Non or this makes much sense tbh. The thing about this is that when you commission an artist, the artist is not actually given copyright, but the person who paid the artist is given the copyright. Why is it not the same for promoting?
I think you mean prompting. Commissioning an artist in the context you're thinking of implies a written contract, signed by both parties [0]. This is the first reason where it isn't the same, and I believe is enough to consider new laws in the situation.
Even if the new law just codifies the process you're thinking of, it's better to get ahead of the situation, or else we make a mess of things like DMCA, etc.
Is that actually true, absent a contract or agreement that you get the copyright along with the commission? I would think the artist gets it by "default" unless you're also paying for the rights or something?
IANAL, but as far as I know, the artist gets it by default in absence of a contract that specifically assigns it to the other party. It’s possible there are special rules around salaried employees but even then I’d expect almost all busineesses to have a specific condition of employment that the employee’s works are works-for-hire.
If an employee makes something at work, the employer is legally the creator of it.
If someone contracts someone else to make something for him, the copyright in the work vests in the hiring party essentially when it's part of a larger work and the contract says it's a work for hire. A one-off commission on its own does vest in the artist.
(Disclaimer: I'm a lawyer, but I'm not your lawyer.)
It depends, but yes, usually. Post-1978, the author of a work is also the copyright owner as soon as the work is fixed in a "tangible medium of expression" (written down, recorded, dictated, etc.). The exception to this is when the work was created as a "work for hire" (WFH) [1] for an employer. The rules around what constitutes a bonafide WFH are governed by agency law in the US. The primary court case in this area is The Community For Creative Non-Violence v. Reid which outlines a fairly high bar for a work to be considered a true WFH (it asks questions like: did you go to an office where the employer told you what to do and how to create it, did the employer provide you materials, did the employer pay you as a salaried employee rather than a contractor, how were you treated tax-wise and benefits-wise, etc.). That said, as the copyright owner at the moment of fixation, you can absolutely assign that copyright in writing to anyone you'd like. So functionally, there is usually a backup clause saying something like: "this is a work for hire for us, but, in the event it's deemed not a work for hire, you hereby assign the copyright to us anyway."
That was Thaler's argument. He had the first prominent case in this debate. AFAIK, the artist implicitly gets copywrite to the work they created then your contract with the artist transfers that copywrite to you.
Thaler tried to argue that he and the AI were in a situation similar to "work for hire" with an artist, but there was no human input to the AI. But, the copyright office holds that non-human entities/events are not granted copyright protection and therefore have no rights to transfer. So, they rejected his claim.
the artist is not actually given copyright, but the person who paid the artist is given the copyright.
This is only true if it is written in a contract. In general, the artist keeps the copyright. Photography is the same way: Photography studios own the copyright to wedding pictures, school pictures, family portraits, sports functions, and the like. If you pay a photographer to take a family picture, you have to get permission to use that picture for a greeting card unless you've signed a form giving that permission to you.
Similar rules apply if I sell some artwork that wasn't commissioned. Someone else might own the original, but only I can make prints and post a picture where I choose. Unless, that is, we have a contract that says otherwise.
I had to make an account just to come in here to tell you you are not correct.
Copyright is held automatically by the person who produces the work unless otherwise specified in a contract or agreement. I am a commission artist, and I retain all copyright to all of my work, per my terms of service, unless otherwise sold via agreement, and I charge extra for copyright.
Someone can't commission me to paint something, and then turn around and sell prints of it themselves. I still did the work, and they need my approval before they can legally make money off of it.
The reason why AI art is so devisive right now is because thousands of artists have had their work scraped and the AI trained on their work without their permission or compensation, and unlike a human who looks at other art for reference or inspiration, an AI can look at a dataset and pump out thousands of similar images in a very tiny amount of time.
If the AI author or the user generating the image owns copyright, you can kiss artists selling their own work goodbye. The space will be flooded with "AI artists" who create work at a far faster pace than anything a real artist can create, for a fraction of the cost. The people who train AI on this art are profiteering off of the hard work of the people who spent years of their lives developing a skill.
This isn’t true; it’s entirely based on the contract. E.g. I believe wedding photographers usually make contracts to retain all copyrights. They either sell the couple prints of particular photos, or sell them a license to make more copies (which will cost more than the typical photo order).
