> It also seems like we cannot detect whether a piece of text was generated with LLM Based on reading those same reports, I think you can totally can detect it

> It also seems like we cannot detect whether a piece of text was generated with LLM Based on reading those same reports, I think you can totally can detect it, and Daniel also thinks that -- or at least, you can tell when it's very obvious and the user has pretty much just pasted what they got from the LLM into the submit box. Sneaky humans, trying to disguise their sources by removing the obvious tells, make it harder. The curl staff assume good faith and let the submitter explain themselves. Maybe the submitter has a reason for using it -- the submitter may be honest or dishonest as they wish. I like that the curl staff ask submitters to admit up-front if any AI was used, so they can discriminate between people with a legitimate use case (e.g. people who don't speak English but can find valid exploits and want to use machine translation for their writeup), versus others (e.g. people who think generalised LLMs can do security analysis). But even so, the central point of this blog post is that the bad humans waste their time, they can't get that time back, and even directly banning them does not have much of an effect on the bad humans, or the next batch of bad humans.