These are all fair questions, and speak to a difference in approach between US and EU regulators. The EU gets a lot of criticism (from US commentators) of the fact that the GDPR and DMA legislation leave a lot to interpretation. It always comes across to me as people wanting to know the hard lines so they can get as close to them as possible, where the intention seems to be to ensure that companies meet the "spirit" of the law, not the "letter" of the law.
According to the Wikipedia Article on Copyright Law in the United States:
> The United States copyright law protects "original works of authorship" fixed in a tangible medium, including literary, dramatic, musical, artistic, and other intellectual works. This protection is available to both published and unpublished works. Copyright law includes the following types of works: Literary,
Musical,
Dramatic,
Pantomimes and choreographic works,
Pictorial, graphic, and sculptural works,
Audio-visual works,
Sound recordings,
Derivative works,
Compilations,
Architectural works.
If there are confusing edge cases, a judge will decide what counts, since definitions are subject to interpretation. But for the most part, this stuff is pretty clear cut and common sense.
None of this is to say I agree with this new legislation, I just don’t like when people on HN act as if the law falls apart in the face of freshman philosophy questions.
I used to be an artist and then studied science. The first constitution I read was that of the USA, then that of Ireland. I've been reading law since. The decisions of judges are of variable quality and always susceptible to appeal, especially to autonomy. And the Law always falls apart in the face of the philosophy of freshmen, it's how we hold it together in the face of each other that makes it worthwhile.
Absolutely terrible idea. The idea that making a copy of art is somehow theft has lead to nothing but bad outcomes, artificial scarcity, and whining from wealthy institutions complaining about their imagined profits. Intellectual property is one of the most distasteful oxymorons of our age.
Not all artists want their art taken and distributed or modified endlessly.
Regardless, I think it seems important to create a system for tracking attestation of AI produced materials, especially with regard to disinformation campaigns. This may not be the sole/express purpose of the legislation, but we will need systems of attribution in the future to determine what's human made and what's not.
FWIW I don’t want my mind actively polluted by the media and materials derived by people who consider art “content” and all human output “training data” to support their own misguided gambles in the hopes of saving their personal portfolios from self-imposed disaster. And yet here we are. So. I hear you.
The point is, people shouldn't have the right to shove something in front of you that might inspire you to iterate, remix, or derive from that, while penalizing you if you act on those ideas. Intellectual property is mind-policing. Its success hinges entirely on it being allowed to invade public space without the public's consent, while using the state to forbid its own free distribution by the public.
No one can opt-out of copyrighted content, and if they could, copyrighted works would become the slim minority of all content, or cease to exist entirely.
AI doesn't have a will. AI doesn't make art. I think the real conversation is about harm. No one (nice) wants harm to be the outcome at the expense of others. The problem is identifying the harm and preventing that harm. This legislation is about the thought virus equating digital "things" with physical "things". They are not the same. We are deeply prone to skeuomorphism but we should fight this urge so we can appropriately address real harm and that is sadly being missed entirely by these politicians.
If you don't want to have others distribute and modify your art, never show it to anyone else, put it in a box, and bury it five metres under your house.
Even aside from any technology angle, the human mind is fundamentally very good at copying and modifying information-- it's why we can have a coherent cultural zeitgeist rather than individual siloed units of Capital-A Art. You can aspire to be part of that, and be respected for your contribution to it, but expecting everyone else to hobble their potential in service of your fetish for purity is self-indulgent if not downright sociopathic.
TBH, I feel like AI is eventually going to be slotted in as a tool rather than something magical. I can't draw a straight line or a circle, but we don't feel the need to permanently stamp pictures as "he used a straightedge and compasses!" (or "He used Krita!" these days) The ability to make lots of fairly realistic stuff cheap may change some value propositions (see: portrait painters when photography took off) but eventually we'll get past the initial panic phase.
Seems like ambitious legislation to request something from a standards agency and then keep lawing as if said standard existed.
The Content Origin Protection and Integrity from Edited and Deepfaked Media Act (COPIED Act) would direct the National Institute of Standards and Technology (NIST) to create standards and guidelines that help prove the origin of content and detect synthetic content, like through watermarking. It also directs the agency to create security measures to prevent tampering and requires AI tools for creative or journalistic content to let users attach information about their origin and prohibit that information from being removed. Under the bill, such content also could not be used to train AI models.
I don't know if I'm reading this correctly. If this applies to images, and if these watermarks are something like a comment stored in the file's metadata, wouldn't this effectively make it illegal to make a thumbnail of an image? Or to screenshot an image? Or convert it to a .bmp?
> The bill also directs NIST to develop cybersecurity measures to prevent tampering with provenance and watermarking on AI content.
So somehow NIST would have to develop a cybersecurity measure to stop me from making an AI generated image and just screenshotting it, discarding any metadata. I'm not sure how realistic this is.
