But who is reading those tests? Tests to code LOC was 0.15:1 because humans only tested that 0.15. Now you are assuming that the test the machine wrote actually covers the features that should be tested.

Depending on machine-generated tests that nobody reads isn't better than having poorer coverage. The only thing that is guaranteed is that that mass of code returns "true" with the current implementation and "false" if the implementation changes in some ways. Unless people actually read the LLM-generated code, it's impossible to know if it tests anything that actually matters or just tests the same code path that would never be actually used in production in 5 different ways.