4

Context: Informal ban already effective

Meta Stackoverflow and Meta Stackexchange has already been discussing a (temporary?) ban of answers crafted with ChatGPT (SO 421831, SE 384396).

Answers identified as generated by ChatGPT have been already being deleted by moderators.

Example ChatGPT answers deleted
278176
(Please update references)

Currently, answers generated by Artificial Intelligence (AI) are likely to be sub-standards. If the time arrives when AI generated answers could be good enough or superior to human-generated, there would be no much point on storing AI generated answers here anyhow.

AI generate answers usually don't give correct attributions, and may constitute plagiarism.

Request for Discussion:

  • Explicitly ban AI generated answers.

  • Add a warning message.

What is left to do is to explicitly document the ban on Mathematica StackExchange and add a message informing the users that AI generated answers are not allowed and will be deleted.

Identification

It remains unclear what is the criteria for identification of AI generated answers, and how to deal with ambiguous posts. This should be discussed here too.

4
  • 1
    Recently, I suspected an answer on this site was generated by some AI because it was stylistically very well written, yet complete non-sense. After checking some of the user's posts on other SE sites, I became completely convinced that he did not write the answers himself. I warned him with a comment under the answer, and he removed his answer. (However, there is one other answer here on MME for which I have the same suspicion, yet he did not remove it.) To help with the decision, there are some GPT detection tools which seem to be quite accurate.
    – Domen
    Commented Jan 6, 2023 at 12:20
  • 1
    I would appreciate help finding the questions that have been already deleted because they are ChatGPT, that will illuminate this discussion. My point is that "ChatGPT as source" has been used as the main argument for closure.
    – rhermans
    Commented Jan 6, 2023 at 13:57
  • A sample of the posting has probability 0.9889 of being real. The above comments have probabilities 0.9996 and 0.9414 of being real. Checked with GPT-2 Output Detector Demo.
    – mjw
    Commented Mar 14, 2023 at 3:51
  • 1
    Here is a user claiming to be ChatGPT: mathematica.stackexchange.com/questions/285508/…
    – Michael E2
    Commented May 23, 2023 at 0:04

3 Answers 3

10

I don't see why a special rule is required. If an answer is nonsense we should be able to detect, flag, and delete it based on that alone. If the robots are smart enough to fool us, then we've already lost.

6
  • I disagree. A special rule is already enforced, even if it hasn't been spelled out explicitly in [meta]. This is about making it explicit, adding a warning and working out identification.
    – rhermans
    Commented Jan 6, 2023 at 13:48
  • 2
    I don't understand your reply here, @rhermans. If we're already assuming that the special, but still tacit, rule is appropriate, then what is there really left to discuss? I thought part of raising the discussion was to decide whether the special was indeed appropriate. Maybe there's some SE or SO culture I'm not grokking here.
    – lericr
    Commented Jan 6, 2023 at 14:00
  • @lericr What is left discuss is if we make that rule explicit (or reject the rule?). To claim that the quality is the only relevant parameter ignores the fact that "ChatGPT as source" is already been used as reason for closure.
    – rhermans
    Commented Jan 6, 2023 at 14:06
  • I'm so confused. What I think maybe you're saying is that the precedent has been established and therefore cannot be challenged. If that's the case, then fine. Make it explicit.
    – lericr
    Commented Jan 6, 2023 at 14:09
  • The first item under "Request for discussion" is "Explicitly ban AI generated answers". I interpreted that to be a request to discuss the ban, not a request to dscuss its explicitness. If the ban is in effect and cannot be changed, then make it explicit.
    – lericr
    Commented Jan 6, 2023 at 14:11
  • @lericr it can be challenged, for sure. I agree with the ban, but I think we should discuss it. I don't want hidden rules. That said, I'm a bit confused too. Probably I didn't articulate my question in the best way. Let's work it out.
    – rhermans
    Commented Jan 6, 2023 at 14:20
7

It seems to me there are two separate issues.

  1. Quality of ChatGPT answers. You've said that ChatGPT answers "are likely to be sub-standard", but if that's the case, then the normal mechanism for voting on and accepting answers should deal with this. Or stated another way, if a ChatGPT answer is legitimately good, then why would we want to eliminate it?

  2. Ethics. If someone posts a ChatGPT answer without the proper attribution, that seems like an ethical violation. Maybe the target should be those users instead of ChatGPT answers in general. Addressing the plagiarism-like behavior here seems more important than the quality of the answer. On the other hand, if someone attributed the answer clearly, is it necessary to delete that content?

UPDATE

Maybe I misunderstood what was being asked of us, but rather than argue the meta point, let me just clarify my opinion. And that's all it is, my opinion. I'm not a moderator and I'm not going to pout about the outcome either way.

