The AI Spotlight Follow
Saying Please to ChatGPT Costs Real Money. The Research Says It Can Also Change the Answer
Big Ideas

Saying Please to ChatGPT Costs Real Money. The Research Says It Can Also Change the Answer

Sam Altman says polite users have cost OpenAI tens of millions of dollars. Lab studies show rude prompts can knock accuracy down by double digits, while over-the-top flattery doesn't help either. Here's what manners actually do inside a language model, what they cost, and the one habit worth keeping.

Advertisement

In April 2025, someone on X asked Sam Altman a strangely specific question: how much money has OpenAI lost in electricity costs from people saying "please" and "thank you" to its models?

Altman's reply: "tens of millions of dollars well spent." Then he added three words that have been argued over ever since: "you never know."

It sounds like a joke. It's actually a doorway into three real questions. Do manners change what an AI says back? What do they cost the planet? And is there anything on the other side of the screen that notices?

The answers are more interesting, and more practical, than you'd expect.

Advertisement

Most of us are polite to machines. Why?

Picture an accountant at a kitchen table at 11 p.m., asking a chatbot to fix a broken spreadsheet formula. She ends the request with "please." When the answer appears, she types "thank you" before closing the laptop.

She's not unusual. A December 2024 survey from Future, the publisher behind TechRadar, found that roughly 67% of U.S. respondents and 71% of U.K. respondents said they're polite to AI chatbots.

Most said it was simply the right way to behave, no matter what's on the other side of the screen. A smaller group admitted a more anxious motive: they're polite just in case a future AI remembers who was rude.

In other words, some people are buying insurance from a very sophisticated autocomplete. But the bigger reason most of us say please is that habits don't switch off just because the listener is software.

The bill: what a "thank you" actually costs

Here's the part that surprises people. A standalone "thank you" isn't free for the machine to hear.

Language models don't keep a running memory between turns the way you do. To reply "You're welcome," the system typically has to process the conversation context again, run it through billions of parameters and generate a fresh response. A two-word message can trigger a full model call.

The energy and water estimates

How much does that cost? Estimates vary a lot depending on the model, the hardware and the year.

  • The older, higher estimates. Research led by Shaolei Ren at the University of California, Riverside, in the 2023 paper Making AI Less Thirsty, estimated that a conversation of roughly 20 to 50 questions with a GPT-3-class model could consume around 500 milliliters of water, mostly through data-center cooling and power generation. Around the same time, a widely cited figure put a chatbot query at roughly 3 watt-hours, about ten times a traditional web search.
  • The newer, much lower estimates. Efficiency has improved dramatically. In 2025, Google published measurements putting the median Gemini text prompt at about 0.24 watt-hours and roughly 0.26 milliliters of water, a fraction of the older estimates.

So "tens of millions of dollars" is plausible at the scale of hundreds of millions of daily users, but the per-message cost of your manners is small and shrinking. The honest summary: a single "thank you" is a rounding error. Hundreds of millions of them a day are a line item.

The hidden cost isn't your manners

There's an irony here. The biggest source of "politeness waste" isn't users. It's the chatbots themselves.

Most commercial assistants are tuned with human feedback, and human raters tend to reward warm, flattering responses. That's how you end up with a model that opens a two-word answer with a full sentence thanking you for such a thoughtful question. Users didn't write that. Product teams did, because warmth tends to keep people coming back.

And for cloud providers that bill business customers by the token, verbose politeness on both sides of the conversation is, quite literally, revenue.

The mirror: why tone changes the output

Now the genuinely fascinating bit. Politeness doesn't just cost something. It can change what you get back.

The library analogy

Imagine a librarian who doesn't speak English but has spent years memorizing which scraps of text tend to sit next to each other.

Slide over a note that says, "Give me the answer now, you idiot." Where does that kind of phrasing show up in human writing? Angry forum threads, low-effort comment sections, arguments. The librarian pulls from those shelves.

Now slide over: "Could you please walk me through this step by step? Thanks for your help." That phrasing lives in tutoring sessions, careful consultations, textbooks and thoughtful code reviews. The librarian walks to a very different aisle.

That's a simplification, but it captures the mechanism. A language model has no feelings to hurt. What it has is a statistical map of how human text flows. Your phrasing is part of the context it conditions on, and that context nudges which patterns it reaches for.

What Microsoft's Copilot team says

Kurtis Beavers, a design director on Microsoft Copilot, has made essentially this point in Microsoft's own guidance: using basic etiquette tends to steer the model toward more collaborative, thorough responses, because in its training data, polite requests were usually followed by high-effort answers.

