I think that an outright AI ban is the wrong way to go. A good answer is good and a bad answer is bad, regardless of if they were written by AI. Instead, I think that we'd be better off specifically targeting the issues of answers written by AI.
Problem 1: Incorrect/Nonsensical answers
Using AI generators will often produce code that doesn't compile (or compiles only with a large amount of errors) and will often suggest packages/sources that don't exist.
Any AI answers that contain TeX code should also have to include an image of the output. This will show that (1) A human was involved in creating the answer; and (2) that the code compiles and gives the correct output. Most good answers from a human include screenshots of the PDF anyways, so this isn't much of a hardship.
Any AI answers that recommend a package/resource should be required to include a (working) hyperlink to either the CTAN page for that package or the website of the source. Again, any good answers from a human would do this anyways.
Problem 2: Plagiarism
AI generators will often print out parts of its training material verbatim. The entirety of TeX.SE is available in a bulk archive, so most AI generators would have been trained on it. Even when an AI prints out a novel solution, it is the work of the AI, not the answerer.
Answers from AI generators should be clearly attributed to the AI. Placing the response from the AI in
> blockquotes is best, but as long as the answer contains a sentence like:
Here's what ChatGPT has to say:
then I don't think that there's a problem.
Answers that are copied verbatim from another answer without acknowledgement should be flagged, just like if a human did the same.
If you wrote more than, say, 80% of the answer yourself and just used an AI for minor help, then I don't think that it needs to be mentioned.
Problem 3: Correct but confusing/overly complicated answers
This is a problem with human answers too. We have human answers doing things like using
\\ instead of
\par, using generators for TikZ that give unreadable code,
The solution here is downvoting, not flagging.
Problem 4: Users can just use the AI generator themselves
Not all users are aware that these AI generators can answer TeX questions. And as we've seen, they often give incorrect code that still needs some editing from someone knowledgeable in TeX. And most AI generators require either payment or a cell phone number, which isn't something that all users have.
Using similar reasoning, you could also say "users can just read the package documentation to find the answers". Indeed, that is correct for many questions, but no one is proposing to ban those questions.
Official policy is to embrace questions/answers that can be easily Googled so long as they're not duplicates of other questions on the site; I see no reason for our AI policy to be any different.
Problem 5: Users can post answers too quickly
If the answers are bad, then they will be downvoted/flagged and the system will prevent them from posting more. If they're good, then what's the problem?
And besides, anyone writing terrible answers using AI too quickly probably won't be stopped by an "AI ban".
Problem 6: Potentially harmful answers
A bad AI-generated answer could lead to real-world harm on sites like Chemistry, Electronics, Law, Travel, or Medicine. Bad AI-generated answers on Bitcoin, Cryptography, Stack Overflow, Super User, and Unix & Linux could break or cause serious issues to a computer. It may make sense for these sites to ban AI because of the serious effects that a subtly-incorrect answer could have.
However, on TeX, the worst that a bad AI-generated answer could do is causing an infinite loop, and that is easily fixed with
Ctrl-C. Even that is quite unlikely; most bad AI-generated answers just won't compile.
Good answers are good, and bad answers are bad regardless of how they were created
Copy-and-pasting answers directly from an AI generator without any human modifications should be forbidden
AI-generated answers should always clearly indicate that they were made by AI
AI-generated code answers must be compilable. To prove this, they should include an image of the PDF output by TeX.
AI-generated answers mentioning a package or other resource should include a functioning hyperlink to CTAN or another site.