D&D General DMs Guild and DriveThruRPG ban AI written works, requires labels for AI generated art

[Jer]

As I understand the technology, midjourney at least does not copy images. It reads patterns in them on a pixel by pixel basis and connects that data to key words. Note that I am not defending the scraping, just explaining the difference.

Somewhere in that process they made a copy of the image. They had to - there's no way to read an image file and not have a copy somewhere on the computer that you're reading it on. Even if it's just in the RAM and it never touches a disk.

Our entire Web usage has skirted around the issue of copyright and the fact that browsing the web necessarily means making copies of everything we view on it. We've all kind of chosen to ignore it because it wasn't in anyone's real interest to try to figure it out - if I post something for you to read I want you to read it, and if that means you have to download it to your machine to read it, well, then I'm implicitly giving you permission to do that just by the nature of the technology.

Does that mean that I'm also giving you permission to use the same text/image to train a neural network to reproduce things in my writing/art style? I wouldn't think so personally, but that's a question of law that needs to be figured out. And I personally hope they resolve it correctly and say "no, that permission was never granted, pay them for the right to do that or don't do it."

 

[Celebrim]

That’s not necessarily the issue. To train the AI they copy the images so that the ai then is able to repurpose parts of them into a new image.

That's not how it works. The AI doesn't make a mosaic of the images it has copied. The AI looks at images and attempts to categorize them as cats, tables, mountains, whatever. Once it has built up an idea of "catness", "tableness", or "mountainness" it turns out that this level of understanding is not only sufficient to look at a picture and say, "It's a cat" or "It's a table", but to draw completely novel and original cats and tables. That's fundamentally no different than what any artist does. This just codifies that process.

Of course the whole internet essentially works that way as well.

Exactly. Every time you view an image on the internet you make a copy of it.

Humans do to some degree as well via memory - though memory isn’t an exact copy as the ai training data would be.

But the memory of the image that the computer has isn't an exact copy either. It's not like the images have to be stored once the neural net is created. The images themselves are recorded in the form of electronic "memories".

So unless extremely careful lots of the internet could be impacted by decisions around an art as well.

The whole point of copyright law is to encourage creativity. The courts takes a rather dim view of using copyrights to squash creativity. TSR back in the old days won a legally significant case on those grounds. This is true transformative and creative work, and as such it's legally protected. If it isn't legally protected, then things will get really weird.

 

[FrogReaver]

Somewhere in that process they made a copy of the image. They had to - there's no way to read an image file and not have a copy somewhere on the computer that you're reading it on. Even if it's just in the RAM and it never touches a disk.

Our entire Web usage has skirted around the issue of copyright and the fact that browsing the web necessarily means making copies of everything we view on it. We've all kind of chosen to ignore it because it wasn't in anyone's real interest to try to figure it out - if I post something for you to read I want you to read it, and if that means you have to download it to your machine to read it, well, then I'm implicitly giving you permission to do that just by the nature of the technology.

Does that mean that I'm also giving you permission to use the same text/image to train a neural network to reproduce things in my writing/art style? I wouldn't think so personally, but that's a question of law that needs to be figured out. And I personally hope they resolve it correctly and say "no, that permission was never granted, pay them for the right to do that or don't do it."

I guess technically a technology could be devised where that wasn’t the case - the sender could send you the data one pixel at a time and your machine confirms it’s deleted before sending the next pixel. No technology currently does that, and it would be based somewhat on the honor system, but can be done and might theoretically be enough to prevent the receiver from infringing on the copyright if any image data transmitted this way.

 

[TheSword]

The internet is at its best when used to share ideas for free.

It’s also ironic because the anti-capitalism rhetoric being used ignores the fact that most of the rpg products that have been told they can’t use this stuff are extrmely low budget, low margin products that probably wouldn’t have imagery beyond a doodle because the publisher is on a shoe-string budget. What’s the average price of product on DM’s Guild? £2? £3?

 

[Sacrosanct]

The internet is at its best when used to share ideas for free.