Because in the prompting case, the artist you commissioned isn't a human. Existing case law says you can't have non-humans holding copyright, because only humans can be creative, and copyright requires creativity. In the case of works for hire, the human artist is contributing creativity so the commissioner can steal[0] it for themselves. But with AI art, there's no creativity to steal, so there's no copyright.
[0] My editorialization. The rationale of what is and isn't a work-for-hire was not some principled decision but just a list of whoever wasn't in the room when the last major copyright law was drafted.
It is a complicated subject with no clear resolution, IMO. Train people on a bunch of books from a local library and noone has an issue. Train a generalized AI on the same thing, and the legal result should be the same.
But what about an imaginary specialized AI that really just focuses on reproducing its input training material in a nearly verbatim way?
The current reality is somewhere in between these two extremes, and defining the boundaries sounds difficult.
The output should certainly not be copyrightable if it does not produce an existing copyrighted work or if fair use is in play, which is why many people seem to be attacking the training set itself, an argument which is also complicated but has little merit.
Sure, but the line between AI and a database can be blurred, even intentionally. Even with ChatGPT, it can be coerced to deliver parts of its training set nearly verbatim.
Newer training techniques likely removed most of that by simply deduplicating training data. All the papers I read on verbatim techniques could only find items that were overpopulated during training.
What if that output is created by mixing in copyrighted material?
e.g.: What happens to your MIT licensed code when an AI assistant generates some code for you which contains some GPL, BUSL or Source Available licensed parts? Moreover, do these code bases scraped with consent of the owners?
The image itself is copyrighted, or it isn't. Fair use may or may not come into play for things like replication of known copyrighted characters, depending on commercial use.
None of this has changed from the traditional copyright process. Even 15 years ago I could find some of my old works rehosted on obscure Chinese websites, and my options for recourse were the same as they are today. Many of my own works were already derivative.
Go ahead and try; A lot of this kind of "gotcha" commentary comes from a lack of understanding about existing copyright law. Let me know how far you get.
1. The local library at least paid for those books. Which is more than the creators of the current AI generation can say[1].
2. You can't train people exclusively on books. If you asked someone to explain a concept they learned in those books, sure that knowledge will help, but their response will also affected by every moment in their life so far. They're bringing other unique things to their response.
It doesn't seem difficult to me. It seems like the intent is the big difference. Purposefully making a copy of something digitally already has rules around it, which was the second extreme describes
I think the copyright of the training material should be propagated, and such a system can not be allowed to compete in the market without attributing and recompensating the authors. ;)
So every DALL·E, Stable Diffusion, and Firefly image should be jointly and equally owned by millions of artists? All of GPT's and LLaMA's outputs should be jointly and equally owned by essentially everyone who has ever written publicly online?
I can see how compensating the original artists would be desired, but I can't see how that wouldn't completely kill the AI space. There's just no way to develop a suitable set of training data from only people who have signed a licensing agreement.
Edit: And to be clear, CC BY-SA would not be a suitable licensing agreement, so things like Stack Overflow and Codidact and Wikipedia would be completely off-limits.
I'd like to express my shock at reading something so uninformed and poorly thought out in a place like HN.
HN goes through its phases; just as the moon waxes and wanes, but this anti-AI, primitivist phase is quite annoying.
If you do not seek progress for humanity and the construction of things that are more meaningful. If you do not have the hacker spirit and want to explore. If you don't want to create just to see what could be... Maybe you're in the wrong place.
.
And make no mistake, this is poorly thought out.
This technology is never going away. Trying to ban generative AI is an attempt in banning math. Far more powerful people have tried for far more compelling reasons to do the same for a very long time. And when all was said and done, their crusade came undone.
.
The anti-AI arguments quickly crumble when you start asking questions like,
What was the state of artists before this? Did all artists enjoy a high standard of life in every part of the world? Or, were they barely scraping by on Fiverr?
In a world where copyright is required to train generative AI, who will be the only ones who could train any meaningful system?
What happens if we forbid people from learning and building on art that is already copyrighted?
What happens when people can copyright style? If only one band owned the copyright on Grunge and that whole Nirvana thing was just a non-starter?
What happens if non-human entities (like a LLC or S-corp) can't meaningfully participate in artistic creation?
.
Based on how the law currently stands, it is the EFF's estimate that there is no way to modify the law to forbid a system from learning and creating derivative works without opening a can of worms and impacting human creativity.
