47
$\begingroup$

I had a +200 bounty question set to expire in 2 hours when I got an answer to it. I suspected the answer of being written by ChatGPT based on the format, and material of the answer. The answer did not actually quite answer the question, but sounded exactly like how chatGPT might respond to the question. I posted a comment, asking about some of the detail of their answer, but also asking "Was this written by gpt or any other ai?". Soon after this, the OP deleted their answer.

I was wondering if there is a trend of bots posting chatGPT answers to questions with a bounty that is soon to expire. Motive being that a single person or organization could control dozens of separate high-rating accounts on math stackexchange for conspiracy purposes, since the answers being posted are not written by humans but rather ai bots.

$\endgroup$
18
  • 7
    $\begingroup$ Post & account in question for those curious. Two (1, 2) of their five posts are for bounties, I see. They seem to use LaTeX inconsistently as well, which I see ChatGPT often do. I find this especially strange for someone claiming two Ph.D.s in STEM fields. GPTZero flags most of their very few posts with $30$-$40\%$ probabilities of being AI-generated. $\endgroup$ Aug 18, 2023 at 19:45
  • 7
    $\begingroup$ As far as the more general trends beyond this person -- we've certainly noticed an uptick in people using ChatGPT and its ilk on MSE in general, both as methods of verifying work when asking questions, and when answering questions (up to and including the generation of whole posts, e.g. this recent incident). How to handle such people was a critical component of the recent moderator strike. $\endgroup$ Aug 18, 2023 at 19:45
  • $\begingroup$ For bounties specifically, I don't know myself, but I wouldn't be surprised if it's a trend. After all, if one is just going to use ChatGPT/etc. for answers, one can only conclude that their endgame is to get a lot of reputation points with minimal effort (even if they have good intentions, it's not like we can't use ChatGPT/etc.), and bounties have several ways of incentivizing that. $\endgroup$ Aug 18, 2023 at 19:45
  • 27
    $\begingroup$ One of their other answers includes this beauty: "This degree computation might involve the fact that the degree of a product of polynomials is the sum of the degrees of the individual polynomials". $\endgroup$
    – JonathanZ
    Aug 18, 2023 at 21:26
  • 2
    $\begingroup$ @JonathanZ Good catch! it wasn't my intention to call out the specific user, but, yeah, that's exactly the kind of phrasing that pops out to me as suspicious $\endgroup$
    – Snared
    Aug 18, 2023 at 22:02
  • 1
    $\begingroup$ Whats wrong in chatgpt's answer if it answers one question? (Im asking this in general) $\endgroup$ Aug 19, 2023 at 21:38
  • 12
    $\begingroup$ Well, the obvious concern is "what if that answer isn't sufficient for the person asking a question?" Like, it might help with a thoughtless calculation, on the assumption it's even accurate to begin with. But what if the OP or others using it down the road don't understand a step? What if follow-up questions in general are warranted? People using GPT to generate these answers obviously aren't sufficiently knowledgeable to answer them without GPT, so they can't help much further than that. They, likewise, clearly aren't sufficiently knowledgeable to ensure the claimed answer is right. $\endgroup$ Aug 19, 2023 at 21:46
  • 2
    $\begingroup$ @MathStackexchangeIsNotSoBad - please see my previous answer math.meta.stackexchange.com/questions/35651/… $\endgroup$
    – JonathanZ
    Aug 19, 2023 at 21:51
  • $\begingroup$ This behavior preceded ChatGPT though. Users post some half-assed answers right before a bounty period ends if the question hasn't received an answer that stands out from the rest. Then the bounty gets awarded to the half-assed answer if it receives 1 upvote and is the only answer for example. $\endgroup$
    – dezdichado
    Aug 23, 2023 at 15:38
  • 1
    $\begingroup$ @dezdichado - It's generally harder to discern that a ChatGPT is half assed, and it wastes more of users' time and energy. $\endgroup$
    – JonathanZ
    Aug 23, 2023 at 17:22
  • 5
    $\begingroup$ @JonathanZ I personally find it extremely easy to decide an answer is ChatGPT generated. The half-assed real answers I am referring to are usually the ones that answer a trivial case of the problem, but nevertheless they are complete solutions in their own regard. Meanwhile, ChatGPT answers are almost always some 4-5 paragraph, SAT essay format, thesaurus vomits. This only applies to math questions/answers - an answer generated by an LLM on a history stackexchange for example would be much harder to discern. $\endgroup$
    – dezdichado
    Aug 23, 2023 at 21:34
  • 5
    $\begingroup$ I would note that the profile picture for that user is also clearly AI generated. Their neck region is nonsensical, and the "sharp face but blurry background" is very characteristic of GANs. $\endgroup$ Aug 25, 2023 at 17:06
  • 1
    $\begingroup$ As a criterion for judging answers, I have always been a bit confused as to whether assedness is a positive or negative thing. I mean, if someone described an answer of mine as empty assed or full assed, I would not be sure how to interpret that. Perhaps that is a question for ChatGPT. $\endgroup$
    – copper.hat
    Aug 28, 2023 at 5:46
  • $\begingroup$ @dezdichado: You find it extremely easy. I too. So what? The question is whether chatgpt nonsense is useful for the site or not, not whether some of us (who are sufficiently competent) can identify and ignore it. $\endgroup$
    – user21820
    Sep 1, 2023 at 5:34
  • $\begingroup$ One potential solution could be a restriction stating that a question with bounty can be answered only by accounts with more than 1000 reputation $\endgroup$ Sep 1, 2023 at 12:47

0

You must log in to answer this question.

Browse other questions tagged .