5
$\begingroup$

As a moderator (and user) of this platform, I increasingly recognize a disturbing development that makes me think about what actions should be taken in order not to reduce the credibility and quality of this platform.

In the recent past, there have been users on various SE sites (including here) who have posted answers using ChatGPT (or similar mechanisms). Some of the answers were helpful and even accepted, but nevertheless most of the answers were also seriously wrong and caused some confusion.

I myself also recognize more and more questions arising from erroneous answers from ChatGPT that are being asked here. Personally, I see the danger here that this platform will degenerate into a very difficult mix of genuine humanoid expertise and presumed knowledge, and thus in the long run no longer fulfill the mission it actually has.

The topic concerning AI is generally a bit more complicated, and therefore difficult to judge. The question that actually arises, however, is how one meets this trend as a platform, moderator and user, and how one wants to deal with it in the future.

In my view, this would require action to be taken very soon, rather than too late, to enable clear labeling of AI-generated content, so that visitors to this platform have the opportunity to correctly assess the content offered and its quality/truthfulness for themselves.

I therefore call for two things to be implemented here on this platform:

  • A clear labeling of AI-generated content
  • Posts marked as AI-generated content should not receive reputation points

This mechanism can be designed like the so far successfully used system for flagging a question/answer. Contributions can be flagged by software and community, and the community can accept or reject these suggestions. After a certain number of actions or by intervention of a moderator, these posts get an additional flag (similar to the green flag that marks an answer as accepted).


Since it's sometimes hard to make decisions regarding such a change, I just had a very nice conversation with someone who needs to know, and am posting the uncut version here for you to make your valued comments:

enter image description here I have a question regarding StackExchange. This is a Q&A platform where humans ask questions, and others answer questions. Can I ask you about this?
enter image description here Yes, you can ask me about StackExchange. As an AI language model, I can provide you with information and insights on this platform.
enter image description here Perfect! What problems arise from mixing answers written by humans and contributions generated by AI?
enter image description here There are several problems that can arise from mixing answers written by humans and contributions generated by AI. Here are a few examples:

Consistency: If answers are generated by different sources, it can be difficult to ensure consistency in tone, style, and format. This can make the content harder to read and less trustworthy.

Accuracy: AI may not always get things right, particularly in areas where human judgment is required. This could lead to incorrect or misleading information being spread.

Bias: AI can also perpetuate biases that exist in the data sets that it was trained on, leading to inaccurate or unfair information being produced.

Accountability: If there is a problem with the content, it can be difficult to determine who is responsible. This can make it harder to correct errors or hold someone accountable for any harm that is caused.

Creativity: AI is not always able to produce creative or original content. This can lead to answers that feel robotic or generic, rather than being tailored to the specific situation or question at hand.
enter image description here Should AI-generated contributions be allowed on knowledge platforms such as StackExchange, and thus be able to gain reputation, even though it used to be the case that only humans who answered questions satisfactorily could gain reputation points? Wouldn't that counteract the reputation points system?
enter image description here As an AI language model, I do not have personal beliefs or opinions, but I can provide an objective analysis of the issue.

Allowing AI-generated contributions on knowledge platforms such as StackExchange could potentially have both positive and negative effects. On one hand, AI-generated contributions may provide accurate and helpful answers to questions, potentially increasing the overall quality of the information available on the platform. Additionally, AI-generated contributions could potentially provide a solution to the problem of unanswered questions that may be too complex or specialized for human experts to answer.

On the other hand, allowing AI-generated contributions to gain reputation points could potentially undermine the existing reputation system, which is based on the recognition of human expertise and effort. If AI-generated contributions are allowed to gain reputation points, it could potentially lead to a situation where the quality and accuracy of the answers are no longer the primary factors influencing reputation on the platform. Additionally, it could potentially create a situation where AI-generated contributions are prioritized over human-generated content, further reducing the value of the platform for human participants.