Fundamentally, the point of creative culture is that we add onto each other and that we stand on the shoulders of giants. An artist doesn't have to figure out how to draw in perspective because all the clever people before us have already thought it through for us. An artist doesn't have to figure out how to soften light, because others have done it before.
All culture is fundamentally additive and intellectual property is an artificial restriction that has caused broader cultural stagnation by making sure that nothing comes into the public realm for a hundred years.
The primitivists want to give this system even more power so that now you can't even see the works and be inspired from them. That's a take that's antithetical to HN's ideas of intellectual curiosity.
The problem isn't with tools being developed, but how they are going to be exploited in the current system that the world exists in. There is outrage for AI generated art because it:
A) is trained on work from existing artists without consent or permission
B) will be used to push artists out of their ability to make money to feed themselves
In a perfect world where artists could sustain themselves while practicing thier craft, this wouldn't be such a devisive issue. If you want to think about things from a counter-culture "hacker" perspective, we should be talking not about limiting AI but how we should be elevating artists to be able to exist despite the advancement in technology removing their jobs.
Its insane to think that artists won't be upset that their entire field, one centered on years and years of practice and learning, is about to be disrupted by a program that can create a thousand works in the span of a few minutes, trained on their own hard work without permission. Especially when these people already often work lower-paying jobs in order to make ends meet so they can work on their passion.
Is it really "hacker" to expect people to have to give up their happy lives doing work they enjoy so they can be shoved into the meat-grinder of a 9 to 5 job, because a program replaced them? Since when is it counter-culture to replace the leisure aspects of life with tech just because we can?
The tools themselves aren't evil and can be very helpful with artists, but it's not insane to see why its receiving push back from the art community who stand to lose the most from it.
Add in the fact that many of the people using or pushing for these tools don't even respect the craft of artists they're stealing from.
Not only in terms of not being interested in understanding the effort taken to develop the craft, but also in terms of understanding aspects of the craft itself. Whenever someone brings up the idea that ML is the same as a human learning from referencing other's art, I feel they giveaway their lack of any genuine exposure to art as that's a purely technical interpretation which misses that very little of art is about pure technical skill.
In a way, I feel the AI imagery stuff has just brought enshittification to art. It's the same overall pattern where some tech"bro" comes in citing disruption, technical legality and the future without any interest in understanding what they're talking about, develops a technology that's good enough for attractive marketing but not actually sufficient for the field, but they bring in so many new people that there's no competing and combined with the shit management from the "techbro", the industry is forced to settle for a new normal that grows ever shittier until all the value has been extracted from it.
I used to be very interested in working on generative AI, it's what got me working on a PhD. But I completely lost interest the moment I saw how irresponsible everyone was with it. The tool is fine, but I am increasingly of the opinion that it's a tool we as a society are not ready to ethically wield.
This is absolutely it. I'm a tech lover at heart, I am constantly playing with all the newest stuff that comes out. And when I first saw some of the things people were doing with Stable Diffusion I was very interested. Quickly mocking up textures for simplistic scenes? Quickly filling in repetitive background elements or flat coloring artwork to save artists time? Smart fill, smart lighting, these are some powerful tools, especially for concept art or speeding up commission work.
But the first thing I saw people do is start generating "good enough" art to get their ideas out, and as an artist, I saw how shit the art looked - but it was enough for the people who were not artists. I have seen quite a few friends lose long-time commissioners who are just generating tons of ideas with clicks.
I don't lament people's ability to generate ideas or become invested in art with tools that make it easy for them. I lament the corporate idiots that are going to try to make fully AI generated movies and paintings and books, and flood the market, kill all the real artists, and yeah, enshittify our art and media.
This is HN. How many computer programs have been created based on existing professional workflows that vastly reduced the amount of people needed to do that job?
An absolute ton. Now that it’s happening to art instead of accountants everyone should care?
Absolutely, because art isnt the same as crunching numbers. No accountant created the math they are crunching and then had a 3rd party come in and vacuum up their work and spit out a computer that creates a facimile of their work to replace them.
You pretty much just described Excel as a whole, not to mention any kind of formula/macro/whatever that gets entered into it. Yeah, they didn't come up with those formulas entirely on their own - they adapted what they learned to their current needs.
Artists also didn't come up with their art style entirely on their own, they adapted what they learned to their current needs.
