Generative AI art was used to create illustrations for its upcoming expansion, and it sounds like the developer is all-in on using it for its future output.
I work in generative AI, specifically curated training sets.
The issue is training on “licensed materials”. If that happened with all AI, no one would have a problem. But its disingenuous to suggest that’s how most AI is currently being trained. A lot of materials have been scraped off of the web, especially for image generation, meaning some portion of the training data was used without the author’s consent or, often, even their knowledge. It’s important to note that scraping training data in this way usually breaks a TOS.
The amount of people I’ve seen supporting AI usage in this context is staggering, with one commenter even telling me it was about the “greed” of the artists, whose work may be in a training set without consent, wanting royalties for slightly changing a parameter with their art (that is, of course, a strawman fallacy).
To me, the only issue here is handling the ethics of what goes into training data and what doesn’t. Authors should have the choice of their materials not being used. Adobe understood this, which is why Firefly being trained on explicitly licensed materials makes it a different beast, to which you allude.
But it’s clear a lot of people don’t understand why using data without consent is a bad thing in this context, and for that reason, some other people will choose not to support companies using it until the issue is resolved. It seems quite reasonable to me.
People usually say that’s the issue, until you show them that it’s possible to generate images and whatnot from models trained on “fully licensed” data. Then they come up with some other reason why evil AI is awful and evil. I’ve been involved in these debates for a long time now and those goalposts have well-worn tracks from how frequently they shift that way.
But it’s clear a lot of people don’t understand why using data without consent is a bad thing in this context
No, they don’t agree that using data without consent is a bad thing. Saying “they don’t understand” it is begging the question, in the literal sense. You’re saying that people who disagree about that are simply being ignorant of some underlying “truth.”
No, they don’t agree that using data without consent is a bad thing.
If this developer doesn’t mind taking data without consent, I hope they don’t have an issue with people pirating their game. That’s a slippery slope if I ever saw one.
“Slippery slope” is also a fallacy. Training an AI and copying a game are two different things and it’s entirely reasonable to hold the position that one is ok and the other is not.
I’m not missing the point. Just because they’re both “using data without consent” doesn’t mean they’re the same thing. Playing baseball and smashing someone’s car both involve swinging a bat but that’s where the similarity ends.
There are many ways that you can “use data without consent” that are perfectly legal.
Legal does not necessarily equate to ethical. And the law will eventually change (I think) to mitigate some of these shortcomings that AI training has highlighted.
Of course not. But “ethical” is a matter of subjective debate. You say X is unethical, I say X is ethical, and ultimately there’s no way to tell who’s “right.”
Law’s different, the whole point of it is to have a system that sorts these things out.
And the law will eventually change (I think) to mitigate some of these shortcomings that AI training has highlighted.
So it’s not currently illegal to train AIs like this? That’s been my point this whole time. It’s a different thing from the things that are currently illegal (such as “theft”).
Currently legal, but unethical. I never claimed it was illegal. (I did mention that scraping usually breaks a TOS, but that’s definitely a legal grey area and moot if its publicly accessible data)
I work in generative AI, specifically curated training sets.
The issue is training on “licensed materials”. If that happened with all AI, no one would have a problem. But its disingenuous to suggest that’s how most AI is currently being trained. A lot of materials have been scraped off of the web, especially for image generation, meaning some portion of the training data was used without the author’s consent or, often, even their knowledge. It’s important to note that scraping training data in this way usually breaks a TOS.
The amount of people I’ve seen supporting AI usage in this context is staggering, with one commenter even telling me it was about the “greed” of the artists, whose work may be in a training set without consent, wanting royalties for slightly changing a parameter with their art (that is, of course, a strawman fallacy).
To me, the only issue here is handling the ethics of what goes into training data and what doesn’t. Authors should have the choice of their materials not being used. Adobe understood this, which is why Firefly being trained on explicitly licensed materials makes it a different beast, to which you allude.
But it’s clear a lot of people don’t understand why using data without consent is a bad thing in this context, and for that reason, some other people will choose not to support companies using it until the issue is resolved. It seems quite reasonable to me.
People usually say that’s the issue, until you show them that it’s possible to generate images and whatnot from models trained on “fully licensed” data. Then they come up with some other reason why evil AI is awful and evil. I’ve been involved in these debates for a long time now and those goalposts have well-worn tracks from how frequently they shift that way.
No, they don’t agree that using data without consent is a bad thing. Saying “they don’t understand” it is begging the question, in the literal sense. You’re saying that people who disagree about that are simply being ignorant of some underlying “truth.”
If this developer doesn’t mind taking data without consent, I hope they don’t have an issue with people pirating their game. That’s a slippery slope if I ever saw one.
“Slippery slope” is also a fallacy. Training an AI and copying a game are two different things and it’s entirely reasonable to hold the position that one is ok and the other is not.
You’re missing the point. Both are using data (work of the dev on a game, work of an artist on art) without consent.
I’m not missing the point. Just because they’re both “using data without consent” doesn’t mean they’re the same thing. Playing baseball and smashing someone’s car both involve swinging a bat but that’s where the similarity ends.
There are many ways that you can “use data without consent” that are perfectly legal.
Legal does not necessarily equate to ethical. And the law will eventually change (I think) to mitigate some of these shortcomings that AI training has highlighted.
Of course not. But “ethical” is a matter of subjective debate. You say X is unethical, I say X is ethical, and ultimately there’s no way to tell who’s “right.”
Law’s different, the whole point of it is to have a system that sorts these things out.
So it’s not currently illegal to train AIs like this? That’s been my point this whole time. It’s a different thing from the things that are currently illegal (such as “theft”).
Currently legal, but unethical. I never claimed it was illegal. (I did mention that scraping usually breaks a TOS, but that’s definitely a legal grey area and moot if its publicly accessible data)