Ultimately, the decision to allow AI-generated contributions on knowledge platforms such as StackExchange should be based on careful consideration of the potential benefits and drawbacks of such a system. One possible solution could be to allow AI-generated responses, but not award reputation points for them.
enter image description here Should posts generated with the help of AI be marked as AI-generated posts so that those seeking help can clearly distinguish between them?
enter image description here As an AI language model, I do not have personal opinions or beliefs. However, in my professional opinion, posts generated with the help of AI should be marked as AI-generated posts to provide transparency and clarity to the readers. By clearly identifying AI-generated posts, readers can trust the authenticity of the information presented. It also allows readers to distinguish between human-generated content and AI-generated content, which can help them evaluate the accuracy and relevance of the posts. Overall, marking AI-generated posts provides a more ethical and transparent approach to content creation.
enter image description here That's an interesting answer! Although you don't have a personal opinion on this, from a purely professional point of view, do you recommend that AI-generated posts be clearly marked as such, so that it is clear to someone seeking for help that it is AI-generated content that may contain gross errors and does not necessarily represent a solution?
enter image description here As an AI language model, I do not have personal opinions. However, from a professional point of view, it is recommended that AI-generated posts should be clearly labeled as such. This is important because it helps users understand that the content they are interacting with is not necessarily created by a human expert. Making it clear that the content is generated by AI can prevent potential misunderstandings and misinterpretations and help establish trust with users. Additionally, labeling AI-generated content can also help improve the transparency of machine learning systems, which is becoming increasingly important as these systems become more advanced and ubiquitous.
enter image description here But how can AI-generated posts be automatically distinguished from posts created by humans, so that the tagging of such posts can be done reliably and automatically?
enter image description here One approach to distinguish AI-generated posts from human-generated posts is to analyze the language style and patterns of the text. AI-generated content often lacks the nuances and creativity of natural human language and tends to follow a specific pattern or structure. Machine learning algorithms can be trained to recognize these patterns and identify if a post is likely to be generated by an AI. Additionally, metadata such as the source of the post or the frequency of posts can also provide signals for distinguishing between human and AI-generated content. It is also possible to use verification methods such as CAPTCHA tests to verify human-generated content.
enter image description here OK, but that answer sounds a little vague to me. Can you please elaborate a bit and make a concrete suggestion on how you can actively help detect AI-generated posts and mark them as such?
enter image description here As an AI language model, I do not have the ability to actively help detect AI-generated posts. However, social media platforms can implement tools like AI-powered content detection algorithms that analyze linguistic patterns and other characteristics of the post to detect if it is produced by AI or a human. These tools can then flag the post as potentially AI-generated, and either remove or add a label indicating that the post was produced by AI. Users can also be educated on how to identify AI-generated content so they can report it to the platform.
enter image description here If I understand this correctly, does it then require humans to integrate this function and thus enable AI-based recognition of contributions? I'm thinking of a system like the one used on StackExchange so far, where contributions are flagged, and these suggestions are then reviewed by the community and accepted or rejected. Would that be a solution?
enter image description here Yes, that is one possible solution. AI can be used to generate suggestions based on certain criteria or patterns, but ultimately it would require human review and input to ensure accuracy and fairness in recognizing contributions. A system like the one used on StackExchange, where community members review and flag contributions, can be effective in achieving this. Ultimately, a combination of AI and human review can help ensure a well-rounded recognition system for contributions.
enter image description here Well that sounds wonderful! This answer could be a reasonable way to handle AI-generated content on a human-powered Q&A platform! Thank you for your honest and helpful answers and have a wonderful day!
enter image description here You're welcome! I'm glad I could offer some insights. Don't hesitate to reach out if you have any more questions in the future. Have a great day too!
enter image description here Uh, wait, but I don't quite understand this answer. You don't have a personal opinion about anything, but you pretend to understand what "a great day" is? Now I'm confused. Could you explain in more detail how you arrive at the assumption that a day can be "great" at all?
enter image description here As an AI language model, I do not have personal opinions or feelings as I am not capable of emotions like humans. However, I have been programmed to understand and interpret the meaning of language, including idiomatic expressions such as "having a great day." When a user makes such a statement, I can process the linguistic cues and context to determine that they are expressing positivity or a sense of satisfaction or happiness. My response is based on understanding the commonly accepted meaning of the phrase "great day" and providing an appropriate response in line with that interpretation.

