The Politeness Paradox: Why Being Rude to AI Unlocks Better Results

The Cost of Courtesy

When Sam Altman casually mentioned that people saying "please" and "thank you" to ChatGPT costs OpenAI tens of millions annually, it sparked a firestorm. On the surface, it's an engineer's nightmare: wasted compute on social fluff. But dig deeper, and you hit something far more interesting. That 2024 Waseda University study didn't just measure watt-hours; it measured output quality. Impolite prompts led to more bias, more errors, more refusals. But here's the kicker: moderate politeness consistently beat both extremes. It's not about being nice—it's about triggering the right data patterns.

How the Model Thinks

Think of ChatGPT as a vast library of human text. When you write "Could you help me structure this analysis?", you're pattern-matching to professional, well-reasoned documents—think academic papers or detailed reports. When you bark "give me the answer," you're pulling from Reddit threads or quick Q&A forums. The model isn't understanding politeness; it's retrieving context based on linguistic cues. As Google DeepMind's Murray Shanahan put it, the model role-plays a smart intern. Treat it like a colleague, and you get colleague-quality work. Bark orders, and you get minimum-viable compliance.

"The model is role-playing a smart intern. Treat the intern like a colleague, you get colleague-quality work. Bark orders, you get minimum-viable compliance." — Murray Shanahan, Google DeepMind

The Numbers Behind the Habit

Let's run the math. OpenAI handles over a billion queries daily. Each GPT-4 query uses roughly 2.9 watt-hours—about ten times a Google search. Add a "please" and "thank you," and you're tacking on extra tokens, each burning more compute. But here's the reality check: OpenAI just raised $40 billion at a $300 billion valuation. Those tens of millions in politeness tokens? It's a rounding error on a rounding error. The real cost isn't financial; it's behavioral. 67% of Americans use polite phrases with AI, and 55% say it's "the right thing to do." They're maintaining a habit that governs every human interaction.

• Compute Impact: Extra tokens increase processing load, but at scale, it's marginal compared to overall infrastructure costs.
• Quality Trade-off: Polite prompts access higher-quality training data, reducing errors and bias in responses.
• User Psychology: Habits from human interactions bleed into AI use, reinforcing social norms even in transactional contexts.

The System Architecture View

From a systems perspective, this isn't about optimizing away "please." It's about designing for emergent behavior. AI models are trained on human data, so they inherit our quirks. When you prompt politely, you're essentially doing a more precise semantic search through that data. The model's attention mechanisms weight words differently based on context, and politeness acts as a subtle steering signal. This isn't a bug; it's a feature of how transformer-based architectures work. Ignoring it means missing out on performance gains.

Why This Matters Beyond the Chat Window

Telling 900 million people to stop saying thank you to save 0.01% of operating costs is peak engineer-brain thinking. It misses the bigger picture: we're training ourselves in how to interact with intelligence—artificial or otherwise. The parent teaching their kid to say please to Alexa isn't doing it for the AI; they're doing it to avoid raising someone who thinks rudeness gets faster results. This habit doesn't stay in the chat window. It shapes how we collaborate, delegate, and communicate in an increasingly automated world. The politeness paradox reveals that our social instincts are, ironically, optimizing for better machine output.

// Example prompt comparison
// Polite: "Could you explain the concept of attention in transformers?"
// Impolite: "Explain attention now."
// The first likely retrieves from academic sources, the second from forums.

So, where does this leave us? Politeness in AI prompts isn't a waste—it's a low-cost hack for better results. Yes, it adds compute overhead, but in the grand scheme of OpenAI's resources, it's negligible. The real insight is that human habits are becoming part of the system architecture. As we build more advanced agents, understanding these nuances will be key to designing interfaces that feel natural and perform optimally. Don't stop saying please; just know why it works.