[bogtog]

If I'm understanding this right, the author gave ChatGPT-4o 2000 reviews and asks it "Are you 90% sure that this is a ghost job". Then, the author used those as labeled examples, trained a BERT model to predict the ChatGPT decision, and then applied the BERT model to the rest of the dataset. I guess this is cool, but if the goal is to pinpoint some percentage of ghost jobs overall I'm very skeptical

(it's a bit disappointing that 200 comments into this thread there was only a single mention of either "BERT" or "ChatGPT" per ctrl-f)

[pas]

the naive keyword search shows 1.6%, the GPT-then-BERT shows 21%, there's a 50% estimate from the literature

confounders abound, of course, but the proposed mechanism and other aspects all make sense