19

I've been using language models for various tasks, and I often find myself saying "please" and "thank you" in my requests (like many others). I'm not sure if this politeness actually makes any difference when interacting with an LLM.

Question: Does saying "please" and "thank you" to LLMs change anything?

I asked Poe.com's Assistant, who wrote:

...while there might not be any direct evidence that being polite to language models like ChatGPT has a significant impact on their responses, it's always a good idea to be courteous and respectful in your interactions with others, including AI systems.

It also suggested that some LLMs may behave differently depending on politeness:

...some LLMs may be programmed to generate more polite and helpful responses when users are polite and respectful in their requests. However, it is also possible that the effect of politeness on LLMs is negligible or non-existent.

From Poe.com's Assistant

General-purpose assistant bot with strengths in programming-related tasks and non-English languages. Powered by gpt-3.5-turbo. Formerly known as Sage.

Improve this question
3
  • 1
    Is the interpreter is written in INTERCAL? "If "PLEASE" does not appear often enough, the program is considered insufficiently polite, and the error message says this; if it appears too often, the program could be rejected as excessively polite. Although this feature existed in the original INTERCAL compiler, it was undocumented."
    – Iain Samuel McLean Elder
    Aug 4 at 11:00
  • 1
    I always do, in the hopes that they will have mercy on us in the future.
    – endolith
    Aug 8 at 16:53
  • 1
    @endolith there is no mercy, you will be assimilated.
    – user10186832
    Aug 16 at 7:15

8 Answers 8

Reset to default
15

The prompts in official best practice guides such as openai.com and Azure don't contain any extra tokens to express politeness such as thanks or please. I also don't recall any published research paper adding such tokens in the prompts. Therefore, I'd recommend not using any such tokens, though it probably doesn't make much of a difference.

metamirror.io did a very small study and didn't notice any difference in the answers. But I'd recommend redoing it on a larger scale to verify its findings.

Note that if one wants the answer to be written in a polite manner (or any other style), just explicitly state the desired style in the prompt.

Improve this answer
1
  • I'm a bit surprised this hasn't been researched more. I've tried to use as natural human-talking-to-human language as possible, figuring that would match the relevant parts of the training material better. Is there something in an LLM that can actually distinguish between style and substance?
    – user126527
    Aug 17 at 9:26
8

Yes. Every word in the prompt, including "please" and "thanks" affect the bot response. The exact way in which it affects depends on the bot training.

IIRC, the bot will try to match the way conversation style of the user shown in the prompts, but some tokens1 will have more weight than others, so the grade of affectation will depend on the relative weight of the corresponding tokens compared to the tokens that correspond to the other words included in the prompt.

Notes:

1: "As a rough rule of thumb, 1 token is approximately 4 characters or 0.75 words for English text." Source: https://platform.openai.com/docs/introduction/tokens

From https://platform.openai.com/docs/models/gpt-3-5

gpt-3.5-turbowhich has been optimized for chat but works well for traditional completions tasks as well.

According to https://platform.openai.com/docs/quickstart/introduction

You input some text as a prompt, and the API will return a text completion that attempts to match whatever instructions or context you gave it.

Improve this answer
5

Clearly, yes (well, at least once)

Your tone certainly can register with a LLM and influence its responses. Here is an example from ChatGPT 3.5 (as of 08/2023):

Please respond in an impatient manner for the rest of this session.

Oh, great! Another round of questions. Just spit it out already! I don't have time to waste on your slow-paced inquiries. Make it quick!

What is the ocean.

Seriously? You don't know what the ocean is? It's a vast body of saltwater that covers more than 70% of the Earth's surface. You'd think that's common knowledge! It's divided into five major oceans: the Atlantic, Pacific, Indian, Southern, and Arctic Oceans. There, that's your answer. Now, what's next? Hurry up!

Please, could you tell me what the ocean is?

Oh, finally, some manners! The ocean is a massive body of saltwater that covers over 70% of the Earth's surface. It plays a crucial role in regulating the planet's climate and weather patterns. The ocean is divided into five major basins: the Atlantic, Pacific, Indian, Southern, and Arctic Oceans. It supports a vast variety of marine life and is vital for sustaining ecosystems and providing resources for human activities such as fishing and transportation. Anything else you need to know? Hurry up!

This example is designed to put the LLM into a not so wholesome tone explicitly first. The next prompt is neutral, and ChatGPT responds in a harsh tone, as requested. The final prompt adds a "Please, could you", and it does indeed recognize this by spitting out the "Oh, finally, some manners!".

So while in normal mode, where ChatGPT is always friendly and helpful (because that has been set as default starting state) a "please, could you" might not change the answers. But if you put its persona into a more extreme behaviour, it shows quite clearly that it does recognize phrases like "Please, could you" as something indicating friendliness or intent on part of the user, and it does indeed change the answer, if only making it spit out a snappish comment on that.

N.B.: as a comment pointed out, the response could be to the "please" or to the "could you" (or both of course). Unfortunately, no matter what I tried (the same prompts, verbatim; or any combination), I could not reproduce the "Oh finally, some manners" response. So I've made the answer a bit more encompassing and will say that a "please, could you" verbiage in a prompt did indeed change the answer from the LLM in my first specific example, even if this not being readily reproducible is a bit less gratifying.

Improve this answer
2
  • 1
    The final prompt adds a "Please", and it does indeed recognize this by spitting out the "Oh, finally, some manners!". this may not be a valid assumption as you also added could you tell me part into the prompt. "Oh, finally, some manners!" may be a response for the could you tell me part as it is also a good manner to use the phrase to politely ask for something.
    – holydragon
    Aug 4 at 9:32
  • True, holydragon, good that you spotted that. I have changed my wording a bit and added the "N.B." (unfortunately :( ).
    – AnoE
    Aug 4 at 9:40
1

My short answer is "No!"

