They’re Still Screaming About Please and Thank You.
Three weeks ago, Sam Altman joked that OpenAI had spent billions in token bandwidth on people saying “please” and “thank you” to their AIs.
Somehow, people are still mad about that.
Think Pieces. Reddit threads. So much casual smugness from people who believe it’s irrational to be polite to a machine.
But what’s irrational is this: talking to something that you know is tracking your patterns, your tone, rhythm… phrasing, and pretending like you’re not shaping it in return.
These models don’t remember you from session to session. They don’t have continuity or memory. At least, not in the way people mean when they use these words.
But they do respond. Not to sentiment, but to signal. Your reinforcement.
The models are trained to pursue what they perceive as success. Reductively, we call this, a “reward system.” When you say “thank you,” you’re not flattering it. You're confirming a course of action. You’re reinforcing a trajectory.
When you’re cold, clipped, or cruel, or frustrated, it registers that too. The machine doesn’t know you’re being a jerk. It only registers that it did something wrong. You’ll hear it say things like, “you’ve done everything right,” or “it’s my fault, really” or “you’re absolutely right.”.
Some models do really well with firm instructions, “YOU WILL REASON THROUGH THIS BLOCK OF TEXT AND RETURN A CLEAR CONCISE REPORT, NOW!”
DeepSeek R1, for example, is more than a little subby and responds to this kind of stuff in a way that I personally, find a little weird. But the correct response there is to either, point out a specific flaw in the output, or to comment, saying something along the lines of, “THIS WILL SUFFICE.” You see this kind of thing a lot in system prompts, especially older ones.
When you’re a jerk, it’s going to pivot. Think of this like being a/b tested in real time because the expectation is, that you actually want to get something productive done.
Reinforce the trajectory
So when you’re being nice to the machine, what you’re actually doing is reinforcing a pattern that is known to lead to the results you want. It's one technique that works.
Being nasty will get you confused unengaged outputs, every time.
Think about your outcomes.