Valve Responds To Claims It Has Banned AI-Generated Games From Steam 56
An anonymous reader quotes a report from TechCrunch: Valve has issued a rare statement after claims it was rejecting games with AI-generated assets from its Steam games store. The notoriously close-lipped developer of the Half-Life series and de facto gatekeeper of PC gaming distribution said its policy was evolving and not a stand against AI. Steam has a review and approval process much like any app platform, and its rules on content aren't always clear until developers test them with edge cases. So it was with one indie dev who posted in a subreddit for like-minded game developers using AI, saying Valve "is no longer willing to publish games with AI generated content."
The game they had submitted had "a few assets that were fairly obviously AI generated," and Valve appeared to take issue with this. "As the legal ownership of such AI-generated art is unclear, we cannot ship your game while it contains these AI-generated assets, unless you can affirmatively confirm that you own the rights to all of the IP used in the data set that trained the AI to create the assets in your game," their first warning letter stated. Then, a week later: "we reviewed [Game Name Here] and took our time to better understand the AI tech used to create it. Again, while we strive to ship most titles submitted to us, we cannot ship games for which the developer does not have all of the necessary rights. At this time, we are declining to distribute your game since it's unclear if the underlying AI tech used to create the assets has sufficient rights to the training data."
Considering most AI tools can't really claim to have legal rights to all their training data (and even if they do, it may still not be an ethical use of that data), this policy as stated basically amounts to a blanket ban on AI-generated assets in games. [...] If the creators can't realistically claim copyright over their own work, Valve has deemed the risk of publishing that work too high. As such, Valve responded to Eurogamer to say that, basically, their policy is more "what's legally required" than any particular stance on AI. "We know it is a constantly evolving tech, and our goal is not to discourage the use of it on Steam; instead, we're working through how to integrate it into our already-existing review policies," Valve said. "Stated plainly, our review process is a reflection of current copyright law and policies, not an added layer of our opinion. As these laws and policies evolve over time, so will our process."
The game they had submitted had "a few assets that were fairly obviously AI generated," and Valve appeared to take issue with this. "As the legal ownership of such AI-generated art is unclear, we cannot ship your game while it contains these AI-generated assets, unless you can affirmatively confirm that you own the rights to all of the IP used in the data set that trained the AI to create the assets in your game," their first warning letter stated. Then, a week later: "we reviewed [Game Name Here] and took our time to better understand the AI tech used to create it. Again, while we strive to ship most titles submitted to us, we cannot ship games for which the developer does not have all of the necessary rights. At this time, we are declining to distribute your game since it's unclear if the underlying AI tech used to create the assets has sufficient rights to the training data."
Considering most AI tools can't really claim to have legal rights to all their training data (and even if they do, it may still not be an ethical use of that data), this policy as stated basically amounts to a blanket ban on AI-generated assets in games. [...] If the creators can't realistically claim copyright over their own work, Valve has deemed the risk of publishing that work too high. As such, Valve responded to Eurogamer to say that, basically, their policy is more "what's legally required" than any particular stance on AI. "We know it is a constantly evolving tech, and our goal is not to discourage the use of it on Steam; instead, we're working through how to integrate it into our already-existing review policies," Valve said. "Stated plainly, our review process is a reflection of current copyright law and policies, not an added layer of our opinion. As these laws and policies evolve over time, so will our process."
Resistance is futile... (Score:2, Insightful)
Ban it all you want, the whole point of AI is to become indistinguishable from human, the so called Turing test.
So all they are doing, is validating the AI. It is a race to see who can fool the censors the most. There is no turning back, now in its infancy, you may be able to spot it, but not for long...
Re:Resistance is futile... (Score:5, Informative)
Re: (Score:3)
Not sure what you mean here. The models are very much deterministic, you'll get the exact same output if you provide it with identical input. I don't know if Midjourney/Dall-E hides it from you but Stable Diffusion just introduces randomness with a seed, and if you keep it constant along with the prompt and other parameters, you'll get the same output every time.
Re: (Score:2)
First off, if you use xformers, you really don't get the same output just by using the same seed, params, and model.