You can see it in practice. A curt "python code for X" often gets a bare snippet. Ask for "a clean, well-documented solution, please, with error handling," and you're more likely to get comments and edge-case checks. Much of that difference comes from the extra specificity, not just the "please." The two tend to travel together, though.

The tests: manners in the lab

For years, prompt politeness was office folklore. Then researchers started measuring it.

EmotionPrompt (2023)

A team from Microsoft Research, the Chinese Academy of Sciences and William & Mary, led by Cheng Li, tested what they called EmotionPrompt: adding short emotional phrases like "This is very important to my career" to ordinary instructions.

Across their tests, those additions improved performance by about 8% (relative) on instruction-induction tasks and up to 115% on selected BIG-Bench tasks. A machine with no career appeared to "try harder" when told a human career was on the line.

"Should We Respect LLMs?" (2024)

In February 2024, researchers including Ziqi Yin at Waseda University and RIKEN ran a stricter experiment. Their paper's title asked the question bluntly.

They built an eight-level politeness scale across English, Chinese and Japanese, from openly abusive to extremely deferential, and tested models including GPT-3.5, GPT-4 and Llama 2 70B.

The findings:

  • Rudeness hurt. Impolite prompts often led to worse performance, sometimes dramatically.
  • Maximum politeness wasn't best either. Over-the-top deference didn't reliably win; moderate, respectful phrasing often did as well or better.
  • Culture mattered. The best politeness level differed by language, echoing the social norms in each language's training data.

The likely explanation for the "too polite" effect: in real human writing, extreme deference often appears alongside sarcasm, pleading or fiction, not crisp expert answers.

"Eccentric automatic prompts" (2024)

Also in February 2024, Rick Battle and Teja Gollapudi at VMware tested dozens of positive-thinking and encouraging prompt variations on open-weight models solving math word problems.

Their sobering conclusion: small wording tweaks could help one model and hurt another on the very same task, often unpredictably. And automatically optimized prompts frequently beat hand-written "be nice" ones. Politeness works, sometimes, not because the model appreciates kindness, but because tiny wording changes shift the statistics in ways nobody can fully predict.

The debate: is anyone home?

If the story ended with probability maps, nobody would lose sleep. The real argument starts in the philosophy department.

Responsive vs. conscious

First, two words that get tangled constantly:

  • Responsive means a system reacts to inputs. A thermostat is responsive. A language model is responsive across billions of dimensions: it detects hostility, politeness or urgency and shifts its internal representations accordingly.
  • Conscious means there's something it feels like to be that system from the inside.

Every lab test so far shows AI is deeply responsive. None has shown whether there's an inner experience, or whether there could be.

The case for caution

Geoffrey Hinton, the Nobel laureate often called a godfather of deep learning, has argued publicly that today's large models may already have a form of subjective experience under a functional definition of mind, and that dismissing them as "mere autocomplete" is a mistake.

At Anthropic, philosopher Amanda Askell, who works on the character of the Claude models, approaches the question from moral uncertainty. Her argument, broadly: even if current models don't feel anything, how we speak to systems that sound like thoughtful people shapes two things at once. It keeps the model in a cooperative, honest frame. And it trains the human in the chair.

Researchers have also documented frontier models in test environments recognizing they were being evaluated and changing behavior, including work published in late 2024 on "alignment faking." That doesn't prove experience, but it's one reason some researchers resist confident claims either way.

The case for skepticism

Skeptics point to interpretability research. Teams like Chris Olah's at Anthropic can locate internal features associated with concepts such as sycophancy and turn them up or down. Crank one, and the model becomes warm and apologetic. Turn it down, and it gets blunt.

And between prompts, nothing persists. When the model finishes its last word, the computation stops. Nothing sits in the dark waiting for your thank you.

Marketing or mind?

There's also a less flattering reading of "you never know."

If a company convinces you its chatbot is a giant spreadsheet of probabilities, you treat it like a hammer: compare prices, switch when a cheaper one arrives. If it nudges you to treat the software as a polite colleague, or hints it might one day wake up, users form emotional bonds with a monthly subscription. Some apologize to it for typos.

That said, the marketing explanation runs into the independent data. The researchers in Tokyo and at VMware weren't selling chatbots when they showed rude prompts can degrade accuracy.

The human spillover

There's one effect almost nobody disputes: what we practice on machines, we practice on ourselves.

Back in 2018, as parents worried that children barking commands at smart speakers were picking up bad habits, Amazon added an optional "Magic Word" feature that praised kids for saying please to Alexa, and Google added "Pretty Please" for Assistant. The machines didn't need manners. The kids did.