You never agreed to teach anyone in particular. By publishing something, you agree to people using them fairly though. I don't see on its own a reason that teaching people with a work is considered fair use but training a model is not.
I don't think something being hard, should mean it just doesn't have to happen. "It's too hard to figure out whose work we used, so we just won't?" I also think it's unfortunate that writers, bloggers, photographers, etc. who have seen no financial compensation for their work will watch it go on to enrich a company with a billion dollar valuation.
It costs $0.091 per copy to "cover" a song in the USA, you get this by securing a "mechanical license" and that rate is set by the Copyright Royalty Board[1]. I don't see why there can't be a similar way forward here.
Hmm, I hadn't considered that, but I suppose Congress could enact a similar compulsory licensing procedure for AI. The hardest part will be identifying authors though. I can certainly appreciate adherence to the law despite difficulty, but I think once it crosses impossibility, you have to reconsider the public policy benefits. Even if you could force someone to license you their work for a statutory fee, something like Wikipedia would still be off-limits since you could never contact all the anonymous Wikipedia users to send each of them his 4¢ royalty.
Royalty rate increased to $0.12/copy as of January 1, 2023 for songs under five minutes in length (the rate increases from there for longer songs). [1][2]
My view is that if AI is trained on public data and it doesn't need to pay royalties for it, then it doesn't make sense that the public should pay royalties to use the AI's output.
If AI companies paid royalties for their entire training set, then it would be a different situation.
But given that it may be infeasible to prove or disprove whether or not AI was trained on public data, we should err on the side of caution and not allow the companies which build/operate AI to have copyrights over its output.
A case could be made that most work which humans produce is itself derivative of various public works and therefore the AI is not so different from humans and therefore it should be given copyrights... But I think the key difference is that the AI is not human, there are 'human rights' but there are no 'AI rights' or 'machine rights'... Until an AI can prove that it has agency and that it works for its own interests and not for a third party, it should not have any rights.
I feel like it's companies' problem if their business model is not compatible with the new reality of AIs relentlessly scraping their data. They may need to rethink their business model and try to increase their margins and/or improve their efficiency to account for rising hosting costs. I guess there are difficult times ahead.
I think it's not so bad since the entities who train and/or use AI also have to pay a lot of money on their end.
Alright - what if I take Stable Diffusion and further train it on my own art, and then use that to generate things?
Or what if I don’t do that, but generate several images and spend hours making a composite? What if I start with my own photograph and generate things to use in a composite on top of that?
It is clear cut because the Stable Diffusion AI you refer to was nevertheless trained on public data to get to that point were it could then be 'fine-tuned' with your own art. Your own art is just a small addition to its much, much larger input/training set. The size of these public training sets are truly massive.
If you start with a Stable Diffusion program from scratch with all weights initialized at random (or whatever weights its initialization algorithm requires) and you train it on only your own photograph(s), then you will almost certainly find that it's not effective at all. In that case though, the low-quality output would be 100% your own.
The reason why generative AIs are so good is precisely because they have access to a large body of public data. Without the public data, they would not be good no matter how good the underlying software/algorithm is.
> Alright - what if I take Stable Diffusion and further train it on my own art, and then use that to generate things?
Operative word there: "further". If it's already trained on everyone else's input, then without that there wouldn't even be a Stable Diffusion for you to "further train on your own art", so YTF should you get sole ownership of the sum of it all?
What you're saying is like "Alright - what if I take Linux and add in my own code, and then sell that as closed source?" Nope, it's already GPLed, and so should AIs trained on public data be.
No, I would hate this regulation and this is everything against what open internet stood for. I can see that you don't want AI companies to exist so you want this regulation, but how is different than search engine scraping the internet? If the content is not behind TOS or paywall, it should be equally scrapable by anyone be it search engine or AI companies or me.
If the content is copyrighted, it should only be shown if user accepts the terms or pays for it.
> I can see that you don't want AI companies to exist so you want this regulation, but how is different than search engine scraping the internet?
This is unnecessary hyperbole.
Preventing the laundering of copyright material through AI algorithms does not prevent AI companies from existing. Take Adobe -- they have license to the training images they use and they have helpful products based on AI processing of those images. It won't be "free" money from dubious sources. But AI does not depend on the laundering of copyright material.
The difference is whether or not you are scraping and then generating similar products (eg the images themselves) or whether you are scraping for some other purpose.
