Are “please” and “thank you” just good manners, or are they changing how ChatGPT learns, behaves, and costs OpenAI’s artificial intelligence millions each day?
Saying “please” might be costing millions
It’s something most of us were taught as kids. Say “please.” Say “thank you.” Politeness costs nothing. But with artificial intelligence, that old wisdom may no longer hold true. Being polite to a chatbot might actually come with a price.
In a short exchange on X, OpenAI CEO Sam Altman revealed a curious detail about how AI systems work. When asked how much it costs OpenAI when users include extra words like “please” and “thank you” in their queries to ChatGPT, Altman replied, “Tens of millions of dollars well spent. You never know.”
Each word we type into ChatGPT is processed through vast data centers, where it gets broken into tokens, run through complex computations, and turned into a response. Even small pleasantries are treated the same way. They require computing power.
That means electricity, cooling systems, and more time spent per request. When multiplied across millions of conversations, those few extra tokens stack up into real energy and infrastructure costs.
According to a December 2024 survey by Future, the parent company of TechRadar, 51% of AI users in the U.S. and 45% in the U.K. regularly use AI assistants or chatbots.
Among them, Americans were more likely to be polite. In the U.S., 67% of users said they speak to AI with courtesy. Of those, 82% said it’s because it feels like the right thing to do, regardless of whether the recipient is human or not.
The other 18% have a different motivation. They said they stay polite just in case there’s ever an AI uprising — a long shot, but one they don’t want to risk being on the wrong side of.
Then there’s the remaining 33% of American users who don’t bother with niceties. For them, the goal is to get answers, fast. They either find politeness unnecessary or believe it slows them down. Efficiency, not etiquette, shapes the way they interact.
AI queries and the hidden infrastructure load
Each response from ChatGPT is powered by computational systems that consume both electricity and water. What seems like a simple back-and-forth hides a resource-heavy operation, especially as the number of users keeps rising.
A report by Goldman Sachs estimates that each ChatGPT-4 query uses about 2.9 watt-hours of electricity, nearly ten times more than a single Google search.
Newer models such as GPT-4o have improved efficiency, cutting that figure down to roughly 0.3 watt-hours per query, according to Epoch AI. Still, when billions of queries are made daily, even small differences quickly add up.
OpenAI’s operating costs reflect this scale. The company reportedly spends around $700,000 per day to keep ChatGPT running, based on internal estimates cited across multiple industry sources.
A major reason behind this cost is its massive user base. Between December 2024 and early 2025, weekly users jumped from 300 million to over 400 million, driven in part by viral features like Ghibli-style art prompts. As usage surges, so does the demand on electricity grids and physical infrastructure.
The International Energy Agency projects that data centers will drive over 20% of electricity demand growth in advanced economies by 2030, with AI identified as the primary driver of this surge.
Water is another part of the equation, often overlooked. A study by The Washington Post found that composing a 100-word AI-generated email uses about 0.14 kilowatt-hours of electricity, enough to light up 14 LED bulbs for an hour.
Generating that same response can consume between 40 to 50 milliliters of water, mostly for cooling the servers that process the data.
At scale, this level of consumption raises broader concerns. In Virginia, the state with the highest density of data centers in the U.S., water usage rose by nearly two-thirds between 2019 and 2023. According to an investigation by the Financial Times, total consumption reached at least 1.85 billion gallons in 2023 alone.
As data centers continue to spread across the globe, particularly in areas with cheaper electricity and land, the pressure on local water and energy supplies is expected to grow. Some of these regions may not be equipped to handle the long-term impact.
What your tone teaches the AI
In AI systems trained on large volumes of human dialogue, the tone of a user’s prompt can strongly influence the tone of the response.
Using polite language or complete sentences often results in answers that feel more informative, context-aware, and respectful. This outcome is not accidental.
Behind the scenes, models like ChatGPT are trained on vast datasets of human writing. During fine-tuning, they go through a process known as reinforcement learning from human feedback.
In this stage, real people evaluate thousands of model responses based on criteria such as helpfulness, tone, and coherence.
When a well-structured or courteous prompt leads to a higher rating, the model begins to favor that style. Over time, this creates a built-in preference for clarity and respectful language patterns.
Real-world examples reinforce this idea. In one informal Reddit experiment, a user compared AI responses to the same question framed with and without the words “please” and “thank you.” The polite version often triggered longer, more thorough, and more relevant replies.
A separate analysis published on Hackernoon found that impolite prompts tended to generate more factual inaccuracies and biased content, while moderately polite ones struck the best balance between accuracy and detail.
The pattern holds across languages as well. In a cross-lingual test involving English, Chinese, and Japanese, researchers observed that rude prompts degraded model performance across the board.
Being extremely polite didn’t always yield better answers, but moderate courtesy generally improved quality. The results also hinted at cultural nuances, showing that what counts as the “right” level of politeness can vary depending on language and context.
That said, politeness isn’t always a silver bullet. A recent prompt-engineering review tested 26 strategies to improve AI output. Among them was adding words like “please.”
The results showed that while such phrases sometimes helped, they did not consistently improve correctness in GPT-4. In some cases, adding extra words introduced noise, making responses less clear or precise.
A more detailed study conducted in March 2025 examined politeness at eight different levels, ranging from extremely formal requests to outright rudeness.
Researchers measured results using benchmarks like BERTScore and ROUGE-L for summarization tasks. Accuracy and relevance stayed fairly consistent regardless of tone.
However, the length of responses varied. GPT-3.5 and GPT-4 gave shorter answers when prompts were very abrupt. LLaMA-2 behaved differently, producing the shortest replies at mid-range politeness and longer ones at the extremes.
Politeness also appears to affect how AI models handle bias. In stereotype-detection tests, both overly polite and hostile prompts increased the chances of biased or refusal responses. Mid-range politeness performed best, minimizing both bias and unnecessary censorship.
Among the models tested, GPT-4 was the least likely to refuse outright, but all showed a similar pattern — there seems to be a sweet spot where tone helps the model respond accurately without compromising balance.
In the end, what we say, and how we say it, shapes what we get back. Whether we’re aiming for better answers, less bias, or simply more thoughtful interaction, our choice of words carries weight.
And while politeness might not always boost performance, it often brings us closer to the kind of conversation we want from the machines we’re increasingly talking to.
