23

Do we have a firm policy on questions, answers, or comments that are automatically generated using a tool like GPT-3?

This post follows a previous request for Academia SE community input: Should this site take an official position on answers generated by ChatGPT? See also the related discussions on StackOverflow (see also here) and network wide.

0

2 Answers 2

27

Policy: Questions, answers, and comments may not be auto-generated on Academia.SE.

While the technology is impressive, these answers are not based on references or personal expertise in academia. Rather, our experience is that they only contain contain common sense advice that is often off-topic. There is also significant potential for harm since the AI might produce plausible-looking answers to difficult questions; these might attract upvotes and give the impression of credibility when in fact the answer could well be incorrect.

If there is some reason why you feel that you need to auto-generate text, it's probably best to seek pre-approval on meta. Note that simply citing the AI model used or disclosing that the text was auto-generated is not sufficient under this policy (though it's a step in the right direction).

What should I do if I see something that I think was written by a chatbot?

Flag it! Good-faith flags on questions or answers will usually be marked "helpful." Good-faith flags on comments are also appreciated, but mechanically such flags are always marked as "declined" unless the comment is deleted.

How will you know if someone has used a chatbot?

We deliberately don't reveal our entire bag of tricks. It is possible there will be some false positives. However, answers that are sufficiently low-quality as to be indistinguishable from a chatbot are also not really desirable, so we don't see this as much of a problem. False negatives are a bit more of a problem, since an answer may seem okay but actually be harmful.

What if I'm asking a question about dealing with chatbots in an academic setting?

The question itself may not be auto-generated. However, if quoting the exact words that the chatbot used is important for your question, you may provide a brief quote from the chatbot.

What if I have a suggestion for improving this policy?

Make a new meta post to propose any changes.

19
  • If we find an old answer to an old question that is way before the official start date (11/30/2022) of GPT-3 (I might be wrong about this date) and it looks very suspicious, can we flag it? I asked a question in another Chat room that someone could use SE to test an AI product?. The answer I got was that we don't know one way or another. So, what's our policy on the old stuff? Retroactive or let it ago?
    – Nobody
    Dec 11, 2022 at 9:28
  • 2
    I would say that the bar on the old stuff is a bit higher -- e.g., we probably won't hammer reasonable-looking, well-received answers months after the fact unless there's very good cause to do so. But if you find something that you think might meet this high standard, I would certainly want to take a look! So, my recommendation would be to raise a "custom" flag and make it clear in the text what your concern is.
    – cag51 Mod
    Dec 11, 2022 at 9:48
  • 1
    I don't like using plagiarism as an argument because it is prone to this objection: what if I write an AI that registers to a website and posts answers without human intervention? Then, arguably, the account belongs to the AI, and there is no plagiarism issue. ("AI in its current state produces tons of plausible-looking answer that are incorrect but hard to spot and waste everyone's time" is a much better one.) Dec 11, 2022 at 16:21
  • 4
    Yes, I hesitated at that as well -- whether this is considered plagiarism at all may change in the coming years (e.g., we probably wouldn't consider evaluating an integral on Mathematica to be plagiarism). Took a stab at an update.
    – cag51 Mod
    Dec 11, 2022 at 19:31
  • 3
    @FedericoPoloni There are two concerning aspects of plagiarism. One is with regard to attribution to e.g. ChatGPT - this is solvable by being open that an answer is generated by ChatGPT. The other is a problem inherent in statistical language models which is that nothing prevents them from lifting wholesale from their training material.
    – Bryan Krause Mod
    Dec 12, 2022 at 15:42
  • 1
    @BryanKrause True. Before the edit (version 1 of this answer) I think it was more clear that my comment is about the first of the two problems only. Dec 12, 2022 at 17:54
  • This does not seem to actually set any policy since it does not say what action moderators, stack exchange, or the community will take as a result of the undisclosed "bag of tricks." Dec 16, 2022 at 0:10
  • Does this change anything other than changing the flagging process from "available" to "encouraged?" Dec 16, 2022 at 0:12
  • 2
    This seems like an arbitrary objection; I don't think any of our past policies have codified the punishments for policy-violators. Moreover, the punishments are well known: we usually remove content and give public warnings, and in more serious cases we give formal, private warnings or suspensions. As for what the policy changes: the change is that we now disallow content that was written by an AI even if it otherwise meets our quality standards.
    – cag51 Mod
    Dec 16, 2022 at 4:54
  • How are answers like this one treated under the policy? On the one hand, it contains autogenerated text (at least it is declared as such). On the other hand, that only makes up part of the answer, and is arguably relevant since the question is about ChatGPT and similar AIs. Personally I am in favor of keeping that answer (and answers like it) around, but it is unclear to me whether the current policy agrees or takes a harder line.
    – Anyon
    Jan 4, 2023 at 13:27
  • I agree, I think that answer is fine. Perhaps we should clarify the policy on this point.
    – cag51 Mod
    Jan 4, 2023 at 17:28
  • I am not sure I need to ask another meta question. Anyway, here you go. To the mods, if we see a post which looks like an AI-generated text, provided by a low rep user (not a new contributor), which way would the mods prefer us to flag the post? A custom flag or an r/a flag? Please advise.
    – Nobody
    Mar 29, 2023 at 15:03
  • 4
    @GoodDeeds - good question. At a minimum, the guidance will change to something like "flag it, but don't mention ChatGPT, just say 'low quality' or some other euphemism." At a maximum, this policy may be overturned completely; there is no point banning ChatGPT if we're not allowed to delete things that are consistent with being automatically generated. We'll follow up soon; at the moment, we are still reeling from the terrible policy being dumped on us, and the resulting resignation(s).
    – cag51 Mod
    May 31, 2023 at 17:37
  • 3
    We have never deleted incorrect answers before, that would be a huge change. In the past, incorrect answers have somewhat-reliably been downvoted...but as the chat-GPT answers are plausible and well-written, and more people (with less knowledge of academia) are given the upvote privilege, this may not hold true in future.
    – cag51 Mod
    May 31, 2023 at 20:57
  • 1
    Physics.SE has clarified that their processes are working as intended and they intend to continue following them. I would be particularly interested in a similar type of clarification laying out the position of Academia.SE's moderation team: physics.meta.stackexchange.com/questions/14438/…
    – user137975
    Jun 2, 2023 at 18:18
5

While there may be background, even automated, solutions, I suggest a user-facing change to some of the user dialogs.

Adding "Seems to be auto-generated", or similar, to either or both the flag and site-specific close dialogs would let users help keep things out as needed.

This is, perhaps, fraught if it would flood the mods with false positives, so consideration/thought is needed.

If SE overall makes a policy then such might be added to the more general close dialog, of course.

I don't know how difficult it is to change these dialogs and how much power the mods have to do it on their own. A "feature request" might be needed.

Added: It might be useful to add something to the help center advice.

3
  • 4
    For now the custom/"in need of moderator attention" is fine; it's a good help to the mods if you include whatever evidence you have though I'm nervous to point out examples of the types of things that would be useful. So far it's a relatively minor influence here, I suspect updates to flag reasons, etc. will primarily be motivated by the needs of StackOverflow which is dealing with quite a lot of these posts.
    – Bryan Krause Mod
    Dec 12, 2022 at 15:45
  • @BryanKrause, yes, I understand.
    – Buffy
    Dec 12, 2022 at 15:47
  • 2
    @BryanKrause, I found the following slashdot story interesting: What does ChatGpt say about itself?: slashdot.org/story/22/12/10/060229/…. In essence it says it is just a parrot.
    – Buffy
    Dec 12, 2022 at 15:51

You must log in to answer this question.

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