If the watermark were embedded into the actual pixels of the image then of course it'd be preserved by a screenshot, but my understanding is that those watermarks are not very robust to the image being resized, compressed or cropped
At the end of the day you can also just take a picture of an AI-generated image on your screen using a camera (bonus points if you're in a future where cameras embed digital signatures to verify the image is authentic and not AI generated)
Lots of people use commercial software they didn’t pay for.
But corporations have too much to lose and little to gain by not paying for the software they use, so following the law is legal insurance for them. And so commercial developers get paid by significant users.
Nobody expects proposed laws like this to have much impact on what individuals do with their screenshots and cameras. The hope would be that it curbed mass unpermissioned use.
> Such content also could not be used to train AI models.
And... how would you make sure it won’t happen?
I don’t think they know how the digital world works, let alone generative AI. You might regulate big companies, but never the community, not to mention companies that are outside of the US that can still access and train their models on such materials.
22 comments
[ 3.5 ms ] story [ 57.8 ms ] threadWhat's an AI watermark?
What's a work of art?
How inundated by works containing supposedly unique digital signatures nonetheless susceptible to collision attacks do they wish to suffer?
The Law cannot match Mathematics in formulation, nor can it match it in meaning.
I wish mathematicians spent more time educating legislators so the rest of us can think and communicate in peace and clarity.
According to the Wikipedia Article on Copyright Law in the United States:
> The United States copyright law protects "original works of authorship" fixed in a tangible medium, including literary, dramatic, musical, artistic, and other intellectual works. This protection is available to both published and unpublished works. Copyright law includes the following types of works: Literary, Musical, Dramatic, Pantomimes and choreographic works, Pictorial, graphic, and sculptural works, Audio-visual works, Sound recordings, Derivative works, Compilations, Architectural works.
If there are confusing edge cases, a judge will decide what counts, since definitions are subject to interpretation. But for the most part, this stuff is pretty clear cut and common sense.
None of this is to say I agree with this new legislation, I just don’t like when people on HN act as if the law falls apart in the face of freshman philosophy questions.
I really feel I'm doing a service to the people receiving the books and anyone they hand the books off to . . .
Regardless, I think it seems important to create a system for tracking attestation of AI produced materials, especially with regard to disinformation campaigns. This may not be the sole/express purpose of the legislation, but we will need systems of attribution in the future to determine what's human made and what's not.
No one can opt-out of copyrighted content, and if they could, copyrighted works would become the slim minority of all content, or cease to exist entirely.
Even aside from any technology angle, the human mind is fundamentally very good at copying and modifying information-- it's why we can have a coherent cultural zeitgeist rather than individual siloed units of Capital-A Art. You can aspire to be part of that, and be respected for your contribution to it, but expecting everyone else to hobble their potential in service of your fetish for purity is self-indulgent if not downright sociopathic.
TBH, I feel like AI is eventually going to be slotted in as a tool rather than something magical. I can't draw a straight line or a circle, but we don't feel the need to permanently stamp pictures as "he used a straightedge and compasses!" (or "He used Krita!" these days) The ability to make lots of fairly realistic stuff cheap may change some value propositions (see: portrait painters when photography took off) but eventually we'll get past the initial panic phase.
The Content Origin Protection and Integrity from Edited and Deepfaked Media Act (COPIED Act) would direct the National Institute of Standards and Technology (NIST) to create standards and guidelines that help prove the origin of content and detect synthetic content, like through watermarking. It also directs the agency to create security measures to prevent tampering and requires AI tools for creative or journalistic content to let users attach information about their origin and prohibit that information from being removed. Under the bill, such content also could not be used to train AI models.
> The bill also directs NIST to develop cybersecurity measures to prevent tampering with provenance and watermarking on AI content.
So somehow NIST would have to develop a cybersecurity measure to stop me from making an AI generated image and just screenshotting it, discarding any metadata. I'm not sure how realistic this is.
If the watermark were embedded into the actual pixels of the image then of course it'd be preserved by a screenshot, but my understanding is that those watermarks are not very robust to the image being resized, compressed or cropped
At the end of the day you can also just take a picture of an AI-generated image on your screen using a camera (bonus points if you're in a future where cameras embed digital signatures to verify the image is authentic and not AI generated)
But corporations have too much to lose and little to gain by not paying for the software they use, so following the law is legal insurance for them. And so commercial developers get paid by significant users.
Nobody expects proposed laws like this to have much impact on what individuals do with their screenshots and cameras. The hope would be that it curbed mass unpermissioned use.
Good luck enforcing that in open source tools.
> Such content also could not be used to train AI models.
And... how would you make sure it won’t happen?
I don’t think they know how the digital world works, let alone generative AI. You might regulate big companies, but never the community, not to mention companies that are outside of the US that can still access and train their models on such materials.
I thought the headline was indicating it would be illegal to remove watermarks from content generated by LLMs.