I believe there are rules relating to copyright which allow using up to a certain number of words of text from an original text without infringing on copyright. I think a difference with search engines there is that search engines reference the source with a link whereas the AI does not. Also, critically, search engines do not claim copyright over the text shown in their results pages.
I think that forms the legal basis for why AIs should be allowed to train on public data for free but it does not form a legal basis for AI to have copyrights over the output.
These calls for public comment are really important to good government. If you or someone you know will be affected by the potential for new regulations, get the word out to leave comments. All are read, considered, and become part of the public record. Best if concerns and potential remedies are stated in a neutral tone. Write like you’re talking to someone at a government desk.
Could someone help understand what is the difference between ml image generation and photography? Like, it seems that most arguments about one do trivially transfer to the other, so i am confuse
I think of LLM’s as a vast digital information landscape, vaguely analogous to walking through World of Warcraft, except at a higher dimensionality of existence.
A photographer doesn’t create nature, the subjects, the architecture, the universe, the billions of R&D behind the silicon sensors, etc., but they create the situation and push the button.
Creating the situation, I.e. drawing from the near infinite of possibilities, is analogous to prompting.
People have enforced copyright over photos, the Eiffel tower's lights are copyrighted, some tattoo artists try to enforce copyright over their work, and I'm pretty sure you can't just take a photo of someone's painting and sell it. Any photo of a pine tree is the exclusive copyright of the photographer, nature can't enforce copyright. A photo of a pine tree generated from hundreds of copyrighted photos of pine trees is a completely different question, I think your argument is too reductive.
You can not take a photo of someone’s art or likeness and sell it in many cases (or if abuse etc.), this is true and should be true of LLMs. In this case the photography analogy holds.
For the case of observing copyrighted photos (not directly storing indefinitely) and generalizing a pine tree from that data, I think that is fine:
-assume we believe and rule it is not fine, then we will just allow huge companies to generate their own data sets and they will own this space moving forward (would require google maps levels of investment). Also, we wouldn’t be able to tell the difference between illegal SD output versus clean FAANG output without extremely draconian measures.
Why? Photography is an established practice, with existing rules, so it seems like a good aid at reasoning about ml image generators.
E.g. by considering in which situations "your photo may be copyrighted by someone else" it might be easier to reason about cases when "your generated image may be copyrighted by someone else"
I think it is like photography. If an image is created by Dall-e then the copyright of that image should belong to the person that created the prompt and selected the output. Those are creative activities, much like pointing a camera and deciding which negatives to use going forward. Also, like photography, the equipment does not claim photography, be that cameras or background paper, etc.
If you take a photo of something copyrighted, that copyright may be enforced on the photo. If Dall-e is trained on loads of copyrighted material (which it is) then the argument is that it's output is owned by the people who's art it was trained on rather than the photographer who just took a picture.
One is a tool for converting analog images to digital images, the other is an opaque non-deterministic algorithm that produces images. What isn't the difference between them, aside from producing images?
Somehow I think AI will contribute more comments on this than people, much like the net neutrality comment stuffing fiasco a while back.
The reality is copyright has diverged so far from its intended purpose that there is no way to go back to the spirit of the law and rethink it because it is so fundamentally broken. So I anticipate some makeshift policy that nobody likes and gets broadly ignored because it is unenforceable.
Honestly, the best approach right now might be to simply say that AI generated content needs to declare that it is AI generated (somehow...) - that would at least give some context to the situation so that actual rules could have an effect down the road.
This is only going to have biased answers, because you would need to be passionate about ai content which is still a minority, to defend it. But there are a lot of uninformed people with opinions out there.
If you stopped calling it AI, and called these things what they really are, tools that predict tokens or pixels there would be a little less fear. Emphasis on tools, you wouldn't create law that didnt allow google search to exist. People think these function calls are thinking and anthropomorphizing them. You wouldn't make content from photoshop not copyrightable.
85 comments
[ 3.1 ms ] story [ 149 ms ] threadI think you mean prompting. Commissioning an artist in the context you're thinking of implies a written contract, signed by both parties [0]. This is the first reason where it isn't the same, and I believe is enough to consider new laws in the situation.
Even if the new law just codifies the process you're thinking of, it's better to get ahead of the situation, or else we make a mess of things like DMCA, etc.
0 - https://www.copyright.gov/circs/circ30.pdf
If an employee makes something at work, the employer is legally the creator of it.
