Distributing AI generated content

[ithoriel]

My thoughts, Daeothar?

Time to usher in the Butlerian Jihad!

"Thou shalt not make a machine in the likeness of a human mind" 

  Where did I put those Shai-Hulud miniatures?

[_Si_]

@Daeothar, yup pretty much my thoughts on it. If AI is reassembling Stephen King's work into a vaguely coherent form and saying "I made this" then that's just plagiarism. But it reads everything by King and then writes a King-esque novel then how's that different from a human doing the same?

I get why creatives are super nervous, and there has to be an element of responsibility from the companies who might choose to use AI instead of employing an artist/author, but you'd expect if we actually value human creativity then that will naturally rise to the top.

Definitely interesting times!

[Easy E]

you'd expect if we actually value human creativity then that will naturally rise to the top.

That's the problem.  We don't value human creativity. 

At least, not enough to pay a premium for it.

[ced1106]

> At least, not enough to pay a premium for it.

Well, I'm seeing many a hobbyist use AI, and they don't have money to even pay themselves!

Before AI, I reviewed PDFs, and, imo, there's too much unnecessary art as it is. Yeah, it's pretty to look at, but it often has nothing to do with the text on the page, and, much worse, consumes ink when printed out -- and the page curls as the ink dries. The only excepts would be illustrations for a scene in an RPG scenario, or diagram for a wargame, but that's about it.

Now, as you get to moneybag corporations, yeah, that's another issue because... they have money. I remember *decades* ago about how Hollyweird writers were sued for stealing ideas from other writers. Derivative works may have laws *protecting* them, but you gotta hire lawyers to protect *you* and lawyers gotta be paid. Even if you hire an artist, if they swipe art from another artist without telling you, that may not stop a lawyer from going after you, law or no law. Settlements are a thing, and lawyers may be going after targets for the settlement because actual court costs, including time, are not worth it for the target.

AFAIK, While a human artist can claim that they were only influenced by someone else's art, AI directly relies upon art on the internet, IP-protected or not, so is closer to "swiping". Or so an entity with money may think, regardless of the law. *They* just don't think a possible lawsuit or other hassle is worth the benefit of additional AI content. And, besides, many of us humans will work on hobby content for free, with better results than a computer.

[paspas]

@paspas, I use Affinity Publisher 2. I've used InDesign a lot in the past, but couldn't face the subscription costs these days. This one did everything I needed it to and was pretty easy to use.

Thanks

[katie]

"If AI is reassembling Stephen King's work into a vaguely coherent form and saying "I made this" then that's just plagiarism."

That is not how an LLM works.

LLMs work by taking your prompt, encoding it, turning it into a vector in an extremely high dimensional space and returning neighbouring vectors which are decoded back into text. It isn't the product of "creativity" but it also is not a re-assembly of extant information.

So the prompt is basically asking "convert my question into an XYZ and then find what's next to X,Y,Z" where there's 10,000 coordinates instead of three and you also want "Prefer heading in the direction Stephen King stuff tends to go". This is how you can get it to explain (say) baking in the style of Stephen King...

Note that even the axes of the highD space are not human-comprehensible. There's no "Stephen Kingness" axis that you're moving up and down. We don't understand the axes (that's one of the reasons it's 'deep learning').

It's HIGHLY unclear that any claim of copyright derivation is going to succeed. It's going to make musical claims of derivation look easy to judge. Given the vastness of the inputs, the difficulty of attributing any given important delta to a bit of content and the difficulty of attributing any output effect to that delta... mathematically, it's almost insignificant.

What might succeed are claims that ML processing constitutes a failure to obtain the correct mechanical rights in the work. They might have cases, but it's difficult to see how publishing something doesn't constitute an implied right to process it. After all, your computer has to process it to show it to you. But I'm not a lawyer. (I'm a software engineer working with ML systems.)

[katie]

Image generation works a bit differently. Basically you take a bunch of images, labelled with descriptions. "This is a cat", "This is a plane"

You repeated blur the images, "This is the cat but blurred", "This is the plane but blurred" and let it learn those. Repeat until there is no information in the image (so they're all just an informationless blur). So the system is basically learning to steps to *unblur* images guided by the labels.

Then you give it a random blur and say "if this is a completely blurred picture of cat riding on a plane... what does the original look like?"

Once again, it's going to be quite hard to attribute any of the output elements to any of the input elements..

[_Si_]

That's all really interesting, and definitely how I hoped it would all work. Makes it so much more exciting that way. Thanks for the insight!

The trouble is that for all there will be 'expert witnesses' for both sides, the legalities of this are going to be decided by people who don't have a hope in hell of understanding the tech.

I'll be really interested to see what comes out of the AI summit at bletchley park this autumn. But you get the feeling legislation is never going to keep up with the speed of development here.

Like I said, interesting times