I think you could make the GPT's more undetectable by a computer using GAN techniques but that might well make the thing more detectable by a person. The GPT-style is something like rhetorical devices run together to make various claims unfold smoothly. An alternative approach could make the things inconsistency more obvious to a person.