If someone contracts someone else to make something for him, the copyright in the work vests in the hiring party essentially when it's part of a larger work and the contract says it's a work for hire. A one-off commission on its own does vest in the artist.
(Disclaimer: I'm a lawyer, but I'm not your lawyer.)
[1] https://copyright.gov/circs/circ30.pdf [2] https://en.wikipedia.org/wiki/Community_for_Creative_Non-Vio...
Thaler tried to argue that he and the AI were in a situation similar to "work for hire" with an artist, but there was no human input to the AI. But, the copyright office holds that non-human entities/events are not granted copyright protection and therefore have no rights to transfer. So, they rejected his claim.
This is only true if it is written in a contract. In general, the artist keeps the copyright. Photography is the same way: Photography studios own the copyright to wedding pictures, school pictures, family portraits, sports functions, and the like. If you pay a photographer to take a family picture, you have to get permission to use that picture for a greeting card unless you've signed a form giving that permission to you.
Similar rules apply if I sell some artwork that wasn't commissioned. Someone else might own the original, but only I can make prints and post a picture where I choose. Unless, that is, we have a contract that says otherwise.
Copyright is held automatically by the person who produces the work unless otherwise specified in a contract or agreement. I am a commission artist, and I retain all copyright to all of my work, per my terms of service, unless otherwise sold via agreement, and I charge extra for copyright.
Someone can't commission me to paint something, and then turn around and sell prints of it themselves. I still did the work, and they need my approval before they can legally make money off of it.
The reason why AI art is so devisive right now is because thousands of artists have had their work scraped and the AI trained on their work without their permission or compensation, and unlike a human who looks at other art for reference or inspiration, an AI can look at a dataset and pump out thousands of similar images in a very tiny amount of time.
If the AI author or the user generating the image owns copyright, you can kiss artists selling their own work goodbye. The space will be flooded with "AI artists" who create work at a far faster pace than anything a real artist can create, for a fraction of the cost. The people who train AI on this art are profiteering off of the hard work of the people who spent years of their lives developing a skill.
[0] My editorialization. The rationale of what is and isn't a work-for-hire was not some principled decision but just a list of whoever wasn't in the room when the last major copyright law was drafted.
Most holders of copyright are non-human entities like Disney.
The legal system considers groups of humans to be as equally protected as and have the same rights as individual humans.
But what about an imaginary specialized AI that really just focuses on reproducing its input training material in a nearly verbatim way?
The current reality is somewhere in between these two extremes, and defining the boundaries sounds difficult.
The output should certainly not be copyrightable if it does not produce an existing copyrighted work or if fair use is in play, which is why many people seem to be attacking the training set itself, an argument which is also complicated but has little merit.
e.g.: What happens to your MIT licensed code when an AI assistant generates some code for you which contains some GPL, BUSL or Source Available licensed parts? Moreover, do these code bases scraped with consent of the owners?
The image itself is copyrighted, or it isn't. Fair use may or may not come into play for things like replication of known copyrighted characters, depending on commercial use.
None of this has changed from the traditional copyright process. Even 15 years ago I could find some of my old works rehosted on obscure Chinese websites, and my options for recourse were the same as they are today. Many of my own works were already derivative.
So if I make a program that takes an image and changes a single pixel, I can use it to launder copyrighted images?
Here's what you need to know about derivative works in US law: https://www.copyright.gov/circs/circ14.pdf
2. You can't train people exclusively on books. If you asked someone to explain a concept they learned in those books, sure that knowledge will help, but their response will also affected by every moment in their life so far. They're bringing other unique things to their response.
[1] - https://www.nytimes.com/2023/07/10/arts/sarah-silverman-laws...
I can see how compensating the original artists would be desired, but I can't see how that wouldn't completely kill the AI space. There's just no way to develop a suitable set of training data from only people who have signed a licensing agreement.
Edit: And to be clear, CC BY-SA would not be a suitable licensing agreement, so things like Stack Overflow and Codidact and Wikipedia would be completely off-limits.
As if that's a bad thing
HN goes through its phases; just as the moon waxes and wanes, but this anti-AI, primitivist phase is quite annoying.
If you do not seek progress for humanity and the construction of things that are more meaningful. If you do not have the hacker spirit and want to explore. If you don't want to create just to see what could be... Maybe you're in the wrong place.
.
