16
$\begingroup$

Recently, there are a lot of new pseudonymous users with a peculiar pattern of posts: only answers, no questions, in lots of widely differing areas of mathematics (where it is unlikely that one random pseudonymous person is expert in all of them), posted in quick succession, which often look OK superficially (including MathJax an all), but actually miss some important step or make unwarranted assumptions or abuse notation in odd ways. I strongly suspect that these users are using AI tools to write their answers.

What should one do about these cases? (Especially if some answers are actually not that bad and get upvotes and accepts?)

$\endgroup$
18
  • 5
    $\begingroup$ Maybe this discussion is a bit related: ChatGPT strikes MathOverflow? $\endgroup$
    – Martin Sleziak
    Commented May 22 at 8:18
  • 11
    $\begingroup$ This is pure speculation, but I wonder if someone is running some kind of experiment on MO with these accounts. $\endgroup$
    – Sam Hopkins
    Commented May 22 at 13:59
  • 3
    $\begingroup$ @ToddTrimble: I noticed that one obviously chatbot-generated answer from a user posing as the founder of a major software company was deleted, but several other answers were not deleted. I do not work in the fields of those other answers, so I can't judge whether they should be flagged or not (but I am suspicious). $\endgroup$
    – Andy Putman
    Commented May 22 at 15:11
  • 3
    $\begingroup$ @AndyPutman Yes, it's the same user. I think in such cases you should act on your suspicions anyway -- thanks for asking. $\endgroup$
    – Todd Trimble Mod
    Commented May 22 at 15:37
  • 2
    $\begingroup$ Are you sure that there have been AI generated answers on this site that actually made sense? I haven't seen any, though of course my perspective is limited. In any event, upvotes or acceptance seem insufficient evidence for the quality of an answer (I do remember some upvoted and/or accepted AI answers that were complete gibberish). $\endgroup$
    – Christian Remling
    Commented May 22 at 17:39
  • 1
    $\begingroup$ Can I see which account this is? I'm currious as I've not seen any model with a good success rate on here or MSE. Try running the first 3 pages through any LLM - I guarantee you it will have at least a 95% failure rate. The future of research, I assume AI will eventually replace but not LLMs. $\endgroup$
    – Ethan Splaver
    Commented May 22 at 19:32
  • 1
    $\begingroup$ @ChristianRemling I always assume that an accept means the answer satisfied the asker, and thus presumably was understandable to them ... why would one accept gibberish? $\endgroup$
    – gmvh
    Commented May 23 at 8:52
  • 14
    $\begingroup$ @gmvh, if you have never seen a user so confused as to accept a nonsensical answer, you've been lucky. $\endgroup$
    – Gerry Myerson
    Commented May 23 at 11:12
  • 2
    $\begingroup$ @gmvh If an answer truly makes good sense then I don't know how one can be sure that it is AI-generated. In any case, I don't see why we should take action against good, sensible answers, regardless of their origin (unless maybe the question is so elementary as to be inappropriate for MO). On the other hand, I agree with Christian Remling and Gerry Myerson that just because an answer is highly upvoted or even accepted, it could still be gibberish. Gibberish should be flagged even if it is an accepted answer. $\endgroup$
    – Timothy Chow
    Commented May 23 at 12:29
  • 5
    $\begingroup$ @TimothyChow I don't know how closely you have been monitoring the deleted answers, but all the cases that I have seen and in some cases flagged involve "bullshitting with high powered language". That is, the answer will locally look impressive because it correctly states something about the Hodge-de Rham spectral sequence or a du Val singularity or something, but on inspection this has little or nothing to do with the question, and is applied incorrectly. The Boris Johnson approach, if you will $\endgroup$
    – Yemon Choi
    Commented May 23 at 13:47
  • 2
    $\begingroup$ "Apes don't read philosophy." "Yes they do, Otto, they just don't understand it." $\endgroup$
    – Yemon Choi
    Commented May 23 at 13:48
  • 1
    $\begingroup$ @SamHopkins Presumably the experiment is to see if they can goad me into apoplexy. $\endgroup$
    – Yemon Choi
    Commented May 23 at 13:49
  • 5
    $\begingroup$ Could we have an AI-generated answer to this one? :) $\endgroup$
    – YCor
    Commented May 24 at 0:30
  • 3
    $\begingroup$ @ChristianRemling In at least one case where an answer was flagged as possibly AI, I had to look hard and say "even if it were, this actually seems to hold up", and so the answer passed the Turing test as far as I was concerned. But in all cases I've seen where I was sure it was AI, there has consistently been a fly in the ointment. That may change. $\endgroup$
    – Todd Trimble Mod
    Commented May 26 at 1:44
  • 2
    $\begingroup$ @gmvh Sorry, I meant it as a joke. $\endgroup$
    – Roland Bacher
    Commented May 31 at 17:00

2 Answers 2

Reset to default
23
$\begingroup$

As per a comment by @Todd Trimble [now deleted, on request from a user -- TT], any posts suspected to be AI-generated should be flagged.

$\endgroup$
0
12
$\begingroup$

If AI-generated answers are actually correct - in particular, this will be relevant at a possible (likely? certain?) future date when AI becomes as effective at answering MO-style questions as human experts - I think it is not appropriate to post without mentioning that it was AI-generated, saying which AI system was used, and ideally also giving the prompt that was inputted.

In addition to concerns about honesty as to who deserves credit for an answer, this is because of the general expectation that the goal of answering questions is teaching, so in particular that answers should include information, if it's easy to provide, about how the question-asker can answer further similar questions in the future. If we use a book in our answer we cite the book, if we use a computer program to do calculations we say which one, and so on.

$\endgroup$

You must log in to answer this question.

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