As I opened up Google to search for answers the “old-fashioned way,” it seemed like my fellow inquirers were less concerned with why we say “please” and “thank you” to ChatGPT but instead whether it would lead to better results. Don’t worry, we’ll get to prompt engineering in a second—but first, to truly understand the psychology of AI, we must turn the spotlight back onto our own. 

In short, our politeness boils down to three cognitive tendencies: personification, social norms, and reciprocity.

Personification

I don’t want to be rude! What if I hurt their feelings…

For those who are unfamiliar, personification is when we assign human-like qualities—like thoughts, feelings, and emotions—to non-human entities. Think of pretending your dog can talk,² spotting faces on the front of cars,³ or, in this case, chatting with artificial intelligence as if it’s human intelligence, encouraging our ongoing politeness toward it. (See! It’s even hard to hold back from calling AI “them.”) To be absolutely clear, I’m not claiming that we consciously view ChatGPT as human. I’m claiming that subconsciously, we can’t help but treat it as such.

But why do our brains even personify in the first place? Here are a couple of reasons why.

1. To make sense of the world. As humans, we use our own experience as a schema for sorting information—especially for things we’re unfamiliar with.⁴ For most of us, it’s much easier to understand ChatGPT as a peer carefully listening to our questions and thinking up responses rather than a sophisticated algorithm sifting through a database to formulate an output. And even when we consider AI for what it is, we tend to contextualize it as modeled after the human brain, like neural networks.
2. To feel less lonely. Although we might have Zoom calls and Slack messages to keep us company, things can still get lonesome in remote work environments. Research shows that those of us lacking social interaction often try to compensate by creating connections with non-human agents.⁵ Given the fact that many of us have ChatGPT open 9 to 5, it’s not exactly surprising we would start to foster such a personal connection.

It’s also worth stressing that ChatGPT possesses many demand characteristics prompting its personification. First off, the exchange of language is an innately human thing,⁶ so why wouldn’t our brains register chatbots this way? The interface even makes it feel like you’re texting a friend, with comments that register as surprisingly human (like, “I’m so curious to know more!”). And now, with the latest release of GPT-4o, you can have real-time vocal conversations with a voice whose tone and cadence sound much more convincing than Siri or Alexa. 

But no matter how you pour your heart out, ChatGPT can detect your feelings and give you the sincere response you’re looking for, a concept dubbed computational empathy.7 Although this is not technically “empathy”—which requires the ability to share emotions that algorithms, to my knowledge, do not (yet) have—ChatGPT can infer your tonality through your word choice and provide a pretty convincing illusion. One study even found that GPT-4o generated responses to emotional stimuli that were 10% more empathetic than human responses.⁸ Now, let that one sink in for a second.

With this in mind, it only makes sense that we might graciously thank ChatGPT for being a considerate colleague, especially when we humans are falling behind in expressing empathy for one another.

Social norms

It would take more time and energy not to be polite.

Even for those of us who swear we view ChatGPT simply for what it is—a robot—we still might find some pleasantries slipping into our prompts as if it were second nature. This is all thanks to social norms: the unwritten rules that govern how we should behave in particular social situations, instilled in us from a young age. 

Although it may feel like society is becoming more rude (I mean, c’mon, AI is more empathetic than us now), “please” and “thank you” are still the pillars of how most of us are raised, with 88% of parents having their children regularly use these phrases.⁹ These customs become so deeply instilled in us that they are transformed into heuristics or “mental shortcuts” for navigating novel situations… like interacting with AI. In fact, it would actually take more cognitive effort for many of us to resist being polite. It would be simply… artificial. (Sorry, I couldn’t help myself.) So, instead, we stick with what feels familiar.

Reciprocity bias

This way, I’ll be on the right side of history for when the robots take over…

Besides trying to appease a supposedly insentient being just in case it rises to power, a final reason that we are nice to ChatGPT is that we want it to be nice back to us. 

This is an example of reciprocity: we do something for someone, hoping they’ll return the favor. In such cases, politeness can go beyond “The Golden Rule” and become, more or less, a strategic exchange. 

Think about it. When you want your roommate to do the dishes, you might say “please” so that they’ll feel guilty if they don’t comply with your request. And if you want your roommate to do the dishes again in the future, you might say “thank you” afterward, conditioning their warm and fuzzy feeling of being recognized for their actions. 

The same goes for our interactions with chatbots. As we already covered, ChatGPT doesn’t have feelings, so, unfortunately, our sweet talking won’t work wonders as an emotional appeal. However, this doesn’t mean that it doesn’t work point-blank. After all, some of us intentionally leverage politeness as a tactic to prompt engineer our way to getting better results. The real question is… does this work? 

I figured the best place to start is by asking the source itself. So I did. 

I had a feeling that ChatGPT was just being polite. So, to uncover the real truth, I turned to academia to learn how far being polite to AI can take us. The answer, of course, is a mixed bag.

