In essence detecting which one is fake is a common way how you train the generator, tweaking the generating process to "fix" any detectable flaw; and you train it until (as far as your system is concerned) the generated texts are indistinguishable from the real ones. A better system might distinguish them, but that better system can be relatively trivially adapted to generate better texts which it won't be able to distinguish from real ones.