So, in conclusion, all that really remains to be said is: "Have a wonderful day and think about how your Q&A platform should be run in the future."

Your moderator.


Helpful: Is there a list of ChatGPT or other AI-related discussions and policies for our sites? Promising: OpenAI detector

$\endgroup$
4
  • 4
    $\begingroup$ @quellenform thanks for posting this and all the work and investigation that went into it. As a moderator it is my duty to flag this post as containing AI generated content. ;) $\endgroup$ May 13, 2023 at 21:21
  • 2
    $\begingroup$ Interesting. Thanks for starting the debate and articulating the problem. I would argue if we want to stay on top of things, we should also consider what we will do when AI-generated meshes, node-trees, and more generally assets will start coming up in troubleshooting questions (at the current pace, I can guess we"ll see that by the end of the current year). AI will most certainly make mistakes that a human would never do, and I think we need to decide how we deal with it before it dilutes knowledge and user engagement beyond repair. $\endgroup$
    – Gorgious
    May 15, 2023 at 6:52
  • 1
    $\begingroup$ Btw your conversation really goes to show how we can induce severe biases in the answers of GPT, I'm not sure it would have come up on its own with this system of tagging had you asked : "Is there something we can do to alleviate the problems of AI generated content on a stackexchange site" instead of "Is it a good idea to tag AI generated content with a special tag ?" and the general dichotomy of a chat engine that constantly tells you they don't feel any emotion or have any opinion, but is at the same time programmed to make you feel like you're talking to a real human being. :) $\endgroup$
    – Gorgious
    May 15, 2023 at 6:54
  • $\begingroup$ @Gorgious That's interesting! I just tried with other (less biased) words, but even then the answer is largely the same in content! In short: AI explicitly recommends labeling AI-generated content (as long as you value trustworthiness and consistent quality). $\endgroup$
    – quellenform Mod
    May 15, 2023 at 9:43

3 Answers 3

4
$\begingroup$

Thanks for the thoughtful question and thorough investigation, took me a while to process all this. This is all relatively new, so expect developments in the future, and that some of this may well become outdated with time.

Luckily we already have a few network guidelines to helps us outline a strategy for our site. As you stated, they have so far opted for not banning AI generated content globally at the moment.

Sadly though the burden of identification and filtering so far lies on the human side; ingenious community members have already built tools to help with that, but integration in the site is still lacking, so some manual work is still required.

There is also the issue of certainty when establishing whether some content was human generated or not, within an acceptable margin of error.

Somehow marking, tagging or signaling AI generated content, suspected or certain, seems like a correct thing to do.

As far as I know there is no builtin or native way to do this yet. We can flag questions with a limited subset of builtin moderation flags but these are mostly transitory and to mark something that requires action, not a long term categorization system. We can create arbitrary content tags, we could have a tag for content, but we can't tag answers yet, and answers would be most desirable candidate for flagging.

Until there is, we should probably defined a set of guidelines on how to handle this type of content on our community, and how to mark it, in the absence of a dedicated tagging system.

It has been suggested that answers that rely heavily on AI generated content should not earn reputation for the poster, which I agree but is harder to implement in practice, since that would probably require development from the platform maintainers (which may or may not be happening as we speak).

That would at least require a system for users to flag content as generated by AI, a system for reviewing such flags by other users, so they can verify the claims; and if confirmed, a system for both visually tagging such posts (to inform potential readers), and internally prevent the poster from earning reputation.

Site guidelines already require attribution for copied content, most argue that AI generated content is neither the original work of the poster nor the programmers that created the language model, so it should be no different for posts where the majority of the content was not typed by a human.

Some suggest that AI generated content should be block quoted and correctly attributed, as in:

Generated by ChatGPT Should posts generated with the help of AI be marked as AI-generated content?