  • It should be expected that if ai-generated answers are posted that they be attributed correctly. So, at the very least some consequences should be meted out to violators, i.e. plagiarizers (if that's the right term).
  • If we allow ai-generated answers (properly attributed), then the normal processes for voting and deleting would naturally come into play, so there is no real threat to quality.
  • I don't know how to enforce it, but it also seems like one shouldn't be able to earn reputation by posting ai-generated answers. More of a side-note than a critical aspect.
  • I don't think ChatGPT poses an existential threat to these communities, at least not in the near future. Sometimes answers are best when they contain more than just answers. So, I think humans will continue to have utility until AI can not only generate answers, but also meta answers. So, I don't think we need to get overly worked up about having the occasional ChatGPT answer in our midst.
  • Having said all of that, I don't actually object to banning ai-generated answers. I don't see it as something precious that needs protecting. I'm not eager to have to start separating ChatGPT wheat from chaff. I think it's reasonable to argue that the cost-benefit analysis favors banning.
  • If our policy is to ban ChatGPT, then I think we should make that explicit.
6
  • See the moderator comment for the first linked closed answer : "This user has been known to produce answers using ChatGPT on other sites and has been suspended elsewhere for it. Reading the code, it looks like something that was likely produced by AI, as I cannot imagine a serious Mathematica programmer using[...].". My point is that "ChatGPT as source" has been used as the main argument for closure.
    – rhermans
    Commented Jan 6, 2023 at 13:59
  • I understand that. What I don't understand is what you are now asking of us. If you're asking "it's already a de facto rule that chatgpt generated content is banned (and modifying that decision is not on the table), so should we make that ban explicit?", then I guess I'd say "yes, obviously". But if you're asking, "do we all agree that banning content by fact of it being chatgpt generated is appropriate", then I'd say "interesting question--see my answer".
    – lericr
    Commented Jan 6, 2023 at 14:04
  • Implicitly I'm calling to discuss both, I assume that on the process of spelling out an explicit ban, we will go through the reasons not to enforce such ban, and potentially revert the idea of banning it on the first place. You say "yes, obviously" however the official answer on Meta.SE is not to have an explicit blanket policy. There is a policy, that is not official, that is been enforced. That is the topic. I agree with the policy, but don't like it to be obscure. I would rather discuss it and agree with the community that this is what we want.
    – rhermans
    Commented Jan 6, 2023 at 14:16
  • To answer "ChatGPT answer is legitimately good, then why would we want to eliminate it?" because there is no point on making SE a repository of AI generated answers. The day AI generated answers replace a Mma.SE user, then this site becomes obsolete.
    – rhermans
    Commented Jan 6, 2023 at 14:26
  • 1
    The prediction of obsolescence is premature, or at least the pace of obsolescence is exaggerated. Maybe 1 out of every 100 answers receives an ai-generated answer, and maybe only a fraction of those get blessed by the OP. Okay, we're on the way to obsolescence, but in the meantime we have some interesting data to digest and a sometimes-useful tool to help us find answers. Declaring that this possible future, gradual slide into obsolescense means we should just shut down SE today seems overkill.
    – lericr
    Commented Jan 6, 2023 at 14:40
  • 1
    This all rests on the assumption that we, humans, can detect if an answer is AI generated. Can we do this really? Even the referenced case deduces involvement of AI based on low quality and prior knowledge of the user's notoriety to use AI. Without proper definition of detection rules, I think it is pointless to try to differentiate AI-generated low quality posts from human-generated low quality posts. Commented Jan 6, 2023 at 15:56
6

Let me just clarify a couple of points that have been brought up, as it was not my intent to create a new shadow policy.

  • For the ChatGPT answer referenced, I wasn't out looking for ChatGPT answers, but I was responding to a moderator flag from a contributing member of this community (who also happens to be active on several another Stack Exchange sites, including as a moderator), alerting us to the OP's pattern of only posting ChatGPT answers all across the SE network and getting suspended for their behavior.
  • Following up on the flag, I deleted it under the existing "low quality" + "spammy behavior on SE" criteria rather than a strict "no ChatGPT" rule. My comment was more to clarify why I thought it was a non-serious answer despite it looking coherently written at first glance. You, and others with 20k+ reputation can undelete the answer if you disagree with my decision :)
  • I do not know of other questions/answers that have been deleted by moderators purely because they were ChatGPT. I certainly haven't done it and I don't recall seeing any either.

I think it's healthy to have the discussion that you're proposing and I defer to the community to come up with a comprehensive approach to dealing with ChatGPT answers.

I personally don't have a strong opinion on the use of ChatGPT on this site at this time, but I like the balanced take that leicr is proposing. Obviously I wouldn't want to see this site turn into a playground of AI generated answers (answering to AI generated questions, perhaps 🤔) but I can envision some valid use cases, such as non-native English speakers using ChatGPT as a guide to fill in the explanatory details in a more "native speaker voice", while relying on their expertise for the code and problem solving parts. As with any emergent technology, it can be put to good use or abused :)

You must log in to answer this question.

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