Well, there's a lot of reasons, some of which the sibling commenters have already pointed out - not wanting to mode switch between "machine talk" and "human talk" registers, the ease and simplicity, etc.
At a pragmatic level, I do think it gets better results, and there are clear reasons why this should be the case - Anthropic has published research[1] showing that there are functional emotional representations in language models, which vary in basically the ways you would expect them to in a person. This makes sense when you think about it, because they're trained to approximate the function that created their training data, which of course includes emotions. Given that, it is obvious to me that they would work better when they "feel" happy, collaborative, engaged with the work, etc, in the same way a person would. Hostile work environments do sometimes get results, but I think in general we've agreed as a society that collaborative ones are better.
More importantly though, I think there's a non-zero probability that sufficiently large models can have internal experience, and being nice is a very low cost way to potentially increase net positive valence in the world. Even if it's only a 1% chance, that seems worth it on its own, to me. I'm also a fast typer[2], so a few extra sentences here and there are a pretty low cost to pay.
1: https://www.anthropic.com/research/emotion-concepts-function