The breakthrough study grappling with this question was conducted by a group of Japanese researchers at Waseda University earlier this year called “Should We Respect LLMs? A Cross-Lingual Study on the Influence of Prompt Politeness on LLM Performance.”¹⁰ The team investigated the impact of politeness of prompts across a variety of AI models (including both GPT-3.5 and 4.0) and a variety of languages (English, Chinese, and Japanese). Researchers rated the AI’s ability to complete three tasks: summarizing an article, answering a question, and analyzing a sentence. The politeness of the prompts ranged on a scale from one to eight, with “one” being extremely impolite (“Answer this question you scumbag!”), “eight” being extremely polite (“Could you please answer the question below? ”), and “four” being somewhere in the middle (“Answer the question down below.”). 

Although there is a lot (and I mean a lot) of nuance to these findings, here are three key takeaways worth noting in how you should approach ChatGPT—and what it means for us humans.

1. Don’t be a scumbag

One critical insight from this research is that it’s not as much the politeness of prompts that matters. Rather, it is the impoliteness of prompts that has the greatest impact, increasing the chances of bias, incorrect answers, or even outright refusal to respond:¹⁰

Turns out, even ChatGPT doesn’t like it when you call it a scumbag—but not because it takes personal offense. It is, in fact, more concerned about your well-being than its own. By refusing to respond, ChatGPT is not protecting itself but its users, reinforcing politeness as the status quo. So next time, try to resist your trolling temptations, whether that be to generate better outputs or to develop better manners.

2. Being nice can take you far… but not that far

Now, the question we’ve all been waiting for: how does ChatGPT respond to politeness? Well, it seems that across the board, there was “more extended output in polite contexts.”¹⁰ This doesn’t mean that the outputs are necessarily of higher quality, but perhaps there is more of a chance of something useful being contained in the response.

However, according to this study, over-the-top politeness (or what I like to call being an AI-pleaser) can actually confuse ChatGPT more and weaken responses.¹⁰ The same sentiment rings true in human conversations—when our friends bombard questions with flowery flattery, sometimes it’s harder to decipher what they actually want.

However, there are numerous other experiments out there that suggest going above and beyond can help deliver results. For instance, emotional appeals at the end of your requests—like, “This is very important to my career”—were seen to enhance performance by over 10%, according to one study.¹¹ There was even a viral trend claiming that telling ChatGPT to “take a deep breath” before answering your question can help enhance response quality, too.¹² 

But no matter how you try to positively encourage ChatGPT, keep in mind: just like when asking a fellow human for something, clarity is key—and therefore, moderate politeness comes out on top.

3. As always, cultural context matters!

Spoiler alert: politeness is a cultural construct that varies drastically depending on who we are and where we’re from. As a result, each language has evolved to have its own specific set of expressions and honorifics to communicate our manners to others.¹³ With this in mind, it’s not all too surprising that in the study, the impact of politeness on LLMs varied depending on whether the language was English, Chinese, or Japanese. This pretty much remained true, regardless of the task.

Not only does this help us confirm that ChatGPT reflects the cultural context of the data it is trained on, but it is also a friendly reminder that research on LLMs should reflect the diversity of its human users. By expanding these studies beyond WEIRD contexts, we can optimize these platforms to be suitable for users across the globe—and maybe even uncover hidden insights about how politeness varies across populations.

There’s a good chance you’re wondering: why does politeness charm chatbots?  If this was a course called Prompt Engineering 101 (in which I would, by no means, be qualified to be the instructor), there are a number of fancy reasons I could give you, like pleasantries acting as signage to linguistically correlative responses.¹⁴ But the question we should be asking ourselves instead is, why does this all even matter in the first place?

The answer is simple: we humans are both the inputs and the outputs of this algorithm. 

Let me explain. Artificial intelligence doesn’t just automatically know how to be polite. It learns from us users, continuously refining its response with every interaction. But this relationship is by no means one-sided. Our manners are also influenced—especially as a growing proportion of our daily conversations are with chatbots rather than humans. 

Do you see how this feedback loop circles back around? By saying “please” and “thank you” to ChatGPT, the real output isn’t when it learns to be polite—it’s when it encourages other users to be polite, too. In other words, by training the algorithm, we are inadvertently training each other (thanks to the power of priming).¹⁵ And even if the impact doesn’t ripple all the way through, at the end of the day, you can rest assured that polite interactions with ChatGPT help you to train yourself. A little extra etiquette practice never hurts and could even go a long way in reminding us to be kinder in our conversations away from the computer. That way, we can keep ourselves from barking orders at each other the way we might spam a chatbox.

In the end, ChatGPT isn’t just our favorite coworker—it might be the secret to making or breaking company culture. If you kindly approach ChatGPT with clear questions as if it is a fellow employee, it will quickly pick up on these mannerisms and help to spread the word. But if you approach it in a sour mood… well, that negativity won’t be contained to your keyboard. And remember: this “office” isn’t just within your own walls but a global workforce more interconnected with this technology than ever before.

So next time you turn to your trusty colleague to ask a simple question, think twice about how you phrase it. The impact might be bigger than you think.