NB: suspishously good grammer is now becuming an AI sign in itself
I moderate for a subreddit where, despite our best efforts to purge them with holy fire, AI spambots are rife (trying to act inconspicuous and get some karma). In addition to the “suspiciously good grammar and bullet points and em dashes” type, there’s conversely also a strain of them that use a gimmicky level of slang.
Where most of the tone of the subreddit is explanatory/informative, and most people are using standard spelling/grammar, they drop in with “ngl bro honestly the vibes are wild fr”
Oh wow! That is good to know haha (counter/counter/counter culture I guess) - do you have rules to spot them or is it just they look so out of place its obvious to anyone who moderates?
We make efforts with the automod, but it’s a fairly rudimentary tool. Mostly a keyword/regex matcher.
So we can have it automatically look for some of the common repeated phrases (relatively few of our human users say “That’s a great question”). If it’s a weak signal maybe it only highlights them for mod review rather than fully automatically removing them.
But otherwise it relies on the human eye to spot them being weird. Especially when the comments are too short for AI detectors to get a good read on it (deciding what’s suspicious enough to run through a detector is itself done by eye).
I moderate for a subreddit where, despite our best efforts to purge them with holy fire, AI spambots are rife (trying to act inconspicuous and get some karma). In addition to the “suspiciously good grammar and bullet points and em dashes” type, there’s conversely also a strain of them that use a gimmicky level of slang.
Where most of the tone of the subreddit is explanatory/informative, and most people are using standard spelling/grammar, they drop in with “ngl bro honestly the vibes are wild fr”
Oh wow! That is good to know haha (counter/counter/counter culture I guess) - do you have rules to spot them or is it just they look so out of place its obvious to anyone who moderates?
We make efforts with the automod, but it’s a fairly rudimentary tool. Mostly a keyword/regex matcher.
So we can have it automatically look for some of the common repeated phrases (relatively few of our human users say “That’s a great question”). If it’s a weak signal maybe it only highlights them for mod review rather than fully automatically removing them.
But otherwise it relies on the human eye to spot them being weird. Especially when the comments are too short for AI detectors to get a good read on it (deciding what’s suspicious enough to run through a detector is itself done by eye).