AI-generated content: Legality, ethics, and the nature of art

Aren’t most AI detectors actually bullshit? I know it’s been picked up in academia that the tools that supposedly check if an easy is AI generated have terrible hit rates.

Maybe they'll be getting better

I'm not really talking about AI detectors.

Like, I'm not going to claim I'm infallible or anything, but like...if I clock something as AI created, I can prove it. I could easily be missing stuff, I'm not going to pretend that "I can always tell", but there's certain hallmarks of AI art and writing that very few people put effort into hiding.

What are these certain hallmarks

For art, it's usually the overly smooth feeling of the picture, but it can also include things like wrong porportions, backgrounds not making sense (particularly visible when a straight line is passing behind a character, the line breaks will usually be at the wrong place), jumbled letters, extra limbs (doesn't happen as much as when it was starting), bad symmetry... Probably other things.

For writing, I don't know as I don't really engage with writing that much these days, but judging by chatbots and certain examples I've seen before, AI tends to be overly wordy to the point of meandering.

Its more that you need a specific one for each AI model, and while combination ones exist, you generally have to build them yourself because they become computationally expensive the more you add to it.

Yeah, basically. Like, say the AI generator is asked to make an image of a room with a table.

Whoops, one pair of legs are square and the other are round. Weird, the table seems to be sagging upwards? There's a string of fairy lights in the background and most of them are missing the actual bulbs. Little weird that the tree outside seems to be growing into the window. Why's the road outside go from outside to inside to outside between two of the other windows? Why do all the windows look different?

It looks okay if you just glance at it, but if you look at the details, nothing quite adds up.

With writing, it tends to be a pervasive feeling of "oh, the AI doesn't know how to not sound like an 11th grader trying really hard to pump up the word count" and very few people using them are good enough or interested enough writers to prevent that from happening. Usually it's a lot of random extraneous details that add nothing to the situation or explanation. The "you pulled this from Wikipedia, huh" thing is the most obvious example, but it's a really telltale sign. Or it's just not well written.

Edited by Zendervai on Mar 27th 2024 at 5:01:26 AM

Which works if you’re looking at a work asserting to be from a professional writer, but it means if you’re trying to differentiate an AI from an 11th grader trying to meet a word count it’s going to be very difficult.

Heh. Actually it isn't. The trick is how you word the instructions.

Genuinely curious, how should you word it then?

We could do a little experiment, actually.

Write up a paragraph and then ask an AI to write a paragraph on the same topic using your prompt tricks. See if we can tell them apart.

And yeah, but I think it points to the general low quality of AI writing. Because no one would say an 11th grader doing that is a wordsmith.

Edited by Zendervai on Mar 28th 2024 at 6:34:14 AM

The BBC has a statement on their efforts to recreate the voice of a dying person for a documentary using AI. Source is Deadline: https://deadline.com/2024/03/mamma-mia-sara-poyzer-replaced-ai-bbc-show-1235870498/

I teach a class in which the assignments are heavily writing oriented. I am careful to word my instructions such that no AI is going to get it right, because I have given definitions or instructions that rely on my own interpretations of theory or research findings. For example, I describe two research designs that, in every textbook or online article I have ever read are combined into one design. Both approaches are technically correct, mine is just makes more distinctions. The point is that Chat GPT doesn't have access to my class lectures, so no one relying primarily on online sources is going to get it right. I have a bunch of tricks like that.

If you input an extremely narrow prompt about a technical subject into Chat GPT, nine times out of ten it just quotes Wikipedia back at you. All my exam questions or essay instructions qualify as "extremely narrow prompts about technical subjects", so the traditional tools used to control plagiarism in general will work.

Finally, none of my students write as well as Chat GPT does, and the adorable little guys generally don't seem to realize this (or, at least, none of them take the time to rewrite the plagiarised essay back into their own writing style). I have enough in-class writing assignments that I have a genuine sample to compare against.

Remember, cheating students are almost always lazy students, so they aren't going to use clever, carefully thought out strategies to cover their tracks.

https://adage.com/article/digital-marketing-ad-tech-news/ai-restrictions-added-ad-agency-contracts/2548696

So it turns out that a bunch of ad agencies are dealing with brands who are heavily restricting or banning the use of AI on any advertising they’re connected with. Because they don’t want to run the risk of their partners giving them something copyrighted or adjacent to copyright that they don’t have the rights for by accident and they really don’t want their stuff to be put into training data because of the risk of it being regurgitated and used by someone else pulling from the same AI.

It might be overcautious, but it’s interesting that ad companies are attempting to go all in but the actual brands are way more leery.

Eh, I think it's the same thing as Facebook and co working with automated moderation despite algorithms being nowhere near ready for the job - people are jumping the gun and want to cut costs even when it's not yet clear that the costs can be cut.

I'd imagine brands are more sceptical because they bear more (perceived) blame than the ad companies.

Edited by SeptimusHeap on Mar 29th 2024 at 2:50:52 PM

Yeah the repetitional damage from bad ads falls on the company whose product is being advertised, not the advertising company.

In automods case a lot of it is out of necicery due to the amount of activity social media has.

Let's look at youtube for example, where 10 hours of video are uploaded every second even if you wanted to rely on human moderation there... how would you.

That's just way too much content for humans to police.

So apparently, Robert Zemeckis’ next movie will use a generative AI process to digitally de-age Tom Hanks and Robin Wright in real time. Source is Wikipedia, under production.

https://en.m.wikipedia.org/wiki/Here_(2024_film)

Wrong thread.

Edited by BigBadShadow25 on Apr 2nd 2024 at 8:43:18 AM

I brought up the discussion a while ago in another ai thread and was redirected here

If an AI gains sapience what does that mean for the art it makes

It'd probably be really alien from our perspective because the experiences a computer program has wouldn't be even remotely analogous to our own.

It would also mean that we just moved into a realm where using AI like we do would be deeply unethical and that we're practicing active and literal slavery until we stop and figure out how to actually talk to it, because the chances of the AI not being able to perceive us properly or understand that we exist are really high too.

seconding the slavery part

that said, the question of "using" ai, or of enslaving it, depends a lot on how it works internally— everything that we've ever seen with any level of intelligence or self awareness evolved through a process designed to maximize individual survival and reproduction. That creates drives and instincts that result in the creature experiencing pain when enslaved. Its not possible at all to know what a sentient creature might want of it came to exist in a totally different way. The concept of "enslaving" it might be totally meaningless, or even not really possible to do

(yes this is more scifi than actually potential, i just find it interesting to think about)

I mean, part of the problem is that we have really fuzzy boundaries on where it's good to enslave vs not.

Like, would it be okay to enslave a bunch of six year olds? The vast majority of people would be like "what the hell, no!"

That's about how smart pigs are and we're pretty fucking awful towards them.

But if an AI starts autonomously putting out art, it should be immediately blocked off from public access to try and determine what's actually happening with it. It could just be a glitch or an error, or there could be actual intent there and letting it be accessible to the public might be a huge mistake.

Would the AI be able to tell the art it's making is getting it prohibited from the public, or that it was prohibited in the first place

Without public training data what would it do