It’s also ironic because the anti-capitalism rhetoric being used ignores the fact that most of the rpg products that have been told they can’t use this stuff are extrmely low budget, low margin products that probably wouldn’t have imagery beyond a doodle because the publisher is on a shoe-string budget. What’s the average price of product on DM’s Guild? £2? £3?

RPG products using AI art are popping up on Kickstarter, not just $2 things on DTRPG. Also, as I mentioned earlier, traditional publishers using real art are getting lost in the shuffle because of the rapid release of AI stuff.

 

[FrogReaver]

That's not how it works. The AI doesn't make a mosaic of the images it has copied. The AI looks at images and attempts to categorize them as cats, tables, mountains, whatever. Once it has built up an idea of "catness", "tableness", or "mountainness" it turns out that this level of understanding is not only sufficient to look at a picture and say, "It's a cat" or "It's a table", but to draw completely novel and original cats and tables. That's fundamentally no different than what any artist does. This just codifies that process.

Even in your pushback here, you define the AI as looking at images. The AI doesn't do this one time and call it a day. It would be a combinatorial nightmare to pre do the required combination of things someone might enter into the prompt and you'd quickly run out of data storage trying to store the necessary 'models'.

Think: 'Elephant', 'African Elephant', 'Animated Elephant', 'Abstract Elephant', 'Elephant in style of insert-famous-artist', etc.

Each of those and the 1,000,000's more prompts like them would require their own stored understanding if it worked as you describe.

But the memory of the image that the computer has isn't an exact copy either. It's not like the images have to be stored once the neural net is created. The images themselves are recorded in the form of electronic "memories".

See above the combinatorial storage problem

 

[bedir than]

What do you mean. As I understand it, there are generative AI trained on public domain and licensed datasets. That is the only ethical standard.

Can you point out any of these?

 

[Jer]

I guess technically a technology could be devised where that wasn’t the case - the sender could send you the data one pixel at a time and your machine confirms it’s deleted before sending the next pixel. No technology currently does that, and it would be based somewhat on the honor system, but can be done and might theoretically be enough to prevent the receiver from infringing on the copyright if any image data transmitted this way.

Maybe, but get a good enough lawyer and you can argue that you've made a copy though. I mean, you have taken it one pixel at a time but how is that that much different from how our TCP/IP technology breaks files down into packets and sends them over the internet? And it's not that much different from saying that if you photocopy a book and send it to someone one page at a time (or even one word at a time) you haven't actually made a copy of the book, which IANAL but I suspect that wouldn't stand up in court.

Copyright is an agreement between the public and the artists who create copyrightable works. It's not a physical law, it's just a way of organizing our laws so that artists and publishers get paid for the work they do. So if someone has come up with "one weird trick" to get around the purpose of copyright, then we need to think about why we have copyright (to promote the creation of new works) and fix things to make sure our laws are still doing that.

And the AI models out there NEED to have new novel works from human beings in order to continue to function. I've now read two papers that talk about "model collapse" where you feed work that an AI has created back into itself and very rapidly (within a few training cycles) the model degrades into garbage. It's true for both the language models and the visual ones. These models are incredibly fragile statistical creations and have a high dependency on distribution of pixels/words in the training sets they're given. If the companies that want to make/use these models want to keep them working into the future, it's in their best interest to not kill the golden goose which is human beings creating new works.

 

[Moonmover]

Most people in this thread are speaking from the point of view of artists and writers, or considering the legality of machine learning in general. That's all well and good, but let's think about this from a customer's perspective as well.

AI-generated products are just spam. They don't provide anything the consumer couldn't have made themselves using the same free software in a few minutes. Creating an adventure module in ChatGPT takes only slightly longer than buying one from DriveThru, and results in a similar level of quality. That goes triple of AI-generated art, which often does not even need to be edited once generated.

So, this new ban is all good news from a consumer's perspective.

 

[TheSword]

RPG products using AI art are popping up on Kickstarter, not just $2 things on DTRPG. Also, as I mentioned earlier, traditional publishers using real art are getting lost in the shuffle because of the rapid release of AI stuff.

What kind of stuff is out there? I got burned on Kickstarter so very rarely use it now.