AI disclosure and policy?

[corpserot]

1. ContentDB probably wants a bit of transparency on AI in the near future. at least, that's the hint Rubenwardy gave on Luanti discord #contentdb.

2. There will likely be people that delegates part of their work to generative AI when making contributions (frog already did prove that AI was involved in SBZ). Contributors need to at least say that AI was involved somewhere in their work.

3. Gauge people's reaction to AI gen stuff. This is different compared to analytical AI uses.

This is quite low priority as cdb hasn't updated to proposed new AI disclosure system yet. there's also no AI-heavy contributions yet (at least, that's what i'm aware of).

[TheEt1234]

i think using LLMs to help with ideas/algorithms/math/whatever is fine, just not with writing the code and copy pasting it

[TheEt1234]

and i guess no AI generated media

i think that would be good AI policy

[ChefZander]

and i guess no AI generated media

i think that would be good AI policy

I agree. Skyblock Zero will never have any AI art.

[TheEt1234]

@ChefZander whats your opinion on this

[TheEt1234]

It would be good to write an AI policy somewhere
What i don't want is sbz to be required to be tagged as "AI Generated" or something

[corpserot]

Generative AI use is about well, generating stuff as part of your contribution/PR like code, documentation, tests, images, models and sounds. Some people use it for auto-completion or make boilerplate code. Some people use it to shortcut the initial work and then refine its generated output by hand afterwards.

• code: promising usecases and is already something people use in luanti (incl engine source code). I think we should allow this and have the contributor transparently declare it all. (btw copy pasting code out of LLM outputs is pretty much outdated practice lol).
• documentation: i use AI to set up the initial documentation then refine it afterwards on some projects. Mostly for RAG tho, not for people. It's not a stretch that someone wants to use it generatively as part of their documentation process.
• tests: personally i find LLMs to be not creative (which makes sense considering how it works and is trained) and writing complete tests needs some imagination. However i do find it useful for boilerplate/common tests stuff.
• images: no. Diffusion SUCKS for pixel art.
• sounds: i believe currently, models are leaning towards generating music as an entertaining activity instead of use as asset maker for games and stuff. No.
• models: probably no? I haven't sewn much news about this.

As for the tag itself, it is unfortunate that cdb doesn't let creators be more specific about AI involvement right now. But i hope it would change soon.

[AbbyRead]

My fear is people using it in place of thinking and testing. The confidence it exudes is contagious and plays into out human biases. It's easy to trust what the AI says because it does a really good job at copying what an expert sounds like. If we're talking usage guideline drafting, a warning like that might be crucial for new contributors to avoid misunderstandings and over-reliance on it. It's easy for a familiar solution to start to feel like the only one.

I think code gen for boilerplate is a very good use case. But you often find it making comments for you in your code that make it sound like you're the one talking. That can be problematic if it gets posted because without context it can be mistaken for your own words. And it kind of is your words in that you're signing off on it when you post something.

LLMs are great, but I would mostly just advise people against getting sucked into the unearned confidence.

[corpserot]

My simple warning for that kind of misuse is:

you are responsible for your code. You can't just say "the AI told me to do that" when we ask questions and request changes. Be accountable for changes you propose.