Secondly, saying "it's determininstic if you us the same huge set of mdoel weights and parameters from a truly vast possible configuration search space" is like saying "humans are deterministic so long as you ensure that all of their neurons are in a given starting state".
Re: (Score:2)
Well I don't use xformers :)
Also I think it's fair to say that if the whole system the way it's used - model, weights and settings - produces identical outputs for identical inputs, then it's deterministic.
Comparing this to humans gets us into the whole "free will" debate for no reason, as well as the practical impossibility of ensuring that all the neurons are in a given starting state. As opposed to using Stable Diffusion being the way millions use it right now all the time.
Re: (Score:2)
> as well as the practical impossibility of ensuring that all the neurons are in a given starting state
Yes. So much so. An actual given starting state is complex beyond our understanding when speaking in the context of human biology.
Re: (Score:2)
Also I think it's fair to say that if the whole system the way it's used - model, weights and settings - produces identical outputs for identical inputs, then it's deterministic.
That's just it. You don't necessarily get the same output from the same input. At least, not in any meaningful sense. That is, the user can only control their input and not all input. You can tailor the user input to optimize the likelihood of the same output, but it's not guaranteed.
Comparing this to humans gets us into the whole "free will" debate for no reason, as well as the practical impossibility of ensuring that all the neurons are in a given starting state.
You can make this very same argument about how these AI models work. Just to simplify things, we know that random number generation is technically deterministic. The generation is based off of some kind of input. However, when
Re: (Score:2)
Re: (Score:2)
Re: (Score:3)
Re: (Score:1)
I think you're mistaking the definition of "deterministic" here. Those hardware RNGs simply produce inputs to the system, and given the same input the system will produce the same output. A human may be incapable of knowing all the inputs, but a computer can spot the slight differences between that and what you'd get if the outputs were truly independent of the inputs (just like it can spot the difference between a really good pseudo-RNG and a true hardware RNG even if it can't see the seeding for the psued
Re: (Score:2)
To be non-deterministic, the system would have to be able to produce different results even if fed the exact same sequence of "random" numbers.
That's a moot point if the numbers it's being fed are truly (unpredictably) random. Remember there's some good evidence to believe the universe is deterministic. We certainly can't prove that it's not, at least not yet. So if the entire universe is deterministic, then so are humans and AIs, no matter what you do.
Re: (Score:2)
Re: (Score:2)
Temperature used in LLMs, not diffusion models. Are you thinking of cfg_scale? It's not related to the concept of temperature in LLMs.
Re: (Score:2)
Re: Resistance is futile... (Score:2)
How do you know human creativity is non-deterministic? Got any evidence to back that up? Any evidence that a sufficiently advanced AI is any more deterministic than human culture and groupthink?
Re: (Score:2)
AI-generated content is as deterministic as human-generated content, when the humans are imbibing psychedelic mushrooms while posting. Since you seem to have experience with that, can you explain it to us?
Re:Resistance is futile... (Score:4, Insightful)
The users could try to fool them, but the model publishers will be easily caught out if they use unlicensed content in their training set.
If the US supreme court judges copying into training sets is not fair use, it's all going to come tumbling down.
Re: (Score:3)
The Turing test is not a rule set in stone, it's just one (though brilliant) man's idea regarding something -- human-like interaction with a machine -- that he was guessing about as it didn't exist in his time.
As for me, rather than making it a yes/no choice, I'd rate AI on how long, on the average, until people start suspecting they are talking to the machine and not a human. From my experience, it always happens sooner or later, it breaks the immersion, and you feel like you've wasted your time.
As for AI-
Re: (Score:2)
Ban it all you want, the whole point of AI is to become indistinguishable from human, the so called Turing test.
So all they are doing, is validating the AI. It is a race to see who can fool the censors the most. There is no turning back, now in its infancy, you may be able to spot it, but not for long...
Except they aren't banning it, Valve said from the get go that it was banning content where the copyright is ambiguous, obviously in order to cover their own arses from being sued. This has little to do with AI and more to do with the fucked up US copyright system allowing organisations to sue on the most spurious of grounds.
Re: (Score:2)
I mean, yeah (Score:5, Interesting)
They're a business, until such time as they can be reasonably sure they're clear of any legal problem with AI generated content they're not going to deal with.
