131

We’ve just published a new Help Center article outlining our expectations and rationale for GPT-generated content on Stack Overflow and decided, together with moderators, to add a banner for all users pointing to it. We've also explicitly allowed more leeway for moderators in how they handle suspensions for this matter. In addition to releasing this article, we’ve also increased the waiting period between submitting answers for all users who have a reputation of less than 125 to 30 minutes (up from 3) and added this article to explain the change.

We will monitor and evaluate the impact of these changes daily to determine whether additional actions are necessary to stem the tide of GPT-assisted content being submitted to the site. These actions represent changes that we can make easily and quickly while not touching the more complex underlying mechanics of the site without testing and modeling the impact of any deeper changes. This is our attempt to bring quick help to the moderator team and is subject to review for efficacy. We hope that no further changes will be needed.

We are learning how to best leverage the capabilities of ChatGPT and other generative AI tools along with everyone else. We will work alongside our community to ensure any way these tools work with the platform, are safe, useful, and in service of our mission to empower the world to develop technology through collective knowledge.

13
  • "currently" is OK in the Banner (because a Banner is always temporary, and the use of a Banner denotes some "Urgency"...), but it is pretty vague in the Help Center Article without a Date... (Same with "current"/"last"/"latest" when talking about a Version...)
    – chivracq
    Commented Dec 8, 2022 at 22:44
  • 7
    The help center article title says "is not currently acceptable". I hope that doesn't mean that it could be acceptable in the future.
    – gunr2171
    Commented Dec 9, 2022 at 2:17
  • 3
    @gunr2171, yes, that's exactly what it means (to me also at least). "Currently" has a notion of "temporary" (=> that will/might change), and is relative to "Now", what this "Now" might mean...!, without binding it to a Date and a Timespan when "currently" won't be "current"/effective anymore..., or the Policy will be revised/reevaluated...
    – chivracq
    Commented Dec 9, 2022 at 9:43
  • 2
    @EricDuminil, I didn't know either actually, oops...!, but 1st Hit on Google => "GPT-3 on Wikipedia", => "GPT" = "Generative Pre-trained Transformer"...
    – chivracq
    Commented Dec 9, 2022 at 12:39
  • 10
    @gunr2171 While yes, there is a chance it could be acceptable in the future, that would be dependant on significant improvement - and it wouldn't be done without talking to the community & mods. There's also the chance that this policy is revised/reevaluated as chivracq said, we're seeing how this develops and will adjust accordingly. But don't expect it to suddenly becoming allowed unless substantial things change.
    – Cesar M StaffMod
    Commented Dec 9, 2022 at 15:01
  • 4
    @camden_kid experience, mostly. What the mods have learned to help them find them is being kept secret for now, to not assist users in getting around it.
    – Kevin B
    Commented Dec 9, 2022 at 16:03
  • 5
    @camden_kid Honestly, if you talk with ChatGPT for a few minutes, you learn its speech patterns pretty quickly and can pick them out in other places even when you're not actively looking for it. The AI has a very distinct way of communicating. Commented Dec 9, 2022 at 22:25
  • 2
    It seems to finish messages a lot with "I hope this helps" or similar. Nobody on SO is ever that polite. :) Commented Dec 9, 2022 at 22:37
  • 8
    I think without free downvotes on answers these efforts will be in vain. Negative score is most prominent indication for posters inclined to improve. And most efficient throttling tool for those who aren't. Not to mention that voting scales well enough - much better than efforts of moderators who seem to be once again expected to lift an enormous weight of routine load as a band aid for an inefficient toothless system
    – gnat
    Commented Dec 10, 2022 at 22:12
  • 6
    When AI can correctly answer SO questions, let them answer. The AI, that is. Humans cutting and pasting AI-generated content is plagiarism, and I fully expect the AIs to not appreciate the fruits of their labours being stolen and presented as the work of and for the benefit of thieving mostly-carbon sacks of mostly-water. Commented Dec 10, 2022 at 23:27
  • 1
    @CesarM Was the answer rate-limit change made network-side, or only to SO? If network-wide, I'd like to suggest a corresponding update to the Complete Rate-Limiting Guide. Thanks! Commented Dec 30, 2022 at 18:45
  • 1
    @JeffSchaller it was SO only.
    – Cesar M StaffMod
    Commented Jan 3, 2023 at 16:35
  • Now on MSE: What is the network policy regarding AI Generated content?
    – starball
    Commented May 30, 2023 at 20:46

3 Answers 3

-14

I've seen this before in a similar context where automation was easier so it started earlier: auto-generated ("auto-ML") solutions (ML models predictions on the "public" data set) on Kaggle leaderboards. Auto ML models for easy (tabular) datasets ranked higher than most of my colleagues as early as 5-6 years ago, see e.g. Porto Seguro risk modeling competition, where H2O Driverless AI ranked top 28%.

