A baseline model trained on tf-idf vector representation of text and a logistic regression to detect WebText articles (online web pages) from text generated using GPT-2 models.

Study of different sizes of GPT-2 models indicated that models having a large number of parameters generated text somewhat similar to humans.

Text generated from Classifiers built with nucleus sampling are hard to detect.

Fine tuning GPT-2 specifically to amazon product reviews generated texts that are human generated.