Consider this answer to this question:

The couple in question is James Morley and his wife, Liu Xiaobo.

hope this helped.

That's the kind of thing that GPT-3 or a similar AI would misinterpret out of this webpage, but a mistake that no human would make:

Nobel Peace Prize Winner’s Wife Detained in China

Added by James Morley on October 11, 2010

A U.S. human rights group stated that the wife of Liu Xiaobo, the Nobel Peace Prize winner, was detained by Chinese authorities in her home not being allowed to leave the premises.

Or this answer to this question:

It seems to me that the world of digital video and audio, even though it was vastly more expensive to store, was even further behind in the compression game than the world of digital images.

(The above was written on May 6th, 2009. Since then, I've learned a few things about video and audio compression, but it still puzzles me. I wrote about this in the article "Video and audio compression - a history".)


In the early days of the Internet, there were three protocols that were used to transfer data between computers: FTP, HTTP and SMTP.

FTP was used to transfer files between computers, and SMTP was used to transfer text messages between computers.

At first, HTML pages were transferred using the FTP protocol, but later the HTTP protocol was developed specifically for this purpose.


The HTTP protocol was developed by Tim Berners-Lee, the father of HTML.

He had this to say about the design of HTML:

HTML was designed to display data and to be as simple as possible. HTML has gone through a number of revisions since Tim Berners-Lee invented it in 1991. HTML is usually written

It's complete nonsense, and again, not the kind that a human would write.

This same user has a ton of answers on Stack Overflow, all posted in very quick succession, and most that look reasonable at first glance to someone who's not an expert in the particular topic, but almost invariably turn out to be wrong when scrutinized or actually used (e.g., Why is no error thrown after resolve() in a Promise?), though some that are more easily noticeable (e.g., https://stackoverflow.com/a/74326200/7509065).

Do we have a rule against AI-generated answers yet? Is there something that can be done about this user other than every community having to delete each of his answers one by one?

  • 4
    Let's flip the premise: ignore the AI angle. What if a user just posts low quality answers? I believe what should be done then, should also apply to a user posting low quality answers by an AI.
    – VLAZ
    Commented Nov 9, 2022 at 7:10
  • 7
    Much worse, almost all of their answer has a +1 vote. Something sneaky may be going under the hood. Commented Nov 9, 2022 at 7:16
  • flagging on smaller sites where appropriate sounds like a good idea. SO's trickier cause of sheer volume but I guess they or the CMs can pick it up from here. Commented Nov 9, 2022 at 8:07
  • 9
    @VLAZ When a user writes low-quality answers, my reaction is "perhaps we can educate them about how to post better answers." When a user posts answers generated in an automated fashion, my reaction is "why did you think this was okay?"
    – Ryan M
    Commented Nov 9, 2022 at 9:34
  • Presumably, SmokeDetector would detect it(?). Commented Nov 9, 2022 at 12:20
  • 2
    How do you know it is AI generated? There may be some selection bias, but for the remaining three on Stack Overflow, it looks more like paid homework, with perhaps some plagiarism thrown in (two are in flawless English while the third wasn't until someone edited it into shape). If it wasn't for the third, it could have been someone who just happened to be a very good writer. Commented Nov 9, 2022 at 13:25
  • cont' - I don't think a machine would use two completely different writing styles. The broken code formatting for the third is consistent with copying off a web page (plagiarism). Commented Nov 9, 2022 at 13:25
  • cont', 2 - Paid homework (or more generally paid commissions for answering questions on Stack Overflow): The questions are all over the place: Haskell (basic. Likely homework), Ruby/Jekyll (web. Fairly advanced, with a bounty), and Java (basic time stuff, but with answers from several very high-rep users). Plagiarism: Straight copy-paste, not generated text. Commented Nov 9, 2022 at 13:35
  • "Do we have a rule against AI-generated answers yet?" only if they're useful, then XKCD #810 Commented Nov 9, 2022 at 14:05
  • 3
    @This_is_NOT_a_forum GPT-3 absolutely does generate text with multiple completely different writing styles. Commented Nov 9, 2022 at 15:07
  • 1
    @This_is_NOT_a_forum On Stack Overflow, there were enough answers (almost 40) that I was able to go through and pick out ones that simply did not make sense. Having seen other AI-written answers on Stack Overflow, I am fairly confident that's what I was looking at. One dead giveaway was an encoding bug: they accidentally include HTML entities in code that should have been just things like < or >.
    – Ryan M
    Commented Nov 11, 2022 at 16:27
  • 1
    It got worse. Commented Dec 5, 2022 at 22:32

2 Answers 2


We don't particularly need a separate policy; we already have processes for handling low-quality content. To the extent that the AI-generated content can be construed as abusing the system (which I don't think is a far stretch), the rules we have against that also kick in here.

Escalating with a meta post and/or mod-flagging problematic content which is not already being downvoted and ultimately deleted by regular users is also part of the processes we have; thanks for doing that, and for bringing our attention to this particular behavior.

  • 5
    I agree with this approach, personally. Consistently low-quality contributors are already within community & moderator purview to manage. When it's possible to identify the cause as AI generation, that removes plausibility that it's just a misguided user - which may inform the ways mods can react. And when mods need support for widespread abuse they're not able to handle independently, they always have the option to escalate to us. We've seen a case or two like this already -- and for now, this standard approach seems to be working.
    – Slate StaffMod
    Commented Nov 9, 2022 at 10:00

Not everything needs to be explicitly spelled out - and if there's something obviously 'wrong', feel free to escalate it.

I've brought it up to the SO mods and custom flagged elsewhere, but here are a few thoughts.

I really don't want to call out specific users on a per site meta in general, but this is messy enough that it could warrant it. Especially on Stack Overflow and its busy flag queue this might be a good idea.

Nice catch - and this is the sort of situation that needs human review.

You must log in to answer this question.

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