Re: AI Wars
Human judges can't reliably classify prose samples as human- or machine-generated. Models trained on corpora of machine- and human-generated prose will almost certainly overfit to particular generators. Adversarial tweaks to output will make it easy to defeat detection models, if the output of those models is available as an oracle – and it has to be, if they're going to be useful for the reading public.
Biting the hand that feeds IT
