How can they know anything they train on is valuable?

At the end of the day it doesn't matter. You got the wrong answer and didn't complain, so why would they care?

In general, I think any human generated content in pre-2022 is valuable because someone did some kind of validation (think about stack overflow answer with user confirming that a specific answer fixed their problem).

If they start to feed the next model with LLM generated crap, the overall performance will drop and instead of getting a useful answer 1 of 5 it will be 1 of 10(?) and probably a lot of us will cancel the subscription ... so in the end I think it matters.