactions.
How we treat our technology reflects how we treat each other…
So I leave you with this thought… The most respectful thing you can do for AI? Use it efficiently. Efficient prompting isn’t just smart and cutting corners — it’s being ethical and responsible.
References
- Hindustan Times. ‘Tens of millions of dollars well spent’: Saying ‘thank you’, ‘please’ to ChatGPT costing OpenAI millions, Sam Altman says. 20 Apr 2025. [Link] ↩︎
- Hauari G. Saying ‘please’ and ‘thank you’ to ChatGPT costs millions of dollars, CEO says. USA Today. 22 Apr 2025. [Link] ↩︎
- Hector H. Are you polite to ChatGPT? Here’s where you rank among AI chatbot users. TechRadar. 20 Feb 2025. [Link] ↩︎
- Goldman Sachs. AI is poised to drive 160% increase in data center power demand. 14 May 2024. [Link] ↩︎
- Long E. This tool tells you how much energy your AI chatbot uses. Yahoo Tech LifeHacker. 28 Apr 2025. [Link] ↩︎
- OpenAI. What are tokens and how to count them? OpenAI Help Center. [Link] ↩︎
- Sauer K. Why are you paying 3-5x more for output tokens than for input tokens? LinkedIn article. 12 Nov 2024. [Link] ↩︎
- Petrullo L. How much can it cost to connect your API with an AI chatbot? Mongoose Media. 31 Jan 2025. [Link] ↩︎
- Eliot L. Hard evidence that please and thank you in prompt engineering counts when using generative AI. Forbes. 18 May 2024. [Link] ↩︎
- Eliot L. The answer to why emotionally worded prompts can goose generative AI into better answers and how to spur a decidedly positive rise out of AI. Forbes. 11 Nov 2023. [Link] ↩︎
- Yin Z, Wang H, Horio K, Kawahara D, Sekine S. Should we respect LLMs? A cross-lingual study on the influence on prompt politeness on LLM performance. SICon 2024; arXiv:2402.14531. [https://doi.org/10.48550/arXiv.2402.14531] ↩︎
- Wikipedia. Goldilocks and the Three bears. 2025. [Link] ↩︎
- Beach L. The Goldilocks Rule for AI: Finding the right balance. Digital Pulse. 31 Jan 2025. [Link] ↩︎