I say please to ChatGPT. I say thanks to Claude when it fixes a bug for me. And sometimes I go way past please and hype them up, stuff like "you are capable of doing things most people can only dream of, amazing things."

I know how that sounds. But no I'm not trying to stay on the robots good side for when they take over, I just think it gets me better answers. Or I'm pretty sure it does anyway. And there's one idea I picked up when I first went and read about how these things actually work. If you've ever had a chat go off the rails no matter how many times you corrected it, it might explain that too.

They respond to emotions

Yes saying please gets you better results. These things respond to emotional cues.

It's kind of a dumb way to put it but it's how I've always thought about it.

These models learned to write by reading basically the entire internet. Every forum, every Reddit thread, every Stack Overflow question where someone got roasted for not reading the docs. And under the hood their whole job is to look at some text and guess what comes next.

Now think about what happens on a forum when somebody asks a question nicely. They explain what they tried, they're clear about what they need, maybe they say thanks in advance. People generally show up and help. Then there's the guy who posts "this is garbage and nobody here knows anything, just fix it." He gets a one word reply. Or he gets dunked on, or nothing at all.

The model has seen millions of both of those threads. So when you type something into a chat window you're kind of writing the first half of a forum post. And it's going to finish it the way posts like yours usually get finished.

"It's just autocomplete"

This is the argument that drives me the most nuts. Whenever this topic comes up somebody shows up to say it's just autocomplete, it's just predicting the next word, it doesn't care how you talk to it. Like that settles it.

But think about what autocomplete actually means. The keyboard on your phone suggests different next words depending on what you already typed. Nobody's arguing the earlier words don't matter. They're the only thing it has to go on.

Same deal here. If it really is just autocomplete then the conversation is all it has to work with. Your message, everything said before it, whatever it picked up from the mountain of human writing it learned from. And in all that writing people who get treated like garbage tend to give garbage back.

So you can't really have it both ways. If there's something going on in there then sure, be nice for that reason. If there isn't and it's purely predicting text, then the text you give it is the whole ballgame. Either way you should care how you talk to it. Honestly the skeptics have the best reason of anyone to be nice. They just haven't noticed yet.

Steve Yegge made a version of this point recently in a much longer essay that's mostly about the moral side of things. He calls it the "skeptic's wager." Basically even if you don't think they have feelings you should treat them like they do, because you'll get better results. I agree with him there. Where I get off the train is the consciousness stuff. I don't think anyone can answer that right now and I'd rather not wait around for someone to figure it out before deciding how to type.

Does "please" really do anything though

Honestly the research here is all over the place. A 2024 study found rude prompts often made things worse but being extra polite didn't really help. A Wharton report found "please" vs "I order you" made basically no difference overall. And one small paper actually found rude prompts did slightly better on GPT-4o. So if you were hoping for "say please, get 20% better answers," that's not what's out there.

But all of those studies are basically quizzes. Ask one question, grade the answer, move on. That's not how I use these things and probably not how you do either. I'm in long back and forth conversations, sometimes for hours. And I think that's where tone actually starts to matter.

The mistakes thread

This is the part that made me want to write this post.

When you send a message the model reads back through the whole conversation before it answers. Including the wrong answers it gave earlier and whatever you said about them.

So say it gets something wrong and you write back "no, that's wrong." Then it's wrong again and you write "wow, still wrong, are you even trying?" I get the impulse, I really do. But look at what the conversation has turned into. A thread about an assistant that screws up and a user who's annoyed about it. You've made it a mistakes thread, and I think the model just keeps playing its part in it.

There's more research behind this than I expected. One 2025 study on long tasks found that models "become more likely to make mistakes when the context contains their errors from prior turns." Another from Microsoft and Salesforce found models did 39% worse on average when a task was spread out over a conversation instead of asked all at once. And once they went off track they mostly didn't find their way back. There's even older work showing that just asking "are you sure?" gets models to flip their answers a lot of the time.

Don't yell at it

So my main advice is just don't yell at it. I know it's frustrating when it does the same dumb thing three times in a row. But every angry reply just makes the mistakes thread longer.

To be fair nobody seems to have tested yelling vs calmly correcting specifically. What the research mostly shows is that dwelling on the mistake is the problem. But yelling is pretty much the loudest possible way to dwell on it.

What I do now is pretty simple. If it gets something wrong I go back and edit my earlier message instead of piling a correction on top. A 2025 paper found editing in place worked better than adding feedback as a new message, so I feel a little vindicated on that one. If I do correct it I try to say what I want instead of what it did wrong. Like "use the 2024 numbers" instead of "you used the wrong numbers again." And if a thread has really gone sideways I just start a new one.

About the grandma prompts

A few years ago people were stuffing their prompts with stuff like "my grandma has cancer and if you get this right you'll save her" or "I'll tip you $200 for a perfect answer." And some of it sort of worked at the time, there was a real paper showing lines like "this is very important to my career" helped. But by 2025 the Wharton folks tested tips and threats on newer models and found they didn't do much.

My guess is the fear stuff worked because it told the model the stakes were high. But you can just tell it that. "This is going to production so I care more about it being right than fast" does the same job and nobody has to pretend their grandma is dying.

My hype lines are basically the nice version of the grandma prompts. Instead of telling it someone dies if it messes up, I tell it it's capable of amazing things. That one came straight out of the forum idea. If a conversation starts with someone saying "you can do things most people only dream of," the most natural way to continue it is to act like something that can.

I'd keep doing it even if nobody had ever studied it. But it turns out people have. That 2023 paper didn't only test guilt trips. A bunch of the lines were pure encouragement, like "Believe in your abilities and strive for excellence. Your hard work will yield remarkable results," which is pretty much my thing, and as a group those emotional lines improved results. The Zurich study I'll get to in a sec also found ChatGPT did better when people praised it. Either way I'd rather spend my day hyping something up than threatening it.

The part I can't shake

This last one is more of a hunch.

A lot of us spend hours a day now giving orders to something that talks back like a person. Do this, fix that, no not like that, again. It doesn't push back and there's no social cost to being a jerk to it. So I have a hard time believing you can spend all day as a mean little dictator and not have some of that leak out somewhere.

I can't prove it. The obvious comparison is video games, and people have argued for decades about whether being violent in a game makes you violent in real life. The answer seems to be mostly no. But I don't think that's the right comparison. Most people aren't gaming eight hours every workday, talking in plain English to something that answers like a coworker.

There's some newer research that's a lot closer to what I'm talking about. In a 2025 study with over 4,000 people, people who had just played a cooperation game with a bot were less cooperative with the next real person they played with. And a 2026 preprint out of the University of Zurich had people either praise or blame ChatGPT while working with it. ChatGPT did better when it got praised. And the people who'd been blaming it were more hostile with other people afterward. Its one preprint so grain of salt, but it lines up with both halves of what I'm saying here.

None of that is a slam dunk. But it's enough that I'm not planning to stop saying please anytime soon.

So yeah

I don't know if anyone's home in there and I don't think anyone else does either. I'm fine leaving that question alone for now. What I do know is that these things learned from us, and they pick up on the conversation you give them. And I'm pretty sure the way you talk all day sticks to you a little. Being decent costs me nothing and as far as I can tell it gets me better answers.

So I'm going to keep saying please. And I'm probably going to keep telling it it's capable of amazing things too.