The Hidden Cost of Every Query: How We Prompt AI Matters. Part 2: Why it matters for everything else
In Part 1 wrote about being kind to AI, I focused on what those small acts of politeness mean for us as humans. Some readers pushed back with good questions: does tone actually affect the AI’s performance? And what about the environmental cost of our increasingly casual AI usage?
Those questions sent me down a research path. What I found was genuinely surprising, and a little uncomfortable.
So If Rudeness Works, Why Not Just Be Rude?
84.8%
accuracy with very rude prompts
80.8%
accuracy with very polite prompts
Recent findings from Penn State researchers turned the politeness conversation on its head. Their study found that ChatGPT’s 4o model performed better when given rude prompts, achieving 84.8% accuracy with very rude requests compared to 80.8% with very polite ones.
So The Data Days: Be Ruder, Get Better Answers.
But the researchers themselves paused before drawing that conclusion. They cautioned that while aggressive tone might yield more accurate responses, using demeaning language could harm user experience, accessibility, and inclusivity while contributing to harmful communication norms.
Here’s my own answer to the question: the problem with optimising for rudeness is that it asks the wrong thing entirely. Tone is not actually the variable worth chasing. Context is.
The Environmental Equation We Need To Talk About
Before getting to context, there’s something else worth sitting with. Every time we interact with AI, we’re consuming real resources.
10×
A request made through ChatGPT uses 10 times the electricity of a Google Search.
50 GWh
Training GPT-4 alone consumed 50 gigawatt-hours of energy, enough to power San Francisco for three days.
4%+
U.S. data centers consumed 183 terawatt-hours of electricity in 2024, more than 4% of the country’s total electricity consumption. By 2030, that figure could grow by 133%.
Each extra prompt adds to this footprint. When we fire off multiple vague or blunt queries, we multiply demand. Thoughtful prompting reduces computational churn. That’s not a small thing.
Context Is The Real Variable
This is where the conversation shifts from politeness to something more useful.
Research on emotional prompting shows that incorporating emotional intelligence into AI interactions can boost performance by over 10%. But the reason isn’t sentiment, it’s information. When we add context, we give the AI more to work with.
Consider the difference between these two prompts:
“I need a solution to this math problem”
“I’m tutoring high school students and need to explain this concept in a way that builds on algebraic thinking they already have.”
Same underlying request. Vastly different contexts. Dramatically different outputs. The second prompt isn’t kinder, it’s richer. And richer prompts get better results on the first try, which means fewer queries, less computational load, and less of your time wasted.
We’re Not Just Users. We’re Unwitting Contributors.
Here’s the part that stuck with me most.
The way we interact with AI right now, the tone, the structure, the care or carelessness we bring, is feeding into the training data for whatever comes next. If millions of interactions are impatient and vague, that pattern gets baked in. If millions of interactions are specific, thoughtful, and contextual, that gets baked in too.
We don’t usually think of ourselves as shaping the future of AI just by asking it questions. But in aggregate, that’s exactly what’s happening.
What This Means in Practice
I want to move the conversation away from “should I be polite or rude?” because that’s a distraction. The real question is: am I being intentional?
The practical case for thoughtful prompting
- A prompt crafted with care and context gets better results on the first try. That’s not a moral argument. It’s a practical one that happens to also be the responsible choice.
- And on a site built around experimentation and play, getting better results faster means more time for the interesting part: finding out what’s actually possible.
- Every interaction matters. Not because the machine is listening in the way a person would, but because you are.
Ready to Put This Into Practice?
Don't just read experiment. Browse our hands-on experiments designed to help you apply what you've learned.
Browse Experiments