Can it be that now they are getting optimized against benchmarks that are valuing logics, rather than human appreciation? (I am not an expert at all, just an idea)