2

Recently, a user generated 10 answers within 45 minutes, and by the formatting they seemed to be all AI generated.

Incidentally, the user asked a genuine question last month, so this is most likely a human account rather than a bot operated account.

Such instances would create a huge burden on the moderators, if many (real) users start following the same trend.

It only happened because Ask Ubuntu allowed the user to consecutively post such answers.

Can we have some mechanism to detect such instances, and stop the users from posting too many likely AI generated answers?

I would still prefer the users are allowed to post a single answer even if the system detects it was lilely AI generated (to prevent false positive, I would prefer the moderators to read and determine if the first answer was AI generated).

However, if there are many such consecutive answers, they will be defintely AI generated, and I propose we create a system to prevent such posts.

9
  • 1
    AFAIK, No bot has yet been created to detect this. In the interim, Is there a Query that can be composed to detect too many answers given within a short period of time? Hopefully someone with much more expertise than me can answer this.
    – stumblebee
    Commented May 25 at 2:11
  • "However, if there are many such consecutive answers, they will be definitely AI generated": How do you distinguish between this and a user who writes answers in an external editor before copy/pasting onto the site? And note that answers are rate limited, see the "Answering" section here: The Complete Rate-Limiting Guide
    – terdon
    Commented May 25 at 12:52
  • @terdon The AI detector will filter them when there are more than one answers likely generated by AI. If someone posts many human-written-answers, that is fine.
    – Archisman Panigrahi
    Commented May 25 at 13:39
  • 1
    AI detectors don't really work though. That's a big part of the problem. If we could reliably detect AI-written posts automatically, none of this would be an issue :/
    – terdon
    Commented May 25 at 13:40
  • I copy pasted some answers by that user to quillbot.com/ai-content-detector, and it showed 100% likely AI generated. I copy pasted my answers, and it showed 0% likely AI generated.
    – Archisman Panigrahi
    Commented May 25 at 13:43
  • 2
    @ArchismanPanigrahi should we take that your example with a single true positive and a single true negative does indeed trump users testing a wide variety of posts with AI detectors and finding them to still produce false positives and false negatives to a sufficient degree?
    – VLAZ
    Commented May 25 at 13:51
  • My proposal is, if there are more than one such posts within a short time, with high AI probability, we ask the OP to stop and post after 24 hours, or something like that.
    – Archisman Panigrahi
    Commented May 25 at 13:58
  • 1
    The Complete Rate-Limiting Guide @terdon shared mentions: Users with < 125 rep must wait 3 minutes between answers on most Stack Exchange sites, but must wait 30 minutes between answers on Stack Overflow. Why don't we do this for AskUbuntu too, changing the wait time from 3 to 30 minutes?
    – sotirov
    Commented May 26 at 15:45
  • @sotirov This is a good idea. Please propose this as an answer.
    – Archisman Panigrahi
    Commented May 27 at 13:42

2 Answers 2

Reset to default
4

According to the The Complete Rate-Limiting Guide:

Users with < 125 rep must wait 3 minutes between answers on most Stack Exchange sites, but must wait 30 minutes between answers on Stack Overflow. The longer wait time between answers on Stack Overflow was implemented as part of the response to a flood of ChatGPT answers.

Why don't we do this for AskUbuntu too, changing the wait time from 3 to 30 minutes for users with less than 125 reputation?

We are facing the same problem, why not use the same solution?

4
  • Are we facing the same problem? Should we really react like this based on one case? Do we have any evidence that it is a serious issue?
    – terdon
    Commented May 28 at 9:17
  • FYI, some mods on SO are in favor for returning the 3 minute rate limit on SO. meta.stackoverflow.com/a/429327 I also agree with that. It helped in the early AI days, but it is useless now.
    – Dalija Prasnikar
    Commented May 29 at 7:53
  • @terdon I do not know whether you had seen the occurrence real time, but just in case, I relay it to you: approximately one third of the front page had been dominated by the posts of this one user, a lot of those posts immediately adjacent to each other, indeed separated by only a few minutes from each other. Additionally, the user had used the Ubuntu logo as their avatar. It had an effect of impersonating the Ubuntu concept to do something so alien to Ubuntu: to blatantly disrespect the Ubuntu values and the community. The incident certainly had an angle of facilitating emotional distress.
    – Levente
    Commented Jun 1 at 13:51
  • Facilitating emotional distress!? That's a bit over the top. But anyway, if this just happened once or at least rarely, I just don't see how it's a problem.
    – terdon
    Commented Jun 1 at 14:01
6

There is currently no foolproof filtering system that can detect AI content.

The best you can do is flag something as suspected AI content and then moderators can review to determine if it is AI.

One of the things discovered during the AI evaluation last year around this time (due to mod strike, etc.) was an analysis of AI detection tools, AND the discovery you can tell AI to reword the same content to evade detection.

This introduces extra complexity to detections, and no detectors are foolproof. The human element of analyzing these posts is still relevant, and moderation does NOT rely solely on tooling.

There's no tooling that would work for this functionality currently, nor is there any foolproof way to detect AI content.

1
  • 1
    To add emphasis to "not foolproof", I've seen figures like "10-20% false positive rate" thrown around for such detection algorithms which if true would literally decimate legitimate content (and/or completely overwhelm moderators). Here's a paper which, if you scroll down far enough, shows false positive rates 10-50% in several of the categories. edintegrity.biomedcentral.com/articles/10.1007/…
    – thomasrutter
    Commented May 28 at 11:19

You must log in to answer this question.

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