I think even this is too generous. These “detectors” are claiming to do something which is impossible from a technical sense. You can’t calculate perplexity without knowing the prior probability distribution, and you can’t know that without knowing which model was used. They’re playing a dangerous game where the consequences for being wrong (which will be borne by innocent 3rd parties) can be extremely dire.
I think even this is too generous. These “detectors” are claiming to do something which is impossible from a technical sense. You can’t calculate perplexity without knowing the prior probability distribution, and you can’t know that without knowing which model was used. They’re playing a dangerous game where the consequences for being wrong (which will be borne by innocent 3rd parties) can be extremely dire.