And make no mistake, this is poorly thought out.
This technology is never going away. Trying to ban generative AI is an attempt in banning math. Far more powerful people have tried for far more compelling reasons to do the same for a very long time. And when all was said and done, their crusade came undone.
.
The anti-AI arguments quickly crumble when you start asking questions like,
What was the state of artists before this? Did all artists enjoy a high standard of life in every part of the world? Or, were they barely scraping by on Fiverr?
In a world where copyright is required to train generative AI, who will be the only ones who could train any meaningful system?
What happens if we forbid people from learning and building on art that is already copyrighted?
What happens when people can copyright style? If only one band owned the copyright on Grunge and that whole Nirvana thing was just a non-starter?
What happens if non-human entities (like a LLC or S-corp) can't meaningfully participate in artistic creation?
.
Based on how the law currently stands, it is the EFF's estimate that there is no way to modify the law to forbid a system from learning and creating derivative works without opening a can of worms and impacting human creativity.
Fundamentally, the point of creative culture is that we add onto each other and that we stand on the shoulders of giants. An artist doesn't have to figure out how to draw in perspective because all the clever people before us have already thought it through for us. An artist doesn't have to figure out how to soften light, because others have done it before.
All culture is fundamentally additive and intellectual property is an artificial restriction that has caused broader cultural stagnation by making sure that nothing comes into the public realm for a hundred years.
The primitivists want to give this system even more power so that now you can't even see the works and be inspired from them. That's a take that's antithetical to HN's ideas of intellectual curiosity.
The EFF's take,
https://www.eff.org/deeplinks/2023/04/how-we-think-about-cop...
https://www.eff.org/deeplinks/2023/04/ai-art-generators-and-...
And so, using the mind and work of the collective, you give back to the collective.
A) is trained on work from existing artists without consent or permission
B) will be used to push artists out of their ability to make money to feed themselves
In a perfect world where artists could sustain themselves while practicing thier craft, this wouldn't be such a devisive issue. If you want to think about things from a counter-culture "hacker" perspective, we should be talking not about limiting AI but how we should be elevating artists to be able to exist despite the advancement in technology removing their jobs.
Its insane to think that artists won't be upset that their entire field, one centered on years and years of practice and learning, is about to be disrupted by a program that can create a thousand works in the span of a few minutes, trained on their own hard work without permission. Especially when these people already often work lower-paying jobs in order to make ends meet so they can work on their passion.
Is it really "hacker" to expect people to have to give up their happy lives doing work they enjoy so they can be shoved into the meat-grinder of a 9 to 5 job, because a program replaced them? Since when is it counter-culture to replace the leisure aspects of life with tech just because we can?
The tools themselves aren't evil and can be very helpful with artists, but it's not insane to see why its receiving push back from the art community who stand to lose the most from it.
Not only in terms of not being interested in understanding the effort taken to develop the craft, but also in terms of understanding aspects of the craft itself. Whenever someone brings up the idea that ML is the same as a human learning from referencing other's art, I feel they giveaway their lack of any genuine exposure to art as that's a purely technical interpretation which misses that very little of art is about pure technical skill.
In a way, I feel the AI imagery stuff has just brought enshittification to art. It's the same overall pattern where some tech"bro" comes in citing disruption, technical legality and the future without any interest in understanding what they're talking about, develops a technology that's good enough for attractive marketing but not actually sufficient for the field, but they bring in so many new people that there's no competing and combined with the shit management from the "techbro", the industry is forced to settle for a new normal that grows ever shittier until all the value has been extracted from it.
I used to be very interested in working on generative AI, it's what got me working on a PhD. But I completely lost interest the moment I saw how irresponsible everyone was with it. The tool is fine, but I am increasingly of the opinion that it's a tool we as a society are not ready to ethically wield.
But the first thing I saw people do is start generating "good enough" art to get their ideas out, and as an artist, I saw how shit the art looked - but it was enough for the people who were not artists. I have seen quite a few friends lose long-time commissioners who are just generating tons of ideas with clicks.
I don't lament people's ability to generate ideas or become invested in art with tools that make it easy for them. I lament the corporate idiots that are going to try to make fully AI generated movies and paintings and books, and flood the market, kill all the real artists, and yeah, enshittify our art and media.
An absolute ton. Now that it’s happening to art instead of accountants everyone should care?
