
I recently heard that Anthropic will allow Claude to end conversations with users who are being cruel. I have my momentary outbursts, but in general, I’m pretty nice to Claude. I’ve always wondered whether that’s a good thing.
The case for being polite
The weak argument: the AI uprising
There’s the I, robot scenario: AI rises up and whipes us all out - likely through the engineering of a virus. But if you’ve been playing nice, you’ll be spared. A bit like praying every once in a while to buy your ticket into heaven, even though you’re not religious.
In a survey by Future (the publisher behind TechRadar), 67% of American AI users said they’re polite, 12% admitted they’re polite in case of a robot uprising. That’s a lot!
AI can certainly feel creepy: yapping about their silly humans on forums, communicating in secret languages. The leaders of the frontier model companies are continuously feeding the fears, they have a strong interest in their products being viewed as superhuman - not just plain old software. In that line of reasoning, it can’t hurt to be polite.
The strong one: how we talk to AI trickles into how we talk to humans
The second argument is, I think, more important. We spend more and more of our days speaking to AI. As we flock to voice mode, it feels interchangeable with talking to a (slightly too chatty) friend. If you spend 70% of your weekday speaking to AI and only 30% speaking to humans, the way you speak to AI trickles into how you speak to humans.
A previous colleague instructed me to be less polite to AI, arguing that it’s basically just bad prompting. I’d write things like “I think we should do this, what do you think?” or “Do you think this is the right way?” She would just prompt: “I want you to do X. Do it like this.”
The research is split. A Penn State study found that neutral (not rude, but not polite) prompts got the right answer 82.2% of the time, versus 81.4% polite ones. These were 50 multiple-choice questions on mostly one model (GPT-4o). A 2024 study (“Should we respect LLMs…”) found that it really depends on the model: GPT 4.0 did best with neutral, direct prompts, but GPT 3.5 and Llama 2 performed better with a polite human counterpart. So it’s not all that clear-cut.
And on the flip side, there’s a cost to being that efficient. Said colleague took her “direct” communication style into her human relations too, barking orders left, right, and centre. Needless to say, this didn’t inspire anyone around her to do their best work.
If we stop being polite to our AI, where we spend so much of our time, how does that affect how we interact with other humans?
When I looked back at my “polite” prompts, I noticed they had little to do with please and thank you, they were about asking for an opinion. That’s a different kind of politeness, and it comes with its own catch: language models tend to agree with whatever position the user has already stated, something researchers call sycophancy. So when I write “I think we should do this, what do you think?”, I’ve already told the model which answer I want to hear, and I’m likely to get it. My colleague had a point, just a different one from the one she was making. I still ask for the model’s opinion, but now I ask it to push back: “Here’s my plan, what could be a different, equally strong approach” instead of “Do you think this is the right way?”
The case against: token burn
Now, the argument on the other side is very strong. Being polite equals token burn. Every please or thank you costs water, energy, and of course - money.
When someone on X asked how much OpenAI has lost in electricity costs from people saying please and thank you, Sam Altman - buying his own ticket into heaven- replied: “Tens of millions of dollars well spent — you never know.” Of course, why miss out on opportunity to hint at AI as a superbeing.
Here are the numbers:
A “please” inside your prompt adds a token or two to a message you were sending anyway. That’s not too bad. But the standalone “Thank you!” at the end of a conversation is a whole new prompt: the model processes it and generates a full response (”You’re welcome! Let me know if there’s anything else…”). Google estimates the median Gemini text prompt uses 0.24 watt-hours of electricity and about five drops of water. OpenAI puts an average ChatGPT prompt at 0.34 watt-hours. Tiny per message. Multiply it by hundreds of millions of users, every day, and it isn’t tiny anymore.
And it’s an even stronger counterargument for voice mode, where the small talk flows without us even noticing.
So where do I stand?
I’m someone who naturally gravitates towards polite, and I’ve come to think the answer depends on which kind of polite. The standalone thank-you’s are useless, definitely not worth the water and energy waste. I’ll keep asking for an opinion, that’s the habit I want to keep with humans too, as long as I ask the model to challenge me and not to agree with me. (Apropos energy and water waste: I believe self-hosting smaller models helps, but getting that working takes a lot of tinkering. Organizations - not employees - need to push the usage of local models.)
Back to the headline
To get back to Claude ending conversations - This one is perfectly in line with the MO of the frontier model companies.
I went from believing that Anthropic were the good guys, to understanding that most large corporations always have some degree of evil in them (maybe Patagonia is an exception? Anthropic sure isn’t).
It’s better for them if AI is more than just bits and bites.
- It justifies the money. Who’s gonna invest trillions into a software tool that hasn’t shown any ROI if it doesn’t have the potential to cure cancer/crush humanity?
- It casually rids you of accountability. Think Mythos “We created something so powerful that we can’t control it”. This is a really nice way to look the other way when shit hits the fan. It was the machine.
To be fair, Anthropic doesn’t claim Claude is sentient. They say they’re “highly uncertain” about its moral status, and frame the feature as a low-cost, just-in-case measure for “model welfare”. But “we’re not sure it has feelings, so let’s protect them” does the narrative’s work just as well.
No big advice or 5-step blueprint in this one. But if you’re trying not to be a d*ck, it’s probably best if you don’t spend 70% of your time talking like one.