8

Recently (today/yesterday) there were several answers posted by a "user" (Jordan Thomas) who appeared to some to be using Chat GPT or another type of artificial intelligence (AI) tool to garner answers and post them. This obviously defeats the purpose of first-hand user experience being shared, and relying on a machine to generate an answer.

Vladimir noted the "smell of" Chat GPT in the responses, so credit goes to him on the initial detection and exposure. Is it Chat GPT? That I am not sure of, but I do agree with Vladimir's intuition.

With Chat GPT's new presence in our world, Google has pushed forward hard to release its own AI answer bot, so this is only going to get worse. Heads up!

What should be done here at BE (and other exchange sites) to deal with this intrusion? Is it welcome? Is it not? If it is not, then we should see about communicating a way to detect it (if we can), and police it consistently by the active members of BE.

1

4 Answers 4

Reset to default
3

The main SO policy is found at https://stackoverflow.com/help/gpt-policy

Please make points for/against adopting this as our site's policy too.

I'll give it a day for responses.

update I have deleted all the flagged posts. Every one was well over 99.96% likely to be chatGPT based on the detector.

If you think something is chatGPT, flag it in whatever way seems appropriate and it'll get actioned.

5
  • 2
    Is this worth a separate question? You’re essentially asking if using ChatGPT answers in their entirety should be allowed, or at least what policies to adopt about the use of ChatGPT. The original question asks how moderators can screen those out.
    – Weiwen Ng
    Commented Feb 15, 2023 at 0:11
  • @WeiwenNg could do - but do we want to make a consultation take even longer? I'm intending to delete chatgpt answers on sight, and then deal with any results afterward. As it stands, these answers have now been published under the SE name for ~4 days. For an example, leaving graffiti on a wall is a sure-fire way to attract more - instead it needs to be gone ASAP.
    – Criggie Mod
    Commented Feb 15, 2023 at 1:24
  • 2
    @WeiwenNg I posted a new question proposing an official ban with specific language. bicycles.meta.stackexchange.com/questions/1506/…
    – jimchristie Mod
    Commented Feb 15, 2023 at 13:40
  • 1
    @Criggie I'm looking at what I think is an AI-generated answer right now. Can you comment on the best flag to use? Option being "Spam," "Rude or abusive," "Not an answer," "Very low quality," or "In need of moderator intervention." DavidW commented below that he uses the last one but I'm curious if there's an official position yet on which to use. Or do we need a new flag for "Looks like AI."
    – SSilk
    Commented Aug 31, 2023 at 7:08
  • 1
    @SSilk hey Hydrogen one? yep I see that it got 6 downvotes and was automatically deleted by Community Bot before any of the mods got to it. As for the best flag to use, I'd pick low quality or in need of mod attention. Though realistically we only get a flag every couple of days, so anything will work. Some of the really busy SE sites see a hundred flags a day.
    – Criggie Mod
    Commented Aug 31, 2023 at 10:52
8

Executive summary: Bot-generated answers create unreasonable demands on the expertise of knowledgeable contributors; they should be flagged and deleted.

I concur with the assessment that they are ChatGPT output. They all have that particular feel.

As to what to do... I think it's important to flag them so the moderators are aware; they have more tools (and connections) to investigate than we, as users, do and they also have more powers, beyond just downvote and delete.

The thing to keep in mind about these answers is that while they are superficially plausible, and written in good English, ChatGPT doesn't know actual specific facts. On the other site I am active, it's sometimes easy to spot generated answers because they refer to works that don't exist, or if they do exist they get the attribution or date wrong. But if those are all correct, then someone with knowledge of the work has to notice that the description doesn't match.

And that, in my opinion, is the biggest problem with ChatGPT answers. Not that they're always wrong, because it does get some things right. The problem is that it takes someone with either domain knowledge or the time to do some research to determine how much is fact and how much is made up. And that cost is an unfair burden on the community.

That any random drive-by poster can spend two minutes to generate reasonable-sounding text that needs a knowledgeable contributor to spend five minutes trying to winnow the wheat from the chaff is a problem. That's not a dynamic that tilts in favour of the community. I think the only real solution is to flag and delete.

3
  • 1
    When flagging, the only applicable flag is “no comment needed”, which is a bit too non-specific. I am worried about this, and it would be nice if a moderator could comment on how we might flag a suspect post.
    – Weiwen Ng
    Commented Feb 12, 2023 at 18:53
  • 3
    @WeiwenNg I always flag "needs moderator attention." Even if I come across the post in the low-quality queue, I'll still flag it since it's not public knowledge if other users have already flagged it. In the comment field I describe what makes me think it's chat-bot output. (Of course those flags may not be cleared fast, because I'm sure it generates discussion behind the scenes. But at least the moderators are aware of what's going on, and if it continues to be an issue they'll have the data necessary to help formulate an official policy.)
    – DavidW
    Commented Feb 12, 2023 at 18:59
  • 1
    Ok, I was thinking of the low quality review queue. Flagging directly from the post itself will produce the option for mod attention. Thanks.
    – Weiwen Ng
    Commented Feb 12, 2023 at 23:24
3

From Criggie's link, first three lines of second paragraph:

Stack Overflow is a community built upon trust. The community trusts that users are submitting answers that reflect what they actually know to be accurate and that they and their peers have the knowledge and skill set to verify and validate those answers. The system relies on users to verify and validate contributions by other users with the tools we offer, including responsible use of upvotes and downvotes.

This captures my thoughts as well (hinted at in the original question). Just put "Bike Exchange" in where Stack Overflow is mentioned. Chat GPT, et al, is not in the spirit of the above, so I would support the SO policy being adopted for BE.

1

Yep - chatGPT has recognisable characteristics, like simple short paragraphs, no images at all, and a stilted meter where there's an intro, 3~4 supporting points, and a conclusion paragraph.

It also can latch onto a word in the question that is important but not the focus of the point. EG, that "ride with disabled sister" question says trailer, so that is more significant in the answer than a normal English speaker.

ChatGPT also makes no spelling mistaeks, and uses no humour, which can be a tell.

Personally I'd delete the answers, but I've consulted the other mods about a policy. More info to come.

5
  • 1
    Hillariously, posting this answer made SE check if I was a robot, with a captcha.
    – Criggie Mod
    Commented Feb 13, 2023 at 9:22
  • 1
    In a comment on a suspect answer, you mentioned a ChatGPT detection tool that estimates the probability that the answer was written by an AI. That may be worth adding, even though it is something we can only apply retroactively for now.
    – Weiwen Ng
    Commented Feb 13, 2023 at 18:56
  • @WeiwenNg openai-openai-detector.hf.space its not perfect but indicative. This answer scores 1.75% fake.
    – Criggie Mod
    Commented Feb 13, 2023 at 19:39
  • 1
    @Criggie So there's still a chance you're a bot, eh? :D (Mine only scores 0.02% likelihood of fake, so nyah.)
    – DavidW
    Commented Feb 13, 2023 at 21:29
  • @DavidW Amusingly, the detector thinks a number of my work emails are bottish too. That says a lot about senior management.
    – Criggie Mod
    Commented Feb 13, 2023 at 21:30

You must log in to answer this question.

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