[yorwba]

They use a "threshold calibrated so pre-ChatGPT papers flag at 0.4%", so these things do work most of the time. It also means that there are known false positives, so for any given paper, scoring above the threshold isn't irrefutable proof of AI usage. But for things like estimating the overall proportion of AI writing, you only need to be correct on average, so individual false positives don't matter much.

[jasonfarnon]

It would be interesting to see how that .4% varies across fields on arXiv. I imagine some areas influenced LLM writing much more than others and are more susceptible to false positives. Anyway it's a little strange to me that these detectors do so poorly (at least by what I hear on this site). There is as much labeled data as you would need to train on pre-LLM human vs LLM text. If their out of sample errors are as good as has been claimed, the only methodological flaw I can think of is that post-LLM human text and LLM text have been converging.

[yorwba]

There's a table in the article breaking down the rate by field.