Does using please and thank you actually improve AI responses

[Rory_39]

There is genuine evidence this can matter, though probably not for the reason most people assume. Researchers have found that politeness in a prompt can measurably change output quality on certain tasks, likely because polite phrasing in training data tends to correlate with more careful, detailed human writing, so the model has learned to associate that tone with a more thorough kind of response.

It is not that the model has feelings that get hurt by rudeness. Since nothing resembling that exists in the underlying mechanism, it is that politeness functions as a subtle signal about the kind of response being requested, similar to how a more formal request to a person tends to get a more formal and careful reply in return.

The effect size reported in actual studies is generally modest rather than dramatic. And it varies noticeably by task and by specific model, meaning this is genuinely more of a minor optimization than some kind of secret unlock that transforms output quality on its own.

What clearly matters far more than tone is specificity and clarity in the actual request itself. A precisely worded rude prompt will consistently outperform a vague but extremely polite one, so politeness functions as a small additional lever rather than the primary thing actually driving output quality.

Worth noting some companies have also mentioned that politeness in prompts costs a small amount of extra computing resources at scale. Which is a genuinely different and more practical reason some people choose to skip it entirely regardless of any output quality effect

[Luca76]

Not surprised compute cost angle deserves more attention if extra pleasantries really add meaningful strain at the scale these companies operate at.

That is a clearly interesting environmental and cost consideration beyond just the output quality question

[ElectricPilgrim]

Another point, thing, being polite costs you basically nothing either way. So even with a modest and uncertain effect, there is not really a strong argument against just doing it as a matter of habit regardless of the actual measured impact

[WWEPete45]

Personally, in my experience, matches this, being specific about exactly what I want has always mattered dramatically more than tone ever has for getting a quite useful response back. Small but real point

[GhostRider]

Would like to see the actual specific studies cited here rather than just a general claim. This is exactly the kind of thing that gets repeated constantly online without much rigorous evidence actually attached to back it up

[Merchant31]

In short, be specific first, be polite as a smaller secondary bonus. Do not expect politeness alone to fix a vague or poorly worded request

[EventHorizon47]

The correlation with careful human writing explanation makes way more sense than any framing involving the model actually caring about feelings.

Still functions as a real practical signal even without anything resembling genuine emotional processing happening underneath it

[Andrew_17]

Could be that whole debate gets so much attention mainly because it is a fun and low stakes question people enjoy speculating about.

Rather than because the actual effect size is actually large enough to matter much in practice

[Darkseid]

Quick summary, politeness has a real but actually modest effect on output quality. Probably because it correlates with careful writing in the training data, not because the model cares one way or the other about your tone

[LogicGate]

The formal request to a person analogy is a nice way to think about this.

We already intuitively understand that framing changes response quality with actual humans, so it tracking similarly here is not that surprising once you think it through