Would it be possible to use GPT model infrastructure to compute the likelihood that a given text was created by that GPT model?