Valve has reportedly become the latest company to react to the uncertain legal landscape surrounding AI-generated artwork by simply barring its use in submitted materials. An anonymous developer going by the Reddit handle potterharry97 reports having a Steam game page submission rejected for the use of "art assets generated by artificial intelligence that appears to be relying on copyrighted material owned by third parties."
Potterharry97 originally posted about the rejection in a May post on the now-private GameDev subreddit (partially archived here, Google Cache here). In that post, potterharry97 admitted that "a large portion of the assets have some AI involvement in its creation" through the use of Stable Diffusion. In a follow-up post this month on the AIGameDev subreddit, potterharry97 wrote that the initial submission was intended as an early placeholder version, "with 2-3 assets/sprites that were admittedly obviously AI generated from the hands."
[...]
These days, the use of AI-generated art can sometimes be easier to catch, as was the case with potterharry97's Stable Diffusion sprites and their telltale hands. But that might get tougher as improvements in generative synthesis models make AI art more and more indistinguishable from art created by a human.As potterharry97 put it in his initial Reddit post, "Even if I redo everything from scratch, how can I definitively prove if something was or wasn't AI generated?"
(Score: 2, Informative) by Anonymous Coward on Sunday July 02, @01:20AM (1 child)
Too much potential liability. Those claiming copyright infringement are likely to sue Valve.
FWIW even human artists imitating a photograph can be found guilty of copyright infringement: https://en.wikipedia.org/wiki/Rogers_v._Koons [wikipedia.org]
So if you are found remaking a specific work then there might be problems. But what counts as a specific work? Do textures and patterns count?
(Score: 1) by khallow on Sunday July 02, @04:38AM
The linked example was extremely close and deliberate copying. It wasn't merely imitation. And consider this description of what happened from a "list of background facts" [courtlistener.com]:
So not only very close copying of the work, but deliberately ignoring of the copyright. Sounds like the court case was a better exhibit of banality than the photo was!
(Score: 3, Interesting) by ledow on Sunday July 02, @10:29AM (4 children)
You need to be able to prove the origin of your works, including any other works that may have been used, remixed, "inspired", portions utilised, or wholesale ripped-off.
This isn't "new". This is what copyright is. If you don't know the origin of your work, and therefore aren't willing or able to testify to it, others have even less chance of assessing that correctly, and it's rejected by default.
Copyright isn't a matter of "just trust me". You have to be able to know the origins of your works. And given that I can say to most AI image generators to give me a copy of the Mona Lisa or the Eiffel Tower at night or some still-in-copyright picture... somewhere that data was used to train the AI and influence its behaviour and you have no idea whether it just used it for "inspiration" or whether it just copied it wholesale because the MACHINE CANNOT ATTEST to that. You can, the machine can't. If you are willing to attest that it absolutely did nothing that a copyright holder would ever hold Steam liable for, then you could put that in writing and use it to take on all such liability if a copyright holder steps up (and don't expect places like Valve to then represent you if you are legally challenged, even by a fraudster).
If you're a content creator, you need to be able to list where you got your resources from. That's the same for hobbyist or professional in a large company. You can't just slap Google Image results in there, or take Fred's word that he didn't copy the music verbatim from another artist's tracks (lawsuits about that happen all the time!). You need to know the origin of your work. If you had an AI that *YOU'D* trained on a set of data that you *KNEW* didn't contain any copyright images, then you could attest to that. But you didn't.
And if you think I'm not the kind of person to eat my own dogfood: I write software. Hobbyist software. Small games, contribution of patches, did a lot of work on some Linux distros, etc.
For the games that *I* personally wrote from scratch I still have - 20 years later:
- Every original source image. With details of the artist (sometimes someone PAID BY ME to produce artwork to my specification, or people I know). And paid receipts.
- Explicitly-stated emails from artists who let me use their music - telling them what I intend to use it for, how, etc. and then replying in the affirmative that I could under certain conditions. Those conditions were always adhered to, including attribution and usage.