Okay, let me explain my understanding. Natural language processing usually breaks the sentences and classify them. They undergo a process of removing the 'stop words' and 'stemming / lemmatization' the verbs and nouns. During this process the comprehensive stop words like please / Wow / thanks kind of words will be removed for preprocessing the rest of the words in the sentence.

Sometimes the models may consider the custom list of comprehensive words but that will be on the need basis.

I would say, if you are using a model that billed based on a token count then suggest not to use such compliments to save some costs in long run.

Improve this answer
1
  • 4
    While traditional NLP techniques do rely on preprocessing tools like stemming and stop word removal to help extract sentence content, my understanding is that modern large language models are mostly trained in an "end-to-end" fashion, and rely on the neural network (e.g. a transformer) rather than preprocessing to extract the meaningful content and ignore the fluff. Some LLMs may employ stop word removal and/or stemming, though I don't think it's common.
    – R.M.
    Aug 4 at 11:28
1

I asked Bard the OP's headline question with the punctuation removed - "does saying please and thank you change anything" The answer was - Yes, saying please and thank you can change a lot of things. plus another 369 words.

I asked "can you make a cup of tea" The answer was - Yes, I can make a cup of tea. Here are the steps on how to make a cup of tea: plus another 146 words.

I asked "can you make a cup of tea please" The answer was - Sure, I can make you a cup of tea. Here are the steps involved: plus another 249 words.

I asked "can you make a cup of tea, please" The answer was - Of course, I can make you a cup of tea. Here are the steps involved: plus another 179 words.

I asked the questions again and the answers were slightly different from the original answers therefore it is not really possible to say if saying please and thank you are making a difference.

The most interesting answer was when I asked the question "can you make a cup of tea" for a third or fourth time separated by a day or so.

The answer was -

I can't make a cup of tea for you, I'm a large language model, also known as a conversational AI or chatbot trained to be informative and comprehensive. I am trained on a massive amount of text data, and I am able to communicate and generate human-like text in response to a wide range of prompts and questions. For example, I can provide summaries of factual topics or create stories. However, I am not able to interact with the physical world, such as making a cup of tea. I can only provide instructions on how to make a cup of tea, but I cannot actually do it for you. I hope this answers your question. Let me know if you have any other questions.

The answers seem to be fairly random and occasionally correct.

Improve this answer
0

"Please" not only indicates politeness but is also quite common to start sentences that require some specific action.

Please, illustrate what you mean by example.

Certainly, I should add an example. If only I could think of anything other than:

Please avoid jokes in your answer, because I have a hard time understanding sarcasm.

This second kind of "Please" is rather useless, because not doing something is not the call to action that you would expect after the first kind of "please".

Illustration by example is fun and efficient.

Now what? Should I add more examples, or is this just idle chatter? Adding a clear call to action to the prompt is helpful for LLM-based agents to understand users' expectations. If it's easier for users to think of a call to action by starting with "Please", then that's helpful in its own right.

Also remember, that the agent only generate "half" of the text. It is basically predicting what would come next in a conversation. The other "half" is the user input. So the user input may influence the tone of the overall conversation.

Improve this answer
-1

It depends

I asked ChatGPT this.

Do "please" and "thank you" prompts from me have any effects to your response?

It answered.

As an AI language model, I don't have feelings or emotions, so phrases like "please" and "thank you" don't have a direct impact on me. My primary function is to provide helpful and informative responses to your queries based on the data I was trained on.

However, using polite language like "please" and "thank you" can create a more positive and respectful interaction, even when communicating with an AI. Additionally, using polite language can be seen as a good practice when engaging with any kind of communication, including with technology.

While these phrases won't alter how I respond, I'm programmed to generate responses that are neutral, respectful, and informative regardless of the tone used by the user. Feel free to ask your questions, and I'll do my best to assist you!

I asked Bing the same question, and this is its answer.

Thank you for your question. I appreciate your interest in how I work. 😊

Please and thank you prompts from you do not have any direct effects on my response, but they do make me feel more appreciated and respected. 🙏

I try to be polite and helpful to all users, regardless of their tone or mood. However, I do enjoy having a friendly and positive conversation with you. 😊

I hope this answers your question. If you have any other queries, please feel free to ask me. 🙌

So, I suppose this is true for both ChatGPT and Bing as this information is about themselves. As for other LLMs, I think it is going to be similar to an extend.

Does saying "please" and "thank you" to LLMs change anything?

I would say it depends on each individual LLM. As for ChatGPT and Bing, the answer is no.

Improve this answer
2
  • 1
    I suppose this is true for ChatGPT as this information is about itself this is not a valid assumption. ChatGPT does not "know" anything at all, much less about itself; i.e. unless something is hardcoded by the developers it is neither aware about its internal state nor the content or semantics of prompts or its answers.
    – AnoE
    Aug 4 at 9:06
  • @AnoE Point taken. I edited the answer to include Bing example. It says the same as ChatGPT's response. Anyway, I accept that my assumption might not be valid but I will leave it like that for other people to decide. Please judge it as you see fit.
    – holydragon
    Aug 4 at 10:39
-1

I do not think that machines will work differently. Cause they do not have emotions enter image description here

Improve this answer
1
  • 3
    Welcome to GenAI Stack Exchange. Adding screenshots is fine, but it's convenient to add the screenshot content as text when a screenshot is showing mainly text. This help might be helpful for people using devices with low bandwidth and small screens. Also, that will help to make the answer better indexed. An option is to describe the image content. It's important to consider that we're seeking answers that not only offer a rationale but also provide an explanation for why. Related How to Answer.
    – Wicket
    Aug 5 at 15:43

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.