I actually tested this at one point with GPT-3.5 just by finding spam and non spam comments on a series of Mr Beast videos and, yeah, it was pretty great at it.

I actually tested this at one point with GPT-3.5 just by finding spam and non spam comments on a series of Mr Beast videos and, yeah, it was pretty great at it. Even ones that I wasn't 100% sure about it echoed that but would lean one way or another. I asked for outputs like Confidence score|Spam/Not Spam|Explanation and never saw it mark a comment that I'd consider genuine as spam and vice versa. Obviously this has a selection bias because I had to choose the inputs but there were some that said things like "I ate a ghost pepper on my channel" and stuff that were clearly spam but, to someone not aware that kind of thing is trying to bait you into looking at their channel, it'd appear as possibly just genuine. Heck it may have been typed by a human who owns the channel but is still spam. GPT got it. I tested this after the video came out a while back from one of the larger channels pleading for Google/YouTube to do something about all the spam comments and the general consensus seemed to be there was "just nothing they could do". Testing this lead me to believe they just don't want to do anything because if it's simple enough that some rando in his house can craft a prompt and get some examples to test in an hour or 2 then a multi-billion dollar company should be able to do SOMETHING.