I replied to a comment with a structured informal proof on HN.
replyA well established user here was adamant that I used AI because apparently humans never ever wrote proofs.
This was a while ago. Any well crafted human output is now being dismissively cast as AI if the reader is challenged by the output intellectually/politically.
It's the irl version of being accused of hacking in Counterstrike.
replyI’ve been accused of being a bot myself. I’ve started leaving some incorrect autocorrects in, hopping it would stave off people for whom it seems making accusations of being a bot is soothing or a coping mechanism when they’re confronted with information that does not fit into their world view or perception of self.
replyIn fairness, the original crop of detectors were based on perplexity scores and were entirely useless (famously, the declaration of independence often came back as 100% AI generated).
replyI'm not convinced we'll ever have full-proof detectors, and certainly the false-positive rate will make them irresponsible for accusations of intellectual/academic fraud, I do think that LLMs are easy for folks to sniff out on average so I imagine it's possible to detect many instances.
Pangram's detector is anecdotally very accurate in my tests. This detector appears to be fine-tuned on a very small dataset (200 papers per subject), and suspect the problem might be in part that.
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.
replyIt 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.
replyThere's a table in the article breaking down the rate by field.
reply