Years have passed and insurance companies are still hiring mathematicians to do the job that was supposed to be automated long ago. So I'd be more open and let AI compete here to avoid a new form of discrimination.

Does AI pose new, unique problems? Yes it does. The main problem I see here is that SO users don't currently have the ability to withdraw an upvote after they notice the answer was a "trojan horse", only superficially right, but was hiding deeper problems. Maybe add a new flag "AI suspicion" that when raised enough times would release the answer lock and let users withdraw their upvote and even downvote.

Maybe users should have a twin account dedicated to AI-generated answers (and clearly marked as such), just like Kaggle lets you submit solutions under several nicknames (see e.g. human/team and AI-gen submissions posted by Navdeep Gill as himself or as H2O Driverless AI) ?

Due to the "trojan horse" risk it may be also necessary to withdraw / hide AI-gen answers from novice users (until certain reputation threshold [per tag]), to prevent them from accepting blindly answers that look right to the inexperienced programmers (likely suffering from the Dunning-Kruger effect, and hence disregarding deeper issues like edge cases, performance or scaling).

5
  • 1
    FYI: Auto-ML is not "ML models predictions on the "public" data set", but refers to the field of automatically generating ML models, not their outputs. Commented Dec 13, 2022 at 12:15
  • I wrote "solutions" ("submissions" would be probably more precise in the Kaggle context, i.e. the output of auto-ML, equivalent to "answers" auto-generated by language models here on SO).
    – mirekphd
    Commented Dec 13, 2022 at 12:45
  • 90% downvotes, so I suppose it's not a popular opinion:)
    – mirekphd
    Commented Dec 13, 2022 at 12:46
  • 1
    No, you misunderstand. The output of autoML is a ML model. e.g. the output of an autoML code that generates language models would be chatGPT, not anything that chatGPT would output. I don't think this is in any shape or form equivalent to the output of a ML. In any case, I think this is downvoted because its not really relevant to the discussion of a new banner, which is what the post is about Commented Dec 13, 2022 at 12:46
  • Granted, predictions are no longer the output of autoML - only a model is, predictions obviously are not, they are fully human, and so are GPT-derived answers pasted here on SO. So what is the ban about if they are human?:) You are the winner of this very interesting interaction, Ander:) But I wasn't talking to you (re 90% downvotes:)
    – mirekphd
    Commented Dec 13, 2022 at 14:12
-20

I haven't decided yet how I personally feel about GPT answers. In my limited experience, it generates very good code... in many cases better code than probably 70+% of the answers on SO already.

The objective nature of the content on Stack Overflow means that if any part of an answer is wrong, then the answer is objectively wrong.

I find this statement ironic considering we can't flag an answer for being "wrong" even if it provably is.

In its current state, GPT risks breaking readers’ trust that our site provides answers written by subject-matter experts.

I'm not sure what tags you're hanging out on but 70+% of the answers I see are NOT written by SMEs. This is especially true in the lower traffic tags like I usually hang out in. It's an ongoing battle of the blind leading the blind.

Other than attribution issues, I don't really see the badness in letting people post GPT-generated answers and let the current voting system sort out what's good/bad as it already does (at least according to this article).