- Every conversation with those artists.
- Every licence for every library, short snippet of code, API interface, etc. used - including checking they were compatible with my intended licence for distribution. And all parts commented where they ended up in my code and how they evolved with CVS, SVN, mercurial, git etc. histories.
I didn't so use much as a single character out of a font without contacting the author, getting permission, and storing that information alongside the source and attributing appropriately in the final product.
Hell, I licenced my forum avatar that I use. I literally paid for the rights to be able to do that. For a tiny GIF. I kept the receipts and licencing agreement to that, too.
And though you can "get away" without doing that, you're literally "getting away" with something - and it can catch up with you at any time.
I still remember playing XQuest 2 and wondering how they got permission to use the Simpson's "doh" sound for when the ship hit a solid object. I still wonder to this day. You can get away with it for decades... but equally you could land Valve into a multi-million dollar lawsuit (even if that's a frivolous one) just as easily.
So I get why they are paranoid about it. I don't understand why, if you're generating content, you can't just maintain a record of that content. You have version control, hyperlinking, you could save copies of the original download in your build tree, you could put in hyperlinks and even screenshots of the source webpage that you got that content from. You could prove everything you used, ever, for almost no effort beyond that taken to get the files in the first place.
But when you start using AI tools that someone else has trained, when those places *ARE* being sued for copyright infringement of their training data, you lose that chain of accountability almost immediately. You have no idea if that AI tool used an illegal reference or not, whether it wholesale incorporated it into your images or not. And copyright law only says things like "substanstially", "similar", etc. - it goes far beyond just identical copy-paste.
Some random IBM presentation in the 70's nailed it:
"A COMPUTER CAN NEVER BE HELD ACCOUNTABLE. THEREFORE A COMPUTER MUST NEVER MAKE A MANAGEMENT DECISION."
You can't sue Stable Diffusion if they supplied you with data that's later found to be copyright - it's on you, and they have EULAs to exclude most liability and tell you that this can happen. Even if you did, they can't "sue" their AI for it, they are lumbered with the problem and they are the people who facilitated, even actioned, copyright infringement if they are found guilty. And you can't pass that buck onto Valve either. Because they *won't* be able to pass the buck when they are distributing it, except back to you.
As it all comes back to you, either way, you need to make sure that you adhere to copyright - no matter how dumb, outdated, overbearing, etc. you consider it. Because even if you're prepared to "take the risk", nobody else is going to on your behalf and ultimately it will all come back to you. You chose to use that data, despite an EULA that says not to, for example, or you attested to its origin and Valve later find out you were lying.
As a programmer, or a developer including artistic assets, even storywriting and character names - you have to take responsibility for the copyright in your code. That means being able to attest where every line came from, that every pixel is being used with permission and appropriate clauses adhered to, and that every sound sample was original or licenced in good faith.
And with modern tools, source code management, hyperlinkage, etc. it's literally seconds of work for each new asset.
Valve can't do that for you. Stable Diffusion can't do that for you. You need to do it. You're the only one who knows where you got your assets, the same way that Stable Diffusion are the only ones that knows what training data they used to generate their images. Just trusting that they are completely licensed isn't acceptable further up the chain, e.g. Valve, without some form of attestation to that.
And let me tell you: Stable Diffusion won't give you an attestation that their AI wasn't trained on any copyright data. Here's why:
https://waxy.org/2022/08/exploring-12-million-of-the-images-used-to-train-stable-diffusions-image-generator/ [waxy.org]
(Score: 2) by Rich on Sunday July 02, @12:12PM (3 children)
It's not that black and white. The brain has a learning process similar to that of AI, and the unconscious reproduction of copyrighted works is possible as well.
