This is ridiculous. I doubt this would work with a general AI, but it surely cannot work with LLMs who understand exactly nothing about human behaviour.
I agree that the approach shouldn’t be done unsupervised but I can imagine it being useful to gain valuable insights for improving the product before real users even interact with it.
But this is completely worthless or even misleading. There is zero value in this kind of "feedback". It will produce nonsense which sounds believable. You need to talk to real users of your software.