> This should be titled "I don't know how to write good tests, so I consider them pointless"

Yes. An LLM certainly doesn't know this. Most of the input data to the LLM will not be "good tests" as these are quite rare in practice. The median test in the input data is not good, so the output of a LLM generating more tests won't be either. Most human oversight won't even recognise this either.

> only change when you change that specific behaviour.

Correct, as Kent Beck observed: "Tests should be coupled to the behaviour of code and decoupled from the structure of code"

bad tests "need constant updates" because they reverse this.