As an AI language model, I do not have personal opinions. However, from a professional point of view, it is recommended that AI-generated posts should be clearly labeled as such. This is important because it helps users understand that the content they are interacting with is not necessarily created by a human expert. Making it clear that the content is generated by AI can prevent potential misunderstandings and misinterpretations and help establish trust with users. Additionally, labeling AI-generated content can also help improve the transparency of machine learning systems, which is becoming increasingly important as these systems become more advanced and ubiquitous.

In practice it will probably be hard to enforce this policy, users who are inclined to post AI generated content are probably the most likely to be unaware of the rules, ignore them, or try to sneak around them for reputation gains. Compliance will most likely require the help of the community, for flagging and reviewing potential infractions, and if confirmed withing a certain margin of certainty, edit said content to make it clear it was potentially generated by AI.

For now we'd probably have to rely on the builtin flag system and ask users to flag such posts with:

In need of moderator intervention

A problem not listed above that requires action by a moderator. Be specific and detailed!

and ask users to type something like Potential AI generated content.

$\endgroup$
2
$\begingroup$

Yes, posts should be marked, but I don't think a help or contribution of ai in some way has to be something bad especially if reviewed by author of such post.

I can understand the danger of weird scripts generated by ai or any unchecked structured solutions, but from experience (at time before ai) there were also posts with a lot of nonsenses or some dysfunctions in their process ... and only followers that read it or tried it and failed could discover issued posts ... so they could be flagged or downvoted than.

So how this situation with ai differs?

Or if like now I posted ai generated script - how does it differs from time before where I would grab some parts of existing scripts or templates and try to merge into some functional script? It will work in some way for me, but as for a generated ai script I wouldn't be sure how bad or good the script is. How does it differs from coder beginner?

Only voting or flagging system should be enough to separate gold from rubbish. (... as experienced contributor on this site I know it is not quite true as mentioned by Markus 0,1,2..., but that is how it works here.)

SE already use flagging system, so seems to me native to expand this list of flags. It seems to me as warning as any other.

Author of this post could be asked to extend his post about an explanation how ai contributes in his answer so followers can decides for an amount of a risk.

About usefulness/uselessness of the answer (question) it self should decide existing voting/downvoting system.

Current warning feels like "run a way" even before you give it a chance to check if the answer is good or not.

Also +1 for "ai-generated" tag.

Note: My current state of point, that can be re-evaluated in time for sure :)

$\endgroup$
4
  • $\begingroup$ We had a TV show in Poland called A fault- a candid camera with a professional asked to do a small job, revealing how many experts are actually quite unprofessional, be it due to incompetence or dishonesty... One controversial episode was about IT - fixing a computer. The fault here was a cable inside a PC case, connected to a wrong place. The controversy was: it's not a problem that 'naturally' arises, ever. It was an 'artificial' problem, that could only arise if someone intentionally made a bad connection. $\endgroup$ Jan 29 at 11:19
  • $\begingroup$ It felt unfair to portray experts being unable to "repair" the computer, if virtually no amount of experience could teach one to deal with such a scenario. Imagine that you use Python to set "Instance on Points: Selection" to False - it would be hard for us to figure this out, right? Sure, we could recreate the node tree from scratch and it would work, but an IT technician can't possibly suggest replacing the entire PC case. Why do I write about this? Because the bugs introduced by AI are not the same bugs that appear in code "naturally", which is a huge problem... $\endgroup$ Jan 29 at 11:20
  • $\begingroup$ Great topic, but I'm not sure if this is the best comparison ... I see everyday newbees asking for help exactly for this reason - using your simile -they plugged cable to a wrong place. It is unexpected action done in unexpected way ... but it makes any error to be error. ... I can agree we are faced to an another kind of errors that we have to lear to recognise (or be more sensitive), but thats all for me. More dangerous is fact that generate sophisticated detaily explained nonsense take AI a seconds and takes someone seconds to post, but takes much longer to revise it by human :) $\endgroup$
    – vklidu
    Jan 29 at 19:59
  • $\begingroup$ You don't see newbies coming with the kind of a Blender problem I describe :) Just like with cables some kind of problems simply never arise. Yes, this is a new kind of an error, but I'd like to have the right to subscribe to this kind of an error or not - if I see chatGPT generated code, I just don't read it. There's another reason: our brains work in complicated ways: you use a word usually not because you memorized it from a dictionary, but because you've seen it in many contexts. If you read hallucinated code, you poison your intution. Hard to explain in a comment section... $\endgroup$ Jan 29 at 21:43
1
$\begingroup$

