In the age of large language models, automation, and AI, even human habits that seem benign have unexpected consequences. One such case is the energy and financial cost triggered when users express courtesy towards OpenAI’s ChatGPT. Although saying “thank you” to an AI model may seem harmless, at the scale of millions of users, these minor niceties result in a measurable and growing burden on infrastructure, cost, and the environment.
At the core, the cost lies in the intensive computational power behind AI systems. ChatGPT, based on large language models like GPT-4, processes every input, whether a complex question or a simple “thank you”, using clusters of high-performance GPUs across data centers. Each response is freshly generated, not pulled from a database, requiring trillions of floating-point operations per token.
An average ChatGPT interaction consumes approximately 0.0029 kilowatt-hours (kWh) of electricity, nearly ten times that of a typical Google search. Each “thank you” processed results not only in handling the input text but also in generating a polite response like “You’re welcome!”, adding further computational demand. These exchanges are token-heavy; even short phrases add multiple tokens to be processed and generated.
Scaling this cost across the user base reveals the hidden enormity. With an estimated 400 million weekly users generating roughly 1 billion messages per day, ChatGPT consumes about 2.9 million kWh of electricity daily. Annually, this amounts to approximately 1.06 billion kWh, or 1.06 terawatt-hours. Assuming an average commercial electricity rate of $0.13 per kWh, this translates to around $138 million spent annually purely on electricity for ChatGPT’s operations.
Sam Altman, CEO of OpenAI, alluded to “tens of millions of dollars” being lost to the added computation resulting from polite interactions. If we conservatively assume that polite phrases, like “thank you,” “please,” etc., and their resulting responses account for 10-20% of ChatGPT’s token generation, the cost attributable to these niceties alone could range from $14 million to $28 million annually.
Consider a simple breakdown — A “thank you” input plus a “you’re welcome” output could add roughly 10–15 extra tokens per interaction. Each token processed requires an incremental slice of computational work, drawing server energy. Across millions of interactions, the energy drain from these tokens becomes significant. If removing niceties could reduce the average interaction length by even 10%, ChatGPT could potentially save around 100 million kWh per year—an equivalent cost saving of $13 million annually.
Generating a single token requires heavy-duty computing. For models like GPT-4, each token generation involves billions to trillions of operations. Each user message triggers a full inference process through the model’s vast neural network. There is no “shortcut” for simpler queries; the model’s architecture engages at full tilt regardless of the message’s brevity or complexity.
The average energy consumption for generating a 100-word email, about 75 tokens, is estimated at around 0.14 kWh, which is enough to power 14 LED bulbs for one hour. Even generating a three-word “You’re welcome” reply consumes a measurable fraction of this energy.
Beyond financial costs, there is a notable environmental footprint. The global AI industry is becoming a significant consumer of resources. Data centers already consume around 2% of global electricity, and AI’s demands are increasing this share.
Each 100-word response requires about 1.4 liters of water for server cooling. ChatGPT’s daily operations, based on current user volumes, could require up to 39 million gallons or 148 million liters of water daily. Water consumption becomes a critical concern in drought-prone areas where many data centers operate.
In terms of carbon emissions, if ChatGPT consumes around 1.06 billion kWh annually, and assuming a moderate grid carbon intensity, the model could be responsible for several hundred thousand tons of CO2 emissions per year. Courtesy-induced additional computation, comprising around 10% of this load, could thus account for tens of thousands of tons of avoidable emissions.
Politeness doesn’t only cause marginal token increases—it often results in more verbose and elaborate responses. ChatGPT, trained on human conversational norms, mirrors politeness. When a user is courteous, the AI tends to respond in a similarly courteous manner, often adding phrases like “I’m glad I could assist you!” or “Hope you have a great day!”. Each such flourish adds tokens, thereby multiplying the computational cost.
This “politeness inflation” results in longer sessions and more intensive resource use per user interaction. It introduces a positive feedback loop where human courtesy leads to AI verbosity, exacerbating energy and financial overhead.
AI systems could be engineered to recognize terminal politeness phrases like thank you and conclude sessions with minimal computation. Lightweight responses for simple, polite interactions could be generated using specialized low-power routines. Infrastructure improvements, such as better cooling systems, renewable energy sourcing, and more efficient model architectures, could mitigate environmental impacts.
Despite these possibilities, there remains a tension between maintaining a human-like, friendly AI experience and optimizing computational efficiency. Sacrificing politeness for efficiency could degrade user satisfaction, which is crucial for mass adoption and engagement.
At this massive scale, even small behavioral nuances like politeness towards AI systems have substantial economic and ecological consequences. Each “thank you” users offer to ChatGPT contributes to a cascade of computational tasks, consuming extra electricity, water, and emitting more CO2. While the cost per interaction is minute, the aggregate impact is undeniable, amounting to tens of millions of dollars and significant environmental strain annually.
Author: Rifat Ahmed
