6

TL;DR

  • Downvote this question, if in favor of allowing AI-generated content
  • Upvote this question, if in favor of banning AI-generated content

The problem

As you may have noticed ChatGPT, a language model by OpenAI, has been made available. So it's AI-driven and you can ask it questions about all sorts of topics and it will give you answers, sometimes honoring and sometimes ignoring the constraints you gave it in your question.

For all the good and intriguing utility this can bring, it has been noticed that suddenly users started answering questions with content that was either clearly or supposedly AI-generated.

Now, what could be wrong about sharing answers from ChatGPT?

  1. typically ChatGPT isn't being attributed
  2. answers are often vague and for topics that have a lot of depth -- such as on our site -- can even be outright misleading

We could probably come up with variations of the above or even new items.

The point is that ChatGPT is (ab)used as an effortless way to answer questions -- but not attribute the "author" -- all the while subverting the gamification aspect of StackExchange, which also means that in the long run it may take away the fun for many of us.

Policy or not? Formalizing our stance ...

"Temporary policy: ChatGPT is banned" is the stance in the StackOverflow community. While all part of StackExchange, we have to make up our own minds about whether to adopt and formalize this very same policy for RE.SE or choose another path.

I think our stance should be formalized and published, such that we can refer to the site help regarding such issues.

How to go about this?

  1. Downvote this question if you are in favor of allowing AI-generated content
  2. Upvote this question if you are in favor of adopting the policy from StackOverflow
  3. Comment on the question for clarifications that may have to be added to the question itself
  4. Write an answer, describing (or detailing) your view(no matter if you voted on this question or not)
    • NB: Answers should and will be upvoted or downvoted to "voice" agreement or disagreement with the views shared in them.

Use option 4 if you lean towards option 2 but wouldn't want to outright adopt the policy from StackOverflow. In this case, please detail what exactly you would want to have differently from their policy.

PS: ChatGPT is just the current fad, so I'd like our discussion to be regarding AI-generated content -- arguably this mainly affects answers -- in general and not only ChatGPT specifically.

3
  • Just stumbled upon this meta post. This is tangential to the topic of ChatGPT, but just a reminder that due to the asymmetrical reputation requirement for upvote (15) and downvote (125), some users can only upvote but not downvote. (That aside, I don't vote to prevent skewing the number since I'm not a regular here :)
    – Andrew T.
    Commented Jan 26, 2023 at 18:59
  • @AndrewT. regular or not doesn't really matter. Many intermittent contributors have contributed a lot of valuable Q&As. But yeah you are right and I wonder if there'd be a way to create a poll that sidesteps this requirement. Perhaps people before us have discussed this and I'm simply not aware of that discussion ...
    – 0xC0000022L Mod
    Commented Jan 27, 2023 at 8:27
  • (*Stack Overflow. *Stack Exchange (the last section).) Commented Apr 26, 2023 at 10:53

1 Answer 1

1

While not generally opposed to the notion of seeing AI-generated on our site or StackExchange in general, I think we should follow the example of StackOverflow and also ban AI-generated answers for the time being.

My main reason is that if no effort or own knowledge or experience went into the creation of an answer, it shouldn't be rewarded at the same rate as those of human posters. The fact that most of the time the AI isn't being attributed doesn't exactly make it any better either.

That said, it's not always easy to detect AI-generated answers or ChatGPT answers specifically (see below for one puzzle piece).

But I am convinced you will develop a spidey-sense that lets you assess this without a tool to help. But maybe you can use the tool to validate (or not) your initial suspicion.

Future watermarking may also be helpful in detection. But this, again, would be ChatGPT-specific.


Glorfindel, who is also member of our community, has written a user script which community members and moderators can employ to gauge whether a question could be AI-generated. This script aims squarely at ChatGPT, though.

Installing the user script via TamperMonkey or another extension of your choice, you will see the "Detect OpenAI" link below each question and answer:

Detect OpenAI link shown on an answer

Once you click it, you will get an assessment such as this:

Detect OpenAI popup with high probability

... showing a probability value for the question/answer being generated via ChatGPT/OpenAI. Obviously we also have questions/answers with a marginal probability shown:

Detect OpenAI popup with low probability

After you used the link, the link will also change to show the probability value:

Detect OpenAI link with probablity

2
  • 2
    In addition to the objection regarding effort (the lack thereof), I'm skeptical of the quality and correctness of the texts produced by such programs. I've seen numerous examples in different domains where there are mathematical or logic errors (mistakes in sets of propositions, logically unsound arguments, etc.). This is to be expected from model-based text generation tools that do not possess reasoning faculties. If we permit low-quality mistake-riddled texts as answers we will quickly find that the utility of the site will drop, undermining its reason for being in the first place.
    – julian Mod
    Commented Jan 20, 2023 at 15:25
  • Hehe, yep I've seen that to. But the answers are usually quite convincing at first glance, right? Unless it outright violates some of the given constraints it sometimes can take a bit until you uncover that this AI is really really really good at making a response sound legit. But at least the policy for technically incorrect answers on SE is usually to downvote them, not to ban or delete them.
    – 0xC0000022L Mod
    Commented Jan 20, 2023 at 21:29

You must log in to answer this question.

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