Computer ratings vs real people ratings?

It is because the rating of a weak Bot is very unprecise. It is very hard to make an engine (Bot) play weak like a human. So they mix good and very bad moves, and the quality of their play is more mixed than of humans. Most times they will play stronger or weaker that their raring.

I think that's some of the story, but not sure it's all the story. Even bots that use machine learning to play more human-like moves like Maia I've found play considerably worse than human players rated similarly. I can usually beat Maia-1100 and I do think it makes more understandable, human-feeling plays and mistakes than normal bots which seem to play really well and then suddenly make the most wild blunders.
I think chess education has changed hugely in the last 5 years as a result of the massive popularity boost, and bot rating/behaviour is still catching up with the experience of what being a newer player is like now. New players have to grow up fast these days because there are so many more good, free resources to help you get a good start than there were before. I think this is also part of the story.

Another part of the story: cc is a company that lives off our time/attention. And people get happy when they beat a higher rated bot. And happiness probably means more time/attention

Humans can sometimes use unconventional moves or create complications that force you to think on your feet, whereas bots might stick to a narrow set of responses.

Bots tend to have one-note playing styles (such as Nelson being too gung-ho with the queen) + they are designed to randomly blunder. If you can find the way to counter the way they play (for example, the most obvious way to beat Nelson is to trade queens ASAP) they generally become pretty easy. Can't say that about (most) people.