108

Following the new temporary policy, content generated by ChatGPT must not be used as a question or answer on Stack Overflow. If I, as a non-moderator user, encounter a single post that I suspect is generated by an AI, what is the correct course of action?

Similarly, if I encounter a user who has multiple answers posted that seem to be generated by an AI, should I respond differently?

(This question was made in response to an answer, posted on the ChatGPT policy announcement. It was requested that a new question be made, to add visibility.)

1
  • If you suspect your neighbor is an AI, or is expressing ideas also shared by an AI, report him! Such individuals shall be identified and censored... using AI. Commented Nov 15, 2023 at 5:40

2 Answers 2

Reset to default
119

Flag it as "[this answer should be] looked at by a moderator". That's all you need to ever do.

Provide as much evidence as you possibly can to a moderator to explain why you think this is a ChatGPT response.

Don't respond/interact with the person directly if you can help it since we don't want to make this a witch hunt. People are going to react aggressively if you accuse them of being a witch using ChatGPT, so flag and let the mods sort that out.

Don't flag it "for not answering the question" or "as very low quality", because 1) that is not the problem with the answers, 2) these need to be reviewed by moderators, not the Low Quality Answers queue, and 3) if these flags are validated, they will result in extremely confusing audits.

If it's spam (real spam, the unsolicited marketing kind, not just unwanted content), then flag it as spam. Lots of spammers are now using generative AI to produce the contents of their spam posts. But if the post is actual spam, then it should be flagged as spam. The fact that it contains AI-generated content is irrelevant and outweighed by the fact that it is spam.

21
  • 46
    But what if the post actually weighs the same as a duck? Commented Dec 9, 2022 at 5:19
  • 2
    In my experience, having one AI answer means they have several. If you're feeling generous, see if you can flag any of the other answers on their network profile. Even if you're not completely certain, it would allow a moderator to investigate it further or make a note for the future. Commented Dec 9, 2022 at 14:15
  • I've a similar question: what if the answer is almost correct? Commented Dec 10, 2022 at 9:07
  • 12
    @shingo: Did I stutter? On the Internet, no less? Commented Dec 10, 2022 at 9:29
  • 30
    @shingo it's a blanket ban. Its correctness is irrelevant. If it's generated by CGPT, it's not allowed and has to go Commented Dec 10, 2022 at 10:05
  • 5
    It's hard to "provide as much evidence as you possibly can" due to the 500 character limit in custom flags. I just flagged an answer that I suspect was AI-generated, but I couldn't manage to fit all my reasons for suspecting it in only 500 characters. This is especially difficult since reasons for suspecting it often have to do with incoherences in the answer, which can sometimes be complicated to explain. Commented Dec 10, 2022 at 17:32
  • 8
    @shingo Verifying an almost-correct answer is more of a waste of time than a clearly wrong (instantly dismissable) one. Also, we don't want to encourage rep farming with a bot that produces wrong answers, so even "completely correct" GPT answers should go. Commented Dec 10, 2022 at 20:59
  • 1
    I was encountered some answers which are acceptable, but the tone, structure and small mistake made me suspect they were AI-generated. Just as @éclairevoyant said this is more of a waste of time so I asked if it is worth to flag. Commented Dec 11, 2022 at 8:36
  • 2
    @MendelG: I didn't misspeak here. If you think it's something generated by ChatGPT, flag it. Whether or not you think it's of quality enough to be voted on is a separate matter altogether. Commented Dec 21, 2022 at 0:28
  • 2
    @Makoto might be worth mentioning in your answer that it's one flag per user regardless of the number of ChatGPT answers they've posted but mention in the flag that other answers by that user need to be considered. Commented Dec 23, 2022 at 15:14
  • 2
    @shingo More important than wasting time, IMHO, is that ChatGPT doesn't cite its sources, so you can effectively consider anything it writes to be plagiarized. Commented Dec 24, 2022 at 23:49
  • 15
    @wjandrea anything copied from ChatGPT without proper citation is plagiarized in itself, and goes against ChatGPT’s own Sharing & Publication Policy Commented Dec 28, 2022 at 23:38
  • 2
    @Makoto your first paragraph here is correct. Your second paragraph is wrong, pointless, confusing and adds confusion. I would suggest you delete the second paragraph. (ChatGPT output, or output from any of the dozen similar bots that hasn't featured on GoodMorningAmerica, is instantly recognizable and anyone who takes the time to flag it need only tap flag and type "chat bot". Much like is you flag, say, swearing.) Commented Jan 1, 2023 at 4:27
  • 3
    @srn: This is still the right course of action. Whether or not it gets actioned is up to Stack Overflow Inc. and a speedy resolution to the moderation strike. Commented Jul 9, 2023 at 19:56
  • 3
    @QuackE.Duck, yes it would be better if ChatGPT answers are deleted. However, most users cannot delete answers, but many users can downvote. (Probably they hope that OP will remove the answer or get banned due to bad reception of their posts.) Commented Aug 14, 2023 at 4:53
-30

Before flagging the post - I would talk to the poster about it. That is, ask explicitly, in a comment, whether they used ChatGPT to generate the post.

