43

Answers generated using ChatGPT are banned on many Stack Exchange sites. This is mainly because they aren't particularly accurate or useful, and are often misleading.

Personally, I've spotted a number of these; none of which were any good, and all of which have been deleted.

As I don't have the Deletion privilege, when I see a ChatGPT answer all I can do is to flag it (and downvote). But, none of the standard answer flags are a great match for this:

  • Spam
  • Rude or abusive
  • Not an answer
  • Very low quality

So, so far I've been using the custom flag:

  • In need of moderator intervention

... with what is basically the same custom message, something along the lines of:

This looks like an absolute rubbish ChatGPT "answer".

But, custom flags can't be handled automatically, and so take up moderators' time which seems unnecessary.


I request that a new flag be created for ChatGPT based answers, something along the lines:

  • ChatGPT
    This answer appears to be generated by ChatGPT (or similar AI) and provides inaccurate/misleading information.
9
  • How about mentioning that the AI content is uncredited? Some users make use of AI translation assistance after all.
    – W.O.
    Commented Apr 21, 2023 at 15:07
  • 21
    "But, with this flag I've had to type out the same custom message over and over." You shouldn't be doing that. Instead, you should be listing the reasons you think a post is generated via ChatGPT.
    – VLAZ
    Commented Apr 21, 2023 at 15:07
  • 5
    @VLAZ I don't understand why I'd need to provide a list of reasons, in my experience ChatGPT answers are self-evidently garbage. Can you post your comment as an answer, and provide a particular example of such a list? Commented Apr 21, 2023 at 20:56
  • 5
    If you've only dealt with "self-evident garbage", then you should know that not all generated content is that. A lot is very hard to pinpoint garbage which requires an SME to find inconsistencies in the explanation. And others still are correct but require expertise in spotting generated content to pinpoint. Some times the evidence is from a chain of answers, not just a single one. So, I don't think "self-evident garbage" is the norm when dealing with ChatGPT content.
    – VLAZ
    Commented Apr 21, 2023 at 21:07
  • 3
    @VLAZ My point is, that if 3 (or whatever number of) standard ChatGPT flags could get a bad answer removed (without needing detailed custom flag messages, or a moderator's attention), it'd be a less taxing system. Commented Apr 21, 2023 at 21:37
  • 2
    As an aside, the vast majority of AI generated content I've encountered is spam. The rest mainly falls into the low quality category (as answers). Easy flagging, but your mileage may vary. Only come across one AI question that wasn't spam, and that was lacking details and clarity, but with that frequency, it's easy to custom-flag there.
    – W.O.
    Commented Apr 23, 2023 at 20:33
  • related on MSO: Add flagging AI generated answers
    – starball
    Commented Nov 22, 2023 at 20:00
  • 2
    @VLAZ I've flagged hundreds of ChatGPT spam answers, and it's onerous to have to explain to mods that have probably seen thousands of such answers why it's ChatGPT. I know explaining why is the policy historically, but it makes about as much sense at this point as explaining why non-English answers should be deleted. Almost all of them stick out like a sore thumb, except for the occasional answer by a spammer that's probably GPT-generated, but the answerer got wise and decided to make it code-only.
    – ggorlen
    Commented Jan 27 at 16:05
  • I would really like to be able to flag answers as 'ChatGPT'
    – CPlus
    Commented Apr 23 at 5:20

2 Answers 2

20

My point is, that if 3 (or whatever number of) standard ChatGPT flags could get a bad answer removed (without needing detailed custom flag messages, or a moderator's attention), it'd be a less taxing system.

This seems to be the crux of your suggestion and unfortunately I think this is impossible (or a bad idea). Please continue to use "In need of moderator intervention" and give as much detail as possible as to why you think it's AI.

As to why the flag you describe won't work, the current process on most sites is as follows:

  1. An "In need of moderator intervention" flag comes in describing why a user thinks an answer is AI-generated. (Flags can be cast by any user with 15 rep.)
  2. A moderator investigates that answer and all of the user's other recent answers. This is where most of the work is, since so many users think they can pass off AI text as their own. Most moderators have chosen to keep the details of this process a secret so that troublemakers can't work around it.
  3. If the suspicions were correct, the answer is deleted.
  4. The moderator sends a moderator message. This will likely come with a suspension if this isn't the first infraction.

First of all, your suggestion prevents the user from giving any details about why the answer is AI generated (as in #1). These details are very important because they can considerably shorten the amount of time a moderator spends on step #2.

Step #2 is important. AI can be tricky to identify and it's easy to get things wrong if you only investigate superficially.

As for step #3, there are existing flags that delete answers directly, the "Spam" and "Rude or abusive" flags (known as red flags). These flags, however, are used on posts that are as obvious as slap in the face — there is no investigation to see if there could be a legitimate answer to a coding problem in a post that recommends you buy pills and supplements. Without moderator intervention, a post needs 6 red flags to be deleted but it's necessary for a group of users to coordinate in order to make sure spam and abusive posts are consistently deleted ASAP. The high number of flags needed makes it less likely that these flags will be used incorrectly. In my experience as a moderator on 3 sites, it's extremely rare to see someone other than myself flag an answer as AI. (Yes, I flag for my own attention. The current record on my sites is 4 other people flagging one answer as AI.)

As for #4, there is no flag that will formally warn the user or suspend them, though that is what's required in most cases. Deletion with no other consequences is too lenient, while red flag deletion is too harsh.


A note: The plagiarism flag on Stack Overflow is very similar in how it works to "In need of moderator intervention" except 1) it forces the user to give specific details and 2) it unlocks a special type of deletion (dissociation) that only moderators can use. It also creates a new category of flag that moderators can filter on, which is only a benefit on sites that have a lot of "In need of moderator intervention" flags (which is likely not very many outside of SO). I'm not sure that this gives enough benefits for an AI version to be worth a spot on the flag dialog; I'd rather keep that simple.

1
  • 2
    If the In need of moderator intervention flag can have custom messages, why couldn't a ChatGPT flag require custom messages too? Also, why couldn't a ChatGPT flag prompt an investigation, and contribute to deletion at say "orange" level? Commented Apr 24, 2023 at 0:30
7

In need of moderator intervention is the best option here.

I've usually described the output of chatgpt as 'barely plausible garbage'. It's believable but often incorrect. In some cases unless you try the answer, are a subject matter expert, or have seen so many of these posts that some part of your brain immediately recognises it, it looks possible.

And while a flag type that automatically deletes it seems 'useful' - it also means that mods don't actually notice the deletion, and might miss a broader pattern of such posts. On some sites, there's also suspensions involved.

Taking care of the 'immediate' problem by automatic deletion also might mean that only way that this gets handled is a potential answer ban.

I'm also hopeful that once the initial excitement wears off, and as the technology matures that there'll be less of these posts, simply cause it's no longer 'interesting' to ask the bot a question and just paste it here.

With some exceptions - we're generally able to deal with it. I'd rather be writing meta posts and such of course, but I don't really see a major benefit over our current options

2
  • I don't know the details of how flags are handled internally, but I would have thought if there was a dedicated flag, it would be easier to log, track patterns, and issue bans. Commented Apr 24, 2023 at 0:20
  • 1
    If they are handled automatically, we are not handling them - and deciding on the 'merits' of the flag. Between the 3 sites I moderate, one site hands out warnings, the other two suspend the user for an entire year. There's no pattern past the initial offense because users tend not to reoffend. Commented Apr 24, 2023 at 1:25

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .