Rafael Harth comments on Language models can generate superior text compared to their input