Is there a longer-form paper on this yet? TPR (P(T|AI)) and FPR (P(T|H)) are useful, but what I really want is the probability that a piece flagged as AI-generated is indeed AI-generated, i.e. P(AI|T). Per Bayes rule I'm missing P(AI), the portion of the challenger set that was produced by AI.
If we assume the challenger set is evenly split 50-50, that means
So slightly better than a 1/3 chance of the flagged text actually being AI-generated.
They say the web-app uses a confidence threshold to keep the FPR low, so maybe these numbers get a bit better, but very far from being used as a detector anywhere it matters.
Precision is impossible to calculate without knowing P(AI), which is use-case specific.
Source: Spent 10 years trying to explain this to government people who insisted that someone tell them Precision based purely on the classifier accuracy without considering usage.
If we assume the challenger set is evenly split 50-50, that means
So slightly better than a 1/3 chance of the flagged text actually being AI-generated.They say the web-app uses a confidence threshold to keep the FPR low, so maybe these numbers get a bit better, but very far from being used as a detector anywhere it matters.