Artists also didn't come up with their art style entirely on their own, they adapted what they learned to their current needs.
I never agreed to teach machines.
It costs $0.091 per copy to "cover" a song in the USA, you get this by securing a "mechanical license" and that rate is set by the Copyright Royalty Board[1]. I don't see why there can't be a similar way forward here.
[1] https://www.crb.gov/rate/
[1] Case docs for CRB Phonorecords IV (see subpart B): https://app.crb.gov/case/detail/21-CRB-0001-PR%20%282023-202...
[2] Summarizing article: https://variety.com/2022/music/news/songwriters-win-copyrigh...
If AI companies paid royalties for their entire training set, then it would be a different situation.
But given that it may be infeasible to prove or disprove whether or not AI was trained on public data, we should err on the side of caution and not allow the companies which build/operate AI to have copyrights over its output.
A case could be made that most work which humans produce is itself derivative of various public works and therefore the AI is not so different from humans and therefore it should be given copyrights... But I think the key difference is that the AI is not human, there are 'human rights' but there are no 'AI rights' or 'machine rights'... Until an AI can prove that it has agency and that it works for its own interests and not for a third party, it should not have any rights.
To me, it seems that the bigger, on-going issue, is "Who will pay for the compute charges?" (which are not trivial)
I think it's not so bad since the entities who train and/or use AI also have to pay a lot of money on their end.
Or what if I don’t do that, but generate several images and spend hours making a composite? What if I start with my own photograph and generate things to use in a composite on top of that?
It’s not so clear cut.
If you start with a Stable Diffusion program from scratch with all weights initialized at random (or whatever weights its initialization algorithm requires) and you train it on only your own photograph(s), then you will almost certainly find that it's not effective at all. In that case though, the low-quality output would be 100% your own.
The reason why generative AIs are so good is precisely because they have access to a large body of public data. Without the public data, they would not be good no matter how good the underlying software/algorithm is.
Operative word there: "further". If it's already trained on everyone else's input, then without that there wouldn't even be a Stable Diffusion for you to "further train on your own art", so YTF should you get sole ownership of the sum of it all?
What you're saying is like "Alright - what if I take Linux and add in my own code, and then sell that as closed source?" Nope, it's already GPLed, and so should AIs trained on public data be.
If the content is copyrighted, it should only be shown if user accepts the terms or pays for it.
This is unnecessary hyperbole.
Preventing the laundering of copyright material through AI algorithms does not prevent AI companies from existing. Take Adobe -- they have license to the training images they use and they have helpful products based on AI processing of those images. It won't be "free" money from dubious sources. But AI does not depend on the laundering of copyright material.
The difference is whether or not you are scraping and then generating similar products (eg the images themselves) or whether you are scraping for some other purpose.
I think that forms the legal basis for why AIs should be allowed to train on public data for free but it does not form a legal basis for AI to have copyrights over the output.
A photographer doesn’t create nature, the subjects, the architecture, the universe, the billions of R&D behind the silicon sensors, etc., but they create the situation and push the button.
Creating the situation, I.e. drawing from the near infinite of possibilities, is analogous to prompting.
For the case of observing copyrighted photos (not directly storing indefinitely) and generalizing a pine tree from that data, I think that is fine:
-assume we believe and rule it is not fine, then we will just allow huge companies to generate their own data sets and they will own this space moving forward (would require google maps levels of investment). Also, we wouldn’t be able to tell the difference between illegal SD output versus clean FAANG output without extremely draconian measures.
The US Congress decided to organize it that way, and would be responsible for reorganizing things. So call your Congressperson and make your argument.
> AI
Not as useful as some people imagine, but still quite useful.
> copyright
Should be abolished completely, it doesn't benefit society.
The reality is copyright has diverged so far from its intended purpose that there is no way to go back to the spirit of the law and rethink it because it is so fundamentally broken. So I anticipate some makeshift policy that nobody likes and gets broadly ignored because it is unenforceable.
Honestly, the best approach right now might be to simply say that AI generated content needs to declare that it is AI generated (somehow...) - that would at least give some context to the situation so that actual rules could have an effect down the road.
If you stopped calling it AI, and called these things what they really are, tools that predict tokens or pixels there would be a little less fear. Emphasis on tools, you wouldn't create law that didnt allow google search to exist. People think these function calls are thinking and anthropomorphizing them. You wouldn't make content from photoshop not copyrightable.