Not looking to debate or stir up anything. But, I will say despite anyone’s thoughts the anti-meta fedi pact at least allowed people to have some information on where instances and people stood on federation with Threads. Some decisions were able to be made based on that information. I think if it doesn’t exist there should be one also regarding AI, which is a polarizing topic especially on fedi. I believe similar to the Threads issue it would allow people to make somewhat informed decisions one way or the other. Are there current lists, sites dedicated to this? If not what are you guys thoughts?

you are viewing a single comment's thread
view the rest of the comments

[–] Pamasich@kbin.earth 5 points 1 day ago (3 children)

And if users refuse to label them, then community mods have tools that help them detect and label them for the users.

And another reason for me to dislike Piefed, I guess.

There is no reliable way to detect AI use automatically, so these tools can't be a source of truth. But if they're not, what's even the point to them for moderation?

[–] 4Robato@lemmy.world 6 points 1 day ago

This issue happens regardless of your way of dealing with it. If you have instances that don't allow AI content, how will they identify when someone breaks the rule?

That's precisely why AI content is so dangerous in the first place.

[–] OpenStars@piefed.social 1 points 1 day ago* (last edited 1 day ago)

This issue is discussed a bit in the linked post.

The situation is a bit similar to that for NSFW: if all content is not perfectly labeled, then is the tag worthwhile to exist at all?

And mods will abuse the system regardless - e.g. look at how often Lemmy.ml removes posts citing "rule 1", despite that rule never anywhere anyway in any language stating that you are not allowed to criticize (or not support strongly enough?) Russia, China, or North Korea (or perhaps soon adding the USA and Israel?). Regardless of platform - Reddit, Lemmy, PieFed, Mbin, nodeBB, Mastodon, Pixelfed, Loops, etc. - mods are gonna mod regardless, and some will be better at that task than others.

Though it is still useful to have "rules", and use them in communities, even if they aren't properly implemented in all situations. The alternative being 4chan where anything and everything goes - not everyone is looking for that?

I guarantee you that Lemmy has for multiple years already abused non-AI content in a far worse manner: see e.g. the (non-AI) moderation tool santabot. Even so, Lemmy does not abolish the ability to post things, or to vote on it. i.e. the worst excesses of a thing do not preclude good uses for it. Knowledge that Santabot exists may encourage you to stop participating in the entire Threadiverse, but it does not need to? Systemically, the Threadiverse is fairly good, overall?

And now PieFed offers tools to detect AI content. I guarantee you that it will be abused, somewhere/way/when/how. Though at least now it exists, and can be improved? And we can be a part of that process too, by offering realistic suggestions to the development team, which demonstrably listens to and actively solicits such feedback all the time.

[–] hendrik@palaver.p3x.de 0 points 1 day ago* (last edited 17 hours ago)

There's good and bad tools out there? Most free internet services that pop up on the first page of Google are in fact rubbish. But that doesn't mean there's good ones. I once tried tools for detecting AI text for a different reason and there were one or two who didn't have any false positives with the 20-30 example texts I tried. Schools and universities have come to use the same tools. Seems they also look at how people are typing, that's pretty much 100% accurate, but people do both.

There's also crazy different approaches out there. Like looking at the probability distribution of the vocabulary and see if that matches ChatGPT. And it'll be a certain unique probability distribution since that's what ChatGPT is. It has to leave a fingerprint, since it's picking the words based on probability. And there's more good strategies. We have one or two open-source tools which demonstrate how it can be done without AI. There's of course also the option to train another AI model / classifier to figure out what constitutes AI text and what isn't. That also works.