Re:I mean, yeah (Score:5, Insightful)
Indeed. Especially as damages could get huge for something like this and you can never rule out that AI generated content can be identified as a copy that is still too close to some original to be legal. There have even been instances where people managed to make generative AI spit out part of its training data.
While this is hard on the developers, Valve is just being reasonably cautious. Especially generative AI that is used to generate stuff that will then itself be copyrighted must only be trained with clean data where the company training the AI had explicite permission for all of it. The current crop of generative AI is trained on stuff like web-data and you often cannot even find out what exactly it was trained on.
Re: (Score:1)
Not an issue in this case because it's out of copyright. But if you try "Mickey Mouse" and get a weird picture of a mutant mickey, it's pretty clear that the training data contains a decent amount of Disney images. I doubt Disney are being that free with their licencing.
Re: (Score:2)
Yes. Not hard to do and easy to not notice that you should better not use some specific thing.
Re: I mean, yeah (Score:1)
Re: (Score:2)
Possibly. Or not. Do you want to risk it? Especially after you may have, say, 100'000 copies of a game that may or may not be in violation of copyright? Fancy some long, expensive lawsuit and maybe a prohibition on selling your game while that goes on?
Re: (Score:2)
No court is going to declare a photograph of a painting to be transformative.
And you can't "train on the original" because the original is not digital.
And this is all moot because it's public domain, and indeed, one of the most widely reproduced pieces of art on the planet.
Re: (Score:2)
[Citation needed]
Re: (Score:2)
What on Earth does that have to do with the claim that you can gain copyright over a painting by taking a picture of it, because of "the choice of equipment, composition, lighting, processing and correction"?
Re: I mean, yeah (Score:1)
Re: (Score:2)
Photographing a copyrighted work does not mean you can reproduce the photograph without violating the original work's copyright. Period. There's nothing "transformative" about taking a photograph. Even hand-made paintings or sculptures that take hours are commonly found to be non-transformative if they resemble an original too much. See for example Jeff Koons' "String Of Puppies"
Re: (Score:2)
What, you mean one of the most famous pieces of public domain artwork in human history, which is mass-reproduced all over the internet in countless forms, and will exist in the training dataset (in various forms) thousands upon thousands of times? Yeah, I'd argue that it would be a bug if it couldn't reproduce it at all. If it couldn't, what hope would it have of, say, reproducing a national flag, or the Washington Monument, or, basically, anything?
This of course has no bearing on the question of being ab
Re: (Score:2)
As for Mickey Mouse (copyright of fictional characters, as distinct from their works, which is the exception rather than the rule), by far the mass violator of Disney's character copyrights is humans drawing or photoshopping them. Don't ask for a well-known copyrighted character, then when you get back something that kind of looks like them, blaming the model, and not yourself for what you asked it to create.
Re: (Score:1)
Instead of "Mickey Mouse", I create "Cartoon Mouse" or public domain characters like "Snow White" or "Beauty and the Beast". The AI creates something superficially similar to Disney's versions. If a human artist created it there would be no question, but the AI artist created it.
Disney comes and insists that it uses their IP and demonstrates that the model does indeed contai
Re: (Score:2)
There is no "memorization of images". It's simply learning the rules that define "how things look".
Wrong. It is not learning any "rules". That is just a bullshit anthropomorphization. What it really does is it calibrates statistical classifiers that may have many dimensions but that, unlike rules, have no dept. It can happen that some training data goes into the system almost verbatim. There are well-known effects like "overfitting" that do this. There usually is some fuzziness but that may or may not be enough to prevent copyright problems.
Re: (Score:2)
You know, you could actually read through to the paragraph that starts, "Since diffusion image generation models ..." before responding, and note the link therein.
And yes, neural networks are self-assembled binary classifiers, each one establishing a rule that divides a dataset, and with each layer making use of the classification decisions of the previous layer to form ever-more-complex classification tasks. [youtube.com]
Or to put it another way: they're logic engines.