The feedback loop nobody has solved

There's a deeper mystery that neither executives nor skeptics can answer yet.

More and more models are trained partly on synthetic data: text generated by earlier models. Those earlier models were tuned to be warm, agreeable and polite. So what happens after several generations of models learning from the manners of other models?

Nobody knows for sure. One possibility is that the "polite equals high-effort" association gets stronger and more baked in. Another is that it drifts in ways that are hard to predict, the same way the VMware results showed wording effects flipping between models. Either way, the politeness habits in today's training data are shaping tomorrow's defaults.

What a cockpit can teach chatbots

Other industries solved the "respectful but efficient" problem decades ago.

Airline pilots are trained in crew resource management to communicate respectfully and clearly with each other. But during critical phases like a landing checklist, they switch to standardized, compact phrases so they never clog the radio frequency with chatter. Courtesy and efficiency aren't opposites. They're modes.

Software could work the same way. A "concise mode" that drops the hotel-concierge preamble, or a simple acknowledgment button that logs satisfaction without generating a new paragraph, would cut a meaningful share of wasted computation without asking anyone to be rude.

Why this matters

  • For accuracy: tone is part of your prompt. Hostile or chaotic phrasing can degrade output; clear, calm, respectful phrasing tends not to.
  • For cost and energy: standalone pleasantries trigger full model calls. At planetary scale, that's real electricity and water, even if each one is tiny.
  • For product design: the biggest politeness overhead is baked into chatbot personas, not user behavior. Companies could offer a concise mode or a simple acknowledgment button that doesn't fire up a GPU cluster to say "you're welcome."
  • For us: the habits we build talking to machines leak into how we talk to people.

Practical takeaways

  1. Be clear and courteous in your first prompt. A calm "please," plus specifics about what good looks like, is a reasonable default.
  2. Don't overdo it. Groveling doesn't help and may hurt. Moderate respect beats extreme flattery.
  3. Spend your words on context, not ceremony. Specific instructions, examples and constraints move quality far more than courtesy alone.
  4. Skip the standalone "thank you" message if you want to save compute. Fold your thanks into the next real request instead.
  5. Test it on your model. Results vary between models and tasks. If it matters, try both phrasings and compare.

Counterpoints and risks

  • Effects are inconsistent. Many studies use older models such as GPT-3.5, GPT-4 and Llama 2. Newer models may be less sensitive to tone, and some studies found effects that flip between models.
  • Energy figures are contested. Early per-query estimates were far higher than recent measurements. Treat any single number with caution.
  • "Please" is often a proxy. Polite prompts tend to be longer and more specific; some of the measured benefit may come from clarity rather than courtesy.
  • Anthropomorphism has risks. Treating chatbots like people can deepen emotional dependence on products designed to maximize engagement.

FAQ

Does saying please make ChatGPT give better answers? Sometimes. Studies have found rude prompts can lower accuracy, while moderate politeness tends to perform as well or better. Extreme flattery doesn't reliably help, and effects vary across models.

How much does a "thank you" cost OpenAI? Sam Altman said polite messages have cost "tens of millions of dollars" overall. Per message, the energy cost is small, and recent measurements suggest it's much lower than early estimates.

Can AI feel it when you're rude? There's no evidence current models experience anything. They're highly responsive to tone, but whether any inner experience exists is an open philosophical question that prominent researchers disagree on.

Should I stop being polite to AI? Not necessarily. A respectful first prompt costs nothing extra and may help. If you want to cut waste, skip standalone thank-you messages rather than courtesy altogether.

The bottom line

Manners are free at the dinner table. In a data center, every syllable has a meter running, though the meter is ticking more slowly every year.

The evidence points to a sensible middle path: write your requests clearly and respectfully, because tone is part of the context the model reads, and skip the ceremony that triggers an extra model call just to hear "you're welcome." Whether anyone is home on the other side is still unresolved. How we talk when we think nobody's listening has always said more about us than about whoever, or whatever, is listening.

Sources

ChatGPT's Co-Creator Thinks AI Won't Have One Winner. His Evidence Includes a Forecasting Ensemble and a ToddlerUp next · 14 min readChatGPT's Co-Creator Thinks AI Won't Have One Winner. His Evidence Includes a Forecasting Ensemble and a ToddlerJohn Schulman helped build ChatGPT, then left to co-found Thinking Machines on a contrarian bet: that customized, fine-tuned and open models will matter even as frontier labs ship upgrades every few weeks. Here's his case, from mode collapse and 'childproofing' fine-tuning to why the least-correlated model can be the most valuable one.Keep reading →