4 ms·
> Why should it matter if an otherwise good answer came from an LLM? Because good as LLMs may be at writing good answers, they're even better at writing good-l
by vintermann 3y ago
> Why should it matter if an otherwise good answer came from an LLM?
Because good as LLMs may be at writing good answers, they're even better at writing good-looking answers. That's pretty much what they're been trained to do, and actually good answers only appear as a fortunate side effect.
Obviously, this makes moderators' jobs a lot harder. I can understand that they get frustrated. It also allows automated climbing in SO's gamified status hierarchy, which makes it even more dysfunctional and threatens those currently on top of it.
So they want to be able to ban posts for looking automated. It's just that whether you decide manually, or with an automated tool of your own, you're going to have a huge false positive rate and quite possibly unfortunate biases in it as well. (Very recently there was a paper showing how automated LM detectors were great at sussing out non-native English speakers and deciding they were bots).
- wilg 3y agoI agree with the rest of your comment, but it's framed as a retort but doesn't really work as one. The person your responding to already said "otherwise good answer" which excludes a wrong but "good-looking" one.
- JenrHywy 3y agoModerators generally aren't checking for correctness - who would have that sort of time?
- komadori 3y agoI would expect the moderators to have some subject knowledge and often it doesn't take much to spot that an answer is blatantly fantastical as LLMs sometimes are. The finer points of some API might need actual checking, but a lot of things you'll know on sight.
- vintermann 3y agoPoint is, you don't know which are wrong unless you do the work. The moderator has to do the actual work, but the LM (or rather, the user running an LM bot on his account) gets the credit. They asked why it is a problem, and I tried to explain why it is in practice.