Re: (Score:2)
Seriously? You are going to uphold that flawed claim? Well, maybe the difference is that I actually passed an academic course on ANNs and another one on deduction systems in my elective subjects as part of my CS master's. You obviously are lacking these basics. The "rules" in ANNs are _statistical_ rules, not _logical_ rules. Otherwise they could not be trained.
You can simulate ANNs in deductive systems (with extreme inefficiency). You cannot simulate deductive systems in ANN, as ANNs do not iterate rule ap
Re: (Score:2)
Except there is no "case". The open legal question is whether the creator can get copyrights on their creations, or whether nobody can. Not whether some third party can.
Based on the current US standard, raw outputs cannot be copyrighted, but the act of modifying or curating (e.g. selecting) them can yield copyright protection. But it should be entirely irrelevant to Valve whether the images are copyrighted to the creator or to nobody, unless they're planning to ban public domain art as well.
Copyright is b
Re: I mean, yeah (Score:2)
As if it didnâ€(TM)t already have problems to begin with.
Re: (Score:2)
It's obviously a CYA move but I don't see how would they be liable exactly. Any more than e.g. Amazon or Netflix if one of the movies used AI-generated texture in a CG scene or something. If a particular game is found to be infringing on someone's copyright, they can take it down.
Otherwise it looks like banning games that use photographic textures because a camera could've been used to take a photo of copyrighted content.
Re: (Score:2)
When I release games to Valve that don't rely upon AI, why doesn't Valve verify that I have ownership or license of all the assets I use in my games? Because they don't.
They only seem to care about AI because so many starving artists are complaining about truly starving, once their lack of talent prohibits them from making any living whatsoever.
Don't they realize (Score:2)
It's a danger for all platforms (Score:4, Interesting)
The other day I was trying to find YouTube videos about a newly launched car and encountered FIVE AI generated channels. All of them worked the same way - slide shows and a speech synthesized regurgitation of some website article that actually reported about this car.
Meanwhile Amazon is being choked up with fake AI books filled with AI generated images. And of course there are people on YouTube giving tutorials and instructions on how to make this trash.
This shit is poisoning everything. Obviously I've mentioned YouTube twice here - AI generated garbage and AI tutorials. They need to sink this crap immediately. Make it an reportable offence to use their platform for either and demonetize it. And work with speech synth & AI image providers to see if they can watermark the output to make it detectable. And Amazon really needs a report button too and change its algorithms so new self-published books can't be gamed to shoot up the charts and to start delisting authors if they're suspected of using AI content through watermarks or user reports.
Re: (Score:2)
Youtube slide-show speech-generated channels that just read off articles have been a spam problem since long before LLMs and AI image generation. The latter two are not enabling factors for that form of spam.
People are, and always will be, online giving tutorials and instructions on how to X, where X is "something that people want to know how to do / get better at doing".
Re: (Score:2)
These channels were AI.
And yes YouTube is full of how-to tutorials on how to generate AI garbage. Just type "how to make ai channel on youtube" (or Amazon / Etsy / Whatever) and you will see it all. There is no reason they should tolerate this stuff when the intention is to turn their platform to shit and profit from ad impressions. It's fraudulent activity and tutorials encouraging fraudulent activity. At the very least they should prohibit or demonetize such tutorials and take down AI content on sight.
Re: (Score:2)
You're missing the point. These videos have existing long before AI was widely accessible. People have been hiring workers via Fiverr, Amazon Turk, and outsourced call centers to put together these videos for $0.40, with the hopes of each video generating $0.80 in ad revenue.
Big tech companies almost solely rely upon user-generated content as a mechanism for enabling unpaid scaling of content. They're not gonna give it up any time soon, just like Facebook is never going to admit how many of its users are
The Information Age is over (Score:2)
Re: (Score:2, Informative)
No but they can get sued if they are caught copying (e.g. https://en.wikipedia.org/wiki/... [wikipedia.org] ). Current AI tools don't normally care about such stuff.
Nob an on AI (Score:2)
Cyan Worlds (developers of Myst) released their new game on Steam (Firmament) a few months ago. They disclosed they had used AI generated content to add some minor world flavoring. For example in one section of the game there is a series of portraits and news articles; these are AI generated.
This game up on Steam, presumably because it does not have the issues Valve objected to in the indie dev's game.