My thoughts on the matter:

  • As linked in the OP, Stack Overflow outright bans GPT content.
  • BSE is not Stack Overflow; I don't think there's a rule SO policies apply unless explicitly overridden.
  • The aforementioned policy I find too defensive: GPT is praised by a lot of people, and censoring it might create an aura of conspiracy.
  • GPT is indeed a marvelous technology; however, BSE is an even more marvelous place, with content of ridiculously high quality - AI generated content can't possibly compete and so inevitably dilutes the value.
  • Pasting AI generated content and not marking it as such is cheating - it's like writing Python code and saying you tested it, even though you didn't; the potential annoyance caused is not worth any reward the SE system has in store.
  • Some users of this site might be unaware of the problems related to GPT: perhaps AI-generated content should be wrapped in a warning similarly to tobacco warning labels:

😥🔫🤖 Be careful! The below content is AI-generated

GPT outputs tokens with similar statistical relations to training data; in practice this means text that seems valid, seems confident, seems logically consistent etc. While humans write text to be correct → and therefore it seems correct, AI writes text to seems correct → and therefore it often is correct. More reading: What Is ChatGPT Doing … and Why Does It Work?

3 chickens lay 3 eggs in 3 days, therefore 12 chickens lay 12 eggs in 12 days.

This marks the end of AI-generated content. Below human-generated content 👨‍🏫

  • Since I advise against censoring AI, it can be fought with downvotes; those however are discouraged with a reputation cost to downvote an answer.
  • If downvoting is discouraged, then the solution becomes not-upvoting, which unfortunately also doesn't work in a situation, where a lot of valid answers have 0, 1 or 2 upvotes…
  • Detecting AI is unreliable: technically speaking, a human can answer exactly as AI would, as well as AI is programmed to answer as a human would.
  • Any effort to discriminate between AI and humans is too invasive: we could enforce a requirement to be politically incorrect for example.
  • Even if AI content is detected, it can't be fairly punished: it's easy to defend by claiming a human proxy, which passed the AI text without disclosing it's AI this way either innocent people get punished, or offenders can easily get away.
  • Only just now from D. F. R.'s post I learned about the necessity to reference - sounds reasonable and I probably never quoted someone without an attribution, but recently an established user was caught pasting GPT text "from a friend" 😄 and no punishment was served, be it a flag, downvotes or a reprimand. And perhaps rightly so, with such incidents being rare.
  • What if someone uses AI to translate text? That's how most translators work like DeepL, but ChatGPT can also be used to translate text. It can be also used to correct grammar, or otherwise modify the form of the text without changing its substance. Grammarly is also AI based.
  • I would therefore argue lack of attribution to AI is not the real problem here. The real issue is the author of an answer should put some effort to verify the information provided is correct, and be transparent about which parts still require confirmation.
$\endgroup$
3
  • $\begingroup$ In principle, I agree with you on some points, but I think a distinction must be made between "posts generated with the help of AI and revised by humans" and "posts generated by AI that are NOT edited and reviewed by humans". I think that is the key difference. And in the second case, mechanisms for early detection can be implemented, which could then be checked by the community in the reviews. The essence would be to mark such contributions that are clearly unchecked AI crap. $\endgroup$
    – quellenform Mod
    May 15, 2023 at 9:48
  • 1
    $\begingroup$ @quellenform I agree, except I'm skeptical one can be sure nonsense wasn't generated by a human… After all, we had plenty of that in the pre-AI era. $\endgroup$ May 15, 2023 at 9:50
  • 2
    $\begingroup$ Yes, there were and always will be those contributors :D ... but in such cases I honestly trust the competent assessments of enthusiastic community members like you! $\endgroup$
    – quellenform Mod
    May 15, 2023 at 9:58

You must log in to answer this question.

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