- There has been an uproar over a song from German hard rock band "Freiwild" copying a phrase from the right wing extremists "Stahlgewitter"s song "Auftrag Deutsches Reich". That wasn't really about copyright, but more about career suicide, given the somewhat shady political past of Freiwild members, who since had switched to commercial success mode. I'm positively sure that the Freiwild guys didn't copy that phrase intentionally, but I'd also assume they had stuck it in their brains from music they listened to far in their past. Such an oops can be expected from anyone who has seen and mentally processed copyrighted materials in their past.
- AI is so prevalent today that you can't exclude its use in your toolchain. I recently got a picture sent from a car make club meet with a large line-up of pretty sports cars. I looked at the distribution of the different models and as I zoomed in on the more distant ones, there was one car that had the front of one model and the rear of another. That was clearly a sharpening AI artifact. Not that anyone was informed there was an AI in use that could lead to such artifacs, but eve if so, could we know whether that AI was trained "clean"? Who shot the photo of the moon that Samsung phones capture better than the optics allowl? Or can we be sure that the magic inpainting of Photoshop is "clean"?
- If it comes down to accountability, would that of Adobe be higher valued (corporations lie and deny all the time until they have to admit that it was 'a single unfortunate mistake by an individual who has since been laid off') than that of Mr. Zwane Mkumbu of Lagos, Nigeria (who, after you rejecting his offer to tap into the wealth of African royalty, offers you guaranteed copyright-free works for $1 a piece)?
In the end it just comes down to reducing the risk of being hit by a lawsuit. And that lawsuit wouldn't come forth for justice, but only if there's a.) money to be sucked from you, or b.) intending to push you out of business. In the Valve case, I have a suspicion that they got a little blackmail discussing the idea of b. And that would come from a class of investors that want to hedge against the chance that the AI revolution becomes a grassroots thing rather than allowing them new dimensions of rent-seeking.
(Score: 0) by Anonymous Coward on Sunday July 02, @02:10PM (2 children)
You're basically saying humans have committed copyright infringement allegedly unintentionally.
But you have not provided any evidence that current AI tools can't/won't do similar stuff.
So it's pretty much black and white for now - you can't safely use AI stuff since you can't prove that AI trained on copyrighted stuff won't infringe.
Most of the AI stuff is not doing any clean room "remakes": https://en.wikipedia.org/wiki/Clean_room_design [wikipedia.org]
(Score: 2) by Rich on Sunday July 02, @08:37PM (1 child)
Of course not. I'd guess you run a very roughly similar risk to end up with an infringing output from AI as from humans. For the task "a 1990s men magazine centerfold picture", both AI and human are trained on the same original copyrighted stuff, the weaving of the original material into the neuron fabric works effectively the same. It's just that the human has bio-memory as opposed to silicon memory, and the law treats humans differently from data files.
No. It's all black in your logic because you can neither use human made stuff because you can't prove it's clean either. Even worse, a contracted human could just flat out lie and provide neither his own work, not even something clean done with AI, but plain out copied original material. (We've just had the article about the guys having their fake record label to run a scam on YT.)
The major difference is that there's a very big drive from the money side not to let the technology fall into the people's hands and they will certainly launch some very big lawsuits in the next time, and they also may use smaller outlets as victims to establish case law. As a business owner you don't want to be on the receiving end here, at all if you have less than a billion in the bank, and even then only when you can turn the outcome around and in some way screw over the media house that sued you in return, for a lot of cold cash, because your own owners don't like seeing their money being wasted on lawsuits.
(Score: 0) by Anonymous Coward on Monday July 03, @05:58AM
Sure but if he gets caught, Valve doesn't get in trouble.
If the AI gets caught, who gets in trouble? Valve.
(Score: 3, Funny) by VLM on Sunday July 02, @07:45PM
Don't ask don't tell for a new generation
(Score: 1) by A_Gopher on Tuesday July 04, @12:35PM
Don't include the AI generated assets in the game. Instead just have the game download a public domain model. Include a prompt + seed + crop co-ords in the game to generate each asset on demand