Now, it's possible, of course, that a poster would try to hide having done this. But - I believe that's not very likely. While some users view for reputation, I believe only a few users are so interested in illicit SO reputation gain that they would secretly use ChatGPT. And even those - after a couple of incidents of this happening, you'll be able to tell from looking at their other questions.

So, I expect that posters are most likely to:

  1. Say "Yes, why?" - in which case you can point them to the relevant policy and ask them to rewrite/delete.
  2. Say "No, I know the answer because XYZ", in which you would probably believe them.
  3. Say nothing for a while, in which case they don't care about their question/answer much anyway.

In all three cases you are in a better position to decide whether to flag, and the moderator is in a better position to act.

11
  • 8
    Or, just use AI detection software... They're pretty accurate. Accurate enough for reasonable doubt. An OP dishonest enough to use ChatGPT to answer, surely won't have issues lying about using it. Commented Feb 20, 2023 at 7:00
  • 2
    I've found commenting to ask for rule violations can lead to the specific evidence just being deleted. While enough rep means you can still access that content and flag it, it makes it much more cumbersome to do so. Commented Feb 20, 2023 at 8:48
  • 3
    @Cerbrus: 1. Link please? 2. Many people use ChatGPT in good faith. Copy-pasting that as a SO answer is poor judgement, but may not necessarily be "dishonest". Commented Feb 20, 2023 at 8:58
  • 4
    The whole reason we have this rule is because many users used ChatGPT in bad fait. Copy-pasting answers in bulk. These detectors are trivial to google... The HuggingFace one is pretty good. Commented Feb 20, 2023 at 9:23
  • @Cerbrus: Well, many users/editors would not DDG these detectors - because it hasn't occurred to them to try. Also, what is "many"? Commented Feb 20, 2023 at 10:29
  • 2
    "DDG"? "Many" is plenty enough to get noticed, enough to flood the mod team with bad answers, and enough to warrant this whole ban in the first place. Commented Feb 20, 2023 at 10:51
  • 1
    1. "DDG?" <- duckduckgo.com , no need to live under Google's thumb. Not that they're perfect, but at least they don't seem to spy on you and use the data for all sorts of nefarious purposes. 2. That "many" is IMHO still so little, that when a user encounters a subjectively-suspicious answer, it is likely to be a false suspicion, or a non-malicious lapse of judgement by the poster. Commented Feb 20, 2023 at 11:17
  • 4
    "Google" is a verb nowadays... "DDG", well, I've never seen it used like that before. Let's not go into politics. Your "IMHO" is completely irrelevant as it clearly was plenty to require such drastic measures. Your perception of the flood of ChatGPT content we got on this site is so far off-base... Maybe it wasn't as prevalent in the tags you frequent, but that doesn't mean the problem doesn't exist, nor does it mean the block is unjustified... Frankly, your suggestion to just "ask the OP" is naïve, and would be a waste of everyone's time. Commented Feb 20, 2023 at 15:05
  • 7
    Regarding the suggestion in the comments to "just use AI detection software", please note that many of these AI detectors are notoriously inaccurate. I've personally found many examples of the Hugging Face detector reporting 99% confidence on clearly non-GPT posts. Moderators will never make a determination solely based on them. Commented Jul 14, 2023 at 0:51
  • 5
    Also yes, users absolutely lie about using ChatGPT/other AI tools. Not all of them, probably not even a majority (the most common response, I believe, is "Sorry, I didn't know it was banned"), but a nontrivial number. Commented Jul 14, 2023 at 0:54
  • @Cerbrus: I suppose I am biased by my tag preferences. TBH, I've only ever seen one answer to a question that I suspected might be AI-generated. Maybe it was because of careful modding (before the strike), maybe it's that some tags get a deluge. This question was phrased in the first person, so I thought about what I would actually do, and wrote that. Commented Jul 14, 2023 at 14:33

You must log in to answer this question.

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.