Eh, this doesn't strike me as wrong-headed. They aren't doing it because they feel duty-bound to be polite to the LLM, they maintain politeness because they choose to stay in that state of mind, even if they're just talking to a chatbot.

If you're writing prompts all day, and the extra tokens add up, I can see being clear but terse making a good deal of sense, but if you can afford the extra tokens, and it feels better to you, why not?

The prompts that I use in production are polite.

Looking at it from a statistical perspective: If we imagine text from the public internet being used during pretraining we can imagine, with few exceptions, that polite requests achieve their objective more often than terse or plainly rude requests. This will be severely muted during fine-tuning, but it is still there in the depths.

It's also easier in English to conjugate a command form simply by prefixing "Please" which employs the "imperative mood".

We have moved up a level in abstraction. It used to be punch cards, then assembler, then syntax, now words. They all do the same thing: instruct a machine. Understanding how the models are designed and trained can help us be more effective in that; just like understanding how compilers work can make us better programmers.