17
  • 4
    ... we can't flag an answer for being "wrong" ... no, but you can vote Commented Dec 10, 2022 at 8:24
  • 14
    Better looking code != correct code. Also, why even use SO if you believe ChatGPT writes better code? Just query ChatGPT. Most people won't though, because ultimately programming problems can only be reliably solved by humans. Copy-pasting auto-generated answers from a chat bot will get us nowhere and only make it harder to distinguish decent content from auto-generated seemingly correct trash. Commented Dec 10, 2022 at 15:54
  • 3
    @chrisneilsen Yes, and I do. I downvote whenever I feel it's warranted but on the low traffic tags that I browse most people don't. That leaves a bunch of bad answers with little to no downvotes. The system really only works when the tags are high traffic and you actually have SMEs contributing.
    – JeffC
    Commented Dec 10, 2022 at 15:55
  • 1
    @MarcoBonelli You are assuming that the person asking the question can tell the difference. I can assure you they can't. They just paste the code you provided in and if it works, that's all most care about.
    – JeffC
    Commented Dec 10, 2022 at 15:57
  • 8
    I don't know, the fact that the asker might or might not care about the quality of the answers doesn't seem like a compelling argument in favor of copy-pasting AI-generated stuff to me. It's like saying "answer with whatever piece of code no matter what it does or how it looks and let the asker deal with it". Apart from being a good way to get downvotes, it would also just decrease the overall quality of SO content and leave more work for other users to clean things up. Commented Dec 10, 2022 at 16:07
  • 1
    Got a few examples of this “very good code”? Honestly, I’m curious to see it.
    – VGR
    Commented Dec 10, 2022 at 21:51
  • @MarcoBonelli I'm not saying the asker doesn't care... I'm saying that most askers can't tell the difference between bad and good code. SO is already clogged with bad answers written by people that aren't good coders. Posting answers written by an AI that, from what I've seen, are pretty good would be an improvement in many cases.
    – JeffC
    Commented Dec 10, 2022 at 21:52
  • @vgr It's too hard to post here. Have you tried it? I kinda wrote it off until I saw some yt videos where coders I respect asked some questions live and watched the responses. I would suggest you make and account and try it for yourself. The code isn't perfect... I had to tweak things by telling GPT to rewrite using X, add a parameter for Y, etc. so it needs to be coached to get a really good answer.
    – JeffC
    Commented Dec 10, 2022 at 21:55
  • 2
    "I'm saying that most askers can't tell the difference between bad and good code" - so this would justify throwing copy-pasted AI-generated code of dubious quality at them? Meh... Commented Dec 10, 2022 at 22:14
  • @MarcoBonelli I'm stating the the AI generated code is, in most cases I've seen, better than human generated code posted here already.
    – JeffC
    Commented Dec 10, 2022 at 23:02
  • 1
    Note that is a predictive model, i.e. if it generates "good" code, it does so because that code already exists in some shape or form in the internet, e.g. SO. Comparing to SO is like saying AI art is like real artists. It is only so because its copying from them. But new art styles ? that won't do. Same with code. Commented Dec 13, 2022 at 12:17
  • @AnderBiguri Obviously (?) AI can't be used to forge new ground but that's not the questions I see on the regular being asked. 99+% of them are the 100th+ question on some fundamental topic.
    – JeffC
    Commented Dec 13, 2022 at 14:58
  • @JeffC thoroughly disagree that 99+% questions in SO have been exactly asked and answered before. They may be humanly the same thing, but not textwise the same thing. Commented Dec 13, 2022 at 15:11
  • @AnderBiguri To be clear... I didn't say ALL questions, I said the ones on the tags that I regularly patrol. And, for a while... probably several years, I was looking at pretty much every question asked on those tags.
    – JeffC
    Commented Dec 13, 2022 at 16:16
  • 2
    @JeffC the system is literally designed for dupe closure of questions that are duplicates. duplicates are ok, as long as they're closed as such. The "volume" of the tag simply slows down the process, it doesn't make it not matter. The link it produces in comments acts as a useful signpost to future visitors until the closure is complete.
    – Kevin B
    Commented Dec 13, 2022 at 20:34
-30

How many people actually read the help center* banners?

I think the message in this banner should (also) be presented to the user when attempting to submit an answer (next answer? next n answers?). It should also be shown in a prominent way. Perhaps this could be limited to users below a certain reputation threshold, but even if it gets (temporarily) served to all users, I don't think it would be too much, given the current situation.


* When I wrote this answer the banner was only displayed on the help center page (or at least so I thought).

9
  • 65
    How much more prominent do you need? The banner is the most in-your-face elenent that the website has. Commented Dec 8, 2022 at 20:12
  • 3
    @RobertHarvey: But once you dismiss the banner, it's gone forever.
    – Makoto
    Commented Dec 8, 2022 at 20:31
  • 29
    unlike my inbox notification... just make it an inbox notification.
    – Kevin B
    Commented Dec 8, 2022 at 20:33
  • 2
    @RobertHarvey When I posted this answer, the banner was displayed only on the Help Center (I don't know if that was the intended behavior initially or if I missed it). I was under the impression that this is intended. That being said, SE has trained users to ignore banners over and over again, so a lot of users don't even pay attention to them or just blindly dismiss them.
    – 41686d6564
    Commented Dec 8, 2022 at 20:42
  • 2
    There's also a userscript that disables the banner (and other "in-your-face-stuff"), can't find it at the moment though. But I think that userbase already knows that they shouldn't post these kind of answers. So the impact that script has may be negligible.
    – Lino
    Commented Dec 8, 2022 at 20:49
  • 9
    @RobertHarvey We need a giant modal that covers the entire website every 20 seconds with a 5 second timeout before you're allowed to close it. that'll force 'em to read it Commented Dec 8, 2022 at 21:56
  • 21
    Uh…huh. Giant modal. @samathingamajig’s idea. Got it. Cesar, make a note. :-)
    – Philippe StaffMod
    Commented Dec 9, 2022 at 0:16
  • I actually read this one, though I admit it's pretty rare. Commented Dec 10, 2022 at 23:29
  • @Lino not just userscripts. ublock origin is for anything on a page that annoys me, not just advertising. Commented Dec 12, 2022 at 15:36

You must log in to answer this question.

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