Policy: Generative AI (e.g., ChatGPT) is banned

https://meta.stackoverflow.com/questions/421831/policy-generative-ai-e-g-chatgpt-is-banned

by RyeCombinator • 2 years ago

22 12 2 years ago

meta.stackoverflow.com

How will they enforce this? There's no easy test if an LLM write a given post. Perhaps they can post honeypot questions.

RobRivera3• 2 years ago

If you knew how to detect AI posts, why would you tell people?

maeil2• 2 years ago

The following points are easy to verify:

1. It's relatively easy to detect "default ChatGPT" output with very few false positives, and I've written a crude version of one myself.

2. It's doable to make LLMs generate output that are impossible to write such a detector for.

I've made a service[1] that does the second part, rewriting content to fit in with a certain community, one of which is StackOverflow.

[1] https://soundnatty.com/

esmy• 2 years ago

I see why they would want to try and enforce this, but it isn't something they'll succeed at.

Especially questions that have one correct code solution. Unless proven otherwise, everything is now written by an AI, the targh is out of the bag.

RobRivera3• 2 years ago

Time to roll over and let stackoverflow die, I suppose

squigz• 2 years ago

Asked 1 year, 9 months ago

karmakaze• 2 years ago

The interesting thought experiment is what will likely happen one way or another depending on the policy they choose/enforce (if possible).

1. They don't ban AI posts => probably a lot of low-effort (often correct) postings; reputation farming because why not; people trusting SO less; may as well skip the middle-man and simply ask ChatGPT directly

2. ban AI posts => how do they enforce this; conflict/complaints/annoyance when mis-flagged; reduction in answer submissions; plenty of generated (mostly correct) answers getting through; reputation farming; quality could still decline until people stop using SO (but maybe not as soon as 1).

3. put more effort into allowing correct answers (regardless of how written), penalize incorrect answers which appear generated => maintain higher quality bar; can sustain for a while until ChatGPT et al is simply just better to ask directly (but not as soon as 1 or 2).

I used up my free credits using GPT-4o the other day asking random stuff that I might have googled or Wikipedia searched before--it was more efficient and engaging. I could come around to paying $20/mo as I'd get way more value than I get from Netflix and a few other services.

A lot of sites are going to have trouble competing with AI so they may as well have the best AI for their domains to stay relevant. Humans will still be providing original content but most of it may be indirectly consumed by others. A lot of content wasn't accurate anyway so there will simply be more pushing an agenda or just selling stuff. Hopefully we will value quality and our preferred truths and there will be some competition to deliver that.

weq• 2 years ago

lol peak AI only exists because of humans answering questions.

SO is positioning itself with a clean dataset, likely with an agreement to sell that data to other compnanies?

throwaway81523• 2 years ago

Lol, Stackoverflow cashes in selling LLM training data to AI companies[1], then finds the companies don't want to pay for recycled LLM spew. Need... fresh... braaaainzzzz!

[1] https://www.wired.com/story/stack-overflow-will-charge-ai-gi...

CaptainFever• 2 years ago

(2022)? The only thing changed is that it's now permanent.

maeil2• 2 years ago

This is pointless in the long run. For now, there's still plenty of people who just simply ask ChatGPT a question straight up and copy-paste the output. In such cases, it's doable to catch a lot of them with very few false positives.

But at some point this will become common knowledge, and then the norm will be using LLM services that generate content that is not recognizable as AI-generated, rendering this policy as useful as one that requires posters to wear red socks while posting.

A trend I've been noticing is that there's people who actually have something useful to add or ask, but for some reason, possibly being scared of downvotes, still use GPT to rewrite their stuff. Maybe the community is to intimidating, maybe they don't know how to phrase stuff properly, but I've seen it multiple times.

This inspired me to develop SoundNatty[1] that shows the long-term fruitlessness of policies like these, by rewriting content to sound natural to a target audience. I'm not offering an API because I hate bot spam as much as anyone else, but honestly it's only a matter of time before someone else does.

[1] https://soundnatty.com

Policy: Generative AI (e.g., ChatGPT) is banned

Related Stories