Saying “Please” And “Thank You” To ChatGPT Is Actually Costing Millions

ChatGPT welcomes kind words, but those “pleases” and “thank yous” are not free. X user, @tomieinlove had tweeted, “I wonder how much money OpenAI has lost in electricity costs from people saying ‘please’ and ‘thank you’ to their models.” Sam Altman responded, “Tens of millions of dollars well spent— you never know.” His post started the conversation all over X by turning a joking question into an unexpected accounting footnote.

Researchers at BestBrokers say ChatGPT needs about 1.059 billion kilowatt‑hours each year, worth about $139.7 million at US retail prices. An Electric Power Research Institute brief adds that every AI query draws roughly ten times the power of a plain Google search, so longer, kinder prompts deepen the draw.

Water leaves the tap as well as the power grid. A University of California, Riverside team found that building a 100 word email uses up to 1.4 litres of cooling water, while a brief “You are welcome” consumes 50ml. Multiply that over half a billion weekly users and the hidden liquid drain looks anything but tiny.

 

What Does The Science Say About Polite Prompts?

 

A cross‑lingual study from Waseda University, RIKEN and NII examined 8 politeness levels across English, Chinese and Japanese tasks. The researchers built graded prompts, from blunt to courtly, then ran them through GPT‑3.5, GPT‑4 and other systems to see how tone shapes output.

Their findings match the power‑bill story… Harsh language drags performance down. Models gave shorter, less accurate replies and sometimes refused awkward queries when the user sounded rude.

Too much flattery, though, did not guarantee top marks. In English runs, middling courtesy worked best, with scores dipping once praise turned flowery.

The authors warn that prompt tuning must account for cultural context. Japanese keigo has a formal hierarchy that Western users may overlook, and the study shows each tongue has its own sweet spot between brusque and obsequious. Ignoring that nuance can make well‑meaning users pour words onto the page with little gain.

 

 

Does Language Change The Sweet Spot For Courtesy?

 

The paper’s data show some contradictions… English models hit their peak with very polite requests, then slid as warmth faded, then the fall grew steeper below level 3. Chinese runs favoured a gentler middle ground and faltered when phrases grew either too sugary or downright rude.

Japanese output followed a hill shape in reverse. Marks went up when users dropped a little formality, dipped in the middle, and climbed again for complete informality, unless etiquette collapsed altogether.

Cultural habits help to explain the curves, in Japan, shop staff answer even casual chat with respectful speech so an LLM trained on that material may treat mid‑level politeness as routine rather than respectful. English data sets, packed with light conversational writing, read extreme courtesy as extra care.

The study also tracked bias. In English, racism and sexism rose at both ends of the politeness scale, hinting that forced nice words or open scorn alike unlock hidden prejudice. Chinese and Japanese patterns echoed this but at milder rates.

Researchers add that reinforcement learning from human feedback narrows these swings. The fully tuned Llama‑2 model showed lower bias than its base twin, though it still stumbled when faced with the coldest insults. That insight matters for prompt writers who juggle fairness with precision.

 

Can They Keep Paying For Nice Words?

 

OpenAI recently closed a $40 billion funding round that placed its valuation at $300 billion, according to company statements, and it reports 500 million weekly users. Thid funding makes the electricity costs Altman joked about look small in comparison.

Even with deep pockets, power planners feel nervous. The Electric Power Research Institute says an AI chat uses 10 times as much energy as a standard Google query. If kind phrasing lengthens each request, grids will feel the strain sooner, and utilities must size next‑generation plants accordingly.

Altman calls courtesy money well spent. The next thing is if developers can trim chips and cooling without asking users to drop their manners. Until then, every extra “please” and “thank you” keeps the costs up.