This sounds ridiculous but hear me out.
For about six months I used ChatGPT the same way I use a terminal. Terse. No greetings, no thanks, just straight commands. "Summarize this." "Fix the bug." "Rewrite this paragraph."
The answers were fast, dense, exactly what I asked for.
Then I read some thread here about people adding "please" and "thank you" and treating it like a person. I figured it was harmless so I started doing it out of habit. "Could you please help me with..." "Thank you, that was really helpful."
Now my answers are longer. Way longer. More hedging. More "of course!" and "great question!" energy in the responses. More caveats I didn't ask for.
I went back to my old terse style for a week and things tightened back up.
I know the obvious explanation is that my phrasing changes are affecting the output style. But it feels weirder than that. When I say "could you please" I'm implicitly giving it permission to be more verbose. When I say "fix this" I'm implicitly demanding precision.
I'm not suggesting it has feelings. I'm saying the linguistic framing of politeness somehow primes a different response mode.
Anyone else noticed this? Is there something in how the model was trained on human conversation that makes polite phrasing activate a different "register"?
[link] [comments]