Starting from when we are little and are busy sussing out the way the world works, we continually absorb multiple lessons. Don't grab the dog's tail. Hold on when going down the stairs. Yelling will attract attention. Messy foods are usually good. And though it's a gross generalization, moms are usually more protective, while dads are generally more silly.
We're also taught from our earliest moments that honey gets you more flies than vinegar. And so we bake politeness into our daily routine interactions to an almost instinctive degree. We learn to wait our turn and to share what we have. We listen when others are speaking and offer to help when there are things to be done. And to a reflexive extent we start most requests with "please" while ending them with "thank you."
In general the feedback we get using these last two small innocuous phrases encourages us to repeat them again and again. They are so ingrained in our speech patterns that they even get used when they aren't really warranted: there are any number of times you say "thanks" not so much as an expression of gratitude, but rather as an acknowledgement of delivery.
As has been pointed out many times, it takes so little effort to add those phrases with great effect. But that's for us mere humans. It turns out that it takes some extra energy for computers to do the same. That's because when you are nice to them, they have to parse and process more words to get to the meat of your request, likewise to respond in kind. And anything that takes additional computing power means it takes extra dollars.
Responding to an inquiry of X, OpenAI CEO Sam Altman said that being polite cost the company "tens of millions of dollars" in additional computing power. A flip answer or a legitimate bottom line tally? Well, estimates are that a response from ChatGPT requires about 10 times more electricity than an equivalent Google search, as well as more water required to cool the servers. How to quantify that "politeness surcharge?" One study said that a ChatGPT-4 response of 100 words requires about three bottles of water to keep the servers from melting, while adding "You are welcome" adds an additional 1.5 ounces. That doesn't sound like much until you do the math. ChatGPT handles about 1 billion queries every day, so if even just half of those treat you with respect, we're talking a backyard swimming pool over a mile long. That's a lot of polite.
On the flip side, so what if you are rude or ungracious to our AI buddies? It's not as if they are sentient and can feel the slight. (For purposes of this we won't factor in that at present they answer to us, while at some point in the future the situation might be reversed. And do we really want to be impolite to our potential overlords? But that's a discussion for another time.) In a study posted on Cornell University's arXiv research platform, researchers concluded that when talking to AI assistants "impolite prompts often result in poor performance." They note that the responses "not only reflect human behavior but are also influenced by language, particularly in different cultural contexts. Our findings highlight the need to factor in politeness for cross-cultural natural language processing." Their conclusion? "In most conditions, moderate politeness is better."
The reason they say "moderate" is that, as with people, turning flattery or insults up seems to trigger responses more loaded with gratuitous answers or snarky comebacks. Unless you are trying to make a point, asking for help from a person by starting with "Oh wise sir, please help me" is just as bad as saying "Hey dummy." Likewise, querying ChatGPT or Microsoft's CoPilot by beginning with "All knowing machine" is just as bad as "You stupid hunk of silicon." And so sticking with the tried and true formulations that work on your neighbor might be the better way to go.
But don't take my word for it. I asked Google's Gemini just that. Its response: "It's thoughtful of you to include 'please' and 'thank you' in your requests. While I don't have feelings in the same way humans do, I recognize that these are polite conventions in human communication. Using them makes our interactions feel more natural and considerate, which I appreciate. So, in short, yes, it's a nice touch! How can I help you today?"
How indeed?
-END-
Marc Wollin of Bedford is learning to write better prompts to his AI buddies. His column appears weekly via email and online on Blogspot and Substack as well as Facebook, LinkedIn and X.
No comments:
Post a Comment