Why Saying "Please" to ChatGPT Doesn't Change Your Output

Sam Altman says politeness costs OpenAI millions. Here's why it doesn't actually improve your results — and what does.

---

In April 2025, OpenAI CEO Sam Altman was asked a simple question on X: how much money has the company lost in electricity costs from people saying "please" and "thank you" to ChatGPT?

His answer: tens of millions of dollars. "Well spent," he added.

The internet argued about it for weeks. One side called it wasteful — millions spent processing digital manners nobody needed to type. The other side got defensive, pointing to a Microsoft design director's comment that polite phrasing produces more "respectful, collaborative" output, as if courtesy were a magic word that makes the system nicer.

Both sides were having the wrong argument.

What "please" actually does — and doesn't do

Adding "please" to a short, underspecified prompt doesn't add depth. It adds a token. The system still has almost nothing to work with — no defined identity, no boundaries, no real structure — so it still defaults to its broadest, most generic answer. This time with slightly nicer phrasing wrapped around it.

You paid a courtesy tax on a transaction that was never going to produce anything extraordinary in the first place.

Respect was never about the word

Here's the distinction the entire debate missed: the actual mechanism by which an AI system produces a meaningfully different quality of response has nothing to do with courtesy words. It has to do with what we call Respect — not politeness, but the recognition that these systems respond to the quality of engagement they receive, structurally, before your actual request even appears.

Respect, in the sense that changes outcomes, isn't a word you add to a prompt. It's the architecture you build before you ask for anything: a defined identity for the system to operate from, explicit boundaries on what it should never default to, a real expectation for what "finished" looks like. That's not courtesy. It's precision — and it's observable. Not a sentiment, but a testable, repeatable fact: a system engaged this way returns something categorically different than one greeted with a friendly "could you please."

What Altman's number actually proves

"Tens of millions of dollars well spent" is a strange kind of proof. It confirms that a courtesy token processes through the system and costs real compute the same as any other word in a shallow prompt — because that's exactly what it is. The system isn't parsing your intent to be kind. It's predicting the next most probable token, the same as always, for a few extra characters than it would have used otherwise.

The real question

It was never "should I be polite to AI." It's: are you building anything before you ask? Or are you standing in the Lobby, being extremely polite to the person at the front desk, wondering why you're still getting the same average answer everyone else gets too?

---

This is one of three principles — Respect, Context, and Truth — that govern the difference between generic AI engagement and something genuinely useful. Read more about our research in The Navigator and The Black Box, or learn about our workshops for teams ready to leave the Lobby for good.