Hacker News Reader

Comment by MagicMoonlight | Hacker News Reader

That seems pretty sensible. There have been lawsuits from artists before. Do you want to risk a game selling 10m copies and then it turns out that all the art was just copied and pasted by the "AI" and Valve is now on the hook.

Also from a store perspective, any game where shortcuts like this are used tend to be shit games. They don't want spam games to be pumped. There's already enough indie trash platformers that nobody wants.

View on HN

There are already successful Steam games known to have used AI art. And this is only in the cases of the developers publicly admitting that.

High On Life used Midjourney for it's ingame posters - https://store.steampowered.com/agecheck/app/1583230/ - https://www.thegamer.com/high-on-life-ai-generated-art/

And Firmament https://store.steampowered.com/app/754890/Firmament/ - https://www.pcgamer.com/firmament-ai-generated-content/

So is Valve now going to remove these games off the store? This seems like a very terrible way to handle this - they need to make clear rules and make a public statement, not just start banning apps that they sense use AI art.

No, the developers just need to “affirmatively confirm” that the own the copyright on all the works in Midjourney’s training set, and they are good.

I'm sure some indie studios will sign whatever, but as soon as a large studio uses an public model Steam will have to roll over on this one.

Can you explain to me how "you must affirmatively state that you own or have licensed rights to the training data (and if you're lying, the legal responsibility is yours and not Valve's)" is not a clear rule?

And yeah, they should kick those games off for using copyrighted materials that they do not own.

This is a rule developers are just finding out now from a game getting rejected. Pretty major deal if multi-million dollar budget games like High On Life should now be banned (even worse if they don't ban it now, making the rules unclear). It should have been a public statement, with a clear change to their developer terms.

> Also from a store perspective, any game where shortcuts like this are used tend to be shit games. They don't want spam games to be pumped. There's already enough indie trash platformers that nobody wants.

I find this hard to agree with. A game engine is a "shortcut" too, I can imagine people saying at some point anything developed with Unity would "tend to be shit games".

Associating quality with visual fidelity anyway is wrong, look at Terraria, I'm pretty sure anyone semi competent with AI generation could produce better assets, but it wouldn't help them produce a better game.

People will use gen AI art in good games, and people will use gen AI art in terrible games.

> Also from a store perspective, any game where shortcuts like this are used tend to be shit games.

Yeah, they indicate that they have already submitted multiple games with AI generated assets, and submitted this one "with a few assets that were fairly obviously AI generated." Maybe I'm being unfair and they are making really good games, but these are not good indicators to me.

These tools have only now become available. I can imagine a game where these tools are used to develop a large world and backstory previously not possible. The main images and text may be hand crafted, but you might walk down a street where the other building are all unique, have names, and descriptions. It could really flesh out some of the procedurally generated games out there. Or it could be terrible. Or it could be good for being so terrible. It shouldn't be rejected entirely just yet.

It's not being "rejected entirely". That is mendacious editorializing. Generative AI products are being rejected unless you affirm that you have rights to the entirety of the training data set.

What's wrong with that?

That you probably don't need rights to the training set in the US, unless Congress changes copyright law. This is being litigated.[1]

Humans can look at a collection of copyrighted images and draw a new picture. The legal basis for holding AIs to a higher standard is weak.

Current litigation: [1]

[1] https://www.theverge.com/2023/1/16/23557098/generative-ai-ar...

> That you probably don't need rights [...]

And if Valve doesn't want to take this risk? To reiterate eropple, what's wrong with that?

> That you probably don't need rights to the training set in the US

Even if that is true--and it's not sure to be--Valve is within their rights to demand additional coverage.

> Humans can look at a collection of copyrighted images and draw a new picture. The legal basis for holding AIs to a higher standard is weak.

Horseshit and worse words. Computers aren't people. They create derivative works from pushing inputs through mathematical models. The inputs are unerasable and the claims to the otherwise by the AI hustler class exist only to be able to profit off human effort without paying for it.

> Valve is within their rights to demand additional coverage.

As a near-monopoly gatekeeper, Valve is vulnerable to antitrust charges.

FTC is apparently about to go after Amazon.[1] US antitrust policy did far too little for too many years. EU competition policy is more aggressive.[2] The EU competition authorities already fined Valve for selling geo-blocked content that only worked in some EU countries. That's a violation of the basic Single Market rules of the EU.

[1] https://arstechnica.com/tech-policy/2023/06/ftc-prepares-the...

[2] https://competition-policy.ec.europa.eu/sectors/ict/cases_en

AI doesn’t copy and paste art though, it generates new art based off of patterns it’s seen in its training material. If it’s training material heavily features red squares, and you prompt it to generate a new piece, chances are you’ll get something with a red square - not because it copied that red square from any particular piece, but because it was a common element. There’s a difference between reproducing common elements in pursuit of adhering to a style, and copying and pasting.

https://arxiv.org/pdf/2301.13188.pdf

Stable Diffusion spits out slightly blurrier versions of the pictures in its training set.

But that’s not the default behaviour of these models at all. Instead you have to work pretty hard to extract these original images.

When using the models normally, they do generate new content, even if the style and subjects are of course interpolated from training data.

It can do that for some images, and that's a training mistake.

There's no way it can do that for more than a tiny fraction.

"we bias our search towards duplicated training examples because these are orders of magnitude more likely to be memorized than non-duplicated example"

If you deliberately try to make it do so. Even then it usually doesn’t work.

Reply on news.ycombinator.com