in their case, 'fair use' seems more fair - stuff like images, if it's a copyrighted work, are featured as a small resolution thumbnail. meanwhile, AI services (image generating ones) grab a high resolution copy off the internet (which could be a ding in itself, and also a ding for other "training data" collections which "only post links" (piratebay, anyone?), create a derivative (a model) off of that copy, and then create even more derivatives off of a model (which is an already legally questionable item in itself) in form of model output (which can be as high res as one wants). text based AI services also do some questionable sourcing of material, also create a derivative, a model, and then subsequent output. and the commercialization is more explicit, "this is a service/product/app, pay us for this product, pay us for compute, etc." all of these things steer AI services further and further away from "fair use".
AI services truly do not have a strong foothold in trying to pass their stuff as "fair use", let alone trying to argue that their "use" is "fairer" than, say, wikipedia's.
Wikipedia also has something to gain by AI not falling under Fair Use.
Wikipedia is most likely a substantial source of ChatGPT's training data. Considering Wikipedia primarily falls under the CC-BY-SA license, if it isn't fair use it could potentially make parts of ChatGPT - and potentially even its output - fall under CC-BY-SA too.
Related to both: how would you tell the difference? More importantly, how would you prove that in a court.
When it comes to copyright infringements, the burden of proof is on the copyright holder. I might claim the work of an AI as my own without informing anyone. And then when somebody copies my work I might sue them. How would they prove my work is not my own but created by an AI? If you fail to do that, a judge is going to rule in my favor. They'd have to.
Or when an AI supposedly infringes on a copyrighted work, who do you drag into court for that (the AI, it's developers, or the user?) and what's your argumentation going to be? You have to prove that the work is clearly yours and not fair use (e.g. a snippet, fragment, or a loose resemblance). And then you have to prove illegal distribution of that work that goes beyond fair use.
This is where the whole thing will fall apart IMHO. I'm sure the legal debates will drag out a long time but you can kind of see where this is going. The cat is out of the bag, and the volume of cases will soon be such that it will be hard to stay on top of them. While there are going to be a multitude of people who feel vaguely cheated, they might not have much of a solid basis for starting legal cases. A few might try but that might just nudge judges to put a stop to those cases with some tactical rulings. Enough people have attempted to abuse copyright for all sorts of reasons ever since the whole notion was formalized that any drastic changes in the way copyright works are both impractical and unlikely to survive a lot of scrutiny if/when that gets escalated to the supreme court.
A lot of amateur lawyers with very fuzzy notions about copyright, the whole notion of property, ownership, intellectual property etc. are not going to really going to be able to change that. And of course copyright law is not just a US thing. Other countries have copyright laws too and there are treaties and trade agreements between countries. All kind of hint at a likely outcome: unless an AI is producing verbatim copies of something copyrighted, it likely is simply fair use, whether people like that or not. And when people copy work produced by an AI, it's up to them to ensure that the work is not copyrighted. And up to the copyright holder to go after them. The AI itself cannot be sued. And a neural network bears no resemblance to any copyrighted work whatsoever; so it's hard to argue that the creators of those can be a party in any such cases. I'm sure people will try. But I'm skeptical of those people getting very far.
Creating some kind of exception for AIs in copyright law is going to be very hard and impractical. Getting international agreement on that even harder. At best it will take very long for anything to happen. By which time there will be so many AIs that outlawing them is not going to be practical or popular.
Copyright and "intellectual property" have always been fuzzy subjects, and the legal frameworks have always struggled to catch up with changes in society and technology. The origins of copyright are (I think) about publishers re-printing stuff printed by other publishers. It has come a long way with photo copiers, audio tape and with "copy" being the essential primitive of any digital computer. And now AIs can recycle creative thoughts in some sense.
I suppose at some point, it makes more sense to look at what kind of law we need in practice, rather than how existing law applies to reality. Should AI models be allowed to learn from publicly available resources, without permission? Probably yes. Should AI output be owned by the AI owner, or anyone? Probably no.
Can we (the world) come to an agreement? From there, can we change laws to make this work the way it should?
> Can we (the world) come to an agreement? From there, can we change laws to make this work the way it should?
I doubt there is any consensus for this. Or even a consensus on this being a problem to begin with. The easiest thing for legislators is to do nothing. Predicting that they will do exactly that is almost a foregone conclusion. After all that is what historically happens any time somebody has some issue with the way copyright works. In light of that, all judges can do is rule consistently with previous rulings. And there are a lot of those. Disregarding those is not something they can do. There's going to be lots of haggling on the details but in the end, they can't just wipe out centuries of previous cases and change e.g. the very loosely defined notion of fair use just because it is convenient for some subset of copyright holders.
Replace the word AI with "child" in the your comment and you kind of have the legal situation in a nutshell. If you stop thinking about AIs as special in any way and just focus on the outcome: a brain like thing producing content that may or may not infringe on some copyrighted work when it is distributed by someone. The distributing part is going to be done by people typically and it's what creates the infringement.
The simplest thing for a judge is to simply look at the content, disregard the way in which it was produced, and compare it to the original work. Either that's going to be a clear infringement under pre-existing notions of that or it isn't. The way in which that copy was created exactly is not really material to the case. There's no need to make any exceptions for there being an AI involved in some way or not. Either the content is an infringement or it is not.
Of course that would require judges and lawyers to be rational and logical, which is not an assumption I'd make lightly. But usually these things have a way of coming to some kind of conclusion that big stakeholders in society can live with.
> Replace the word AI with "child" in the your comment and you kind of have the legal situation in a nutshell.
Very insightful. I had not thought of it that way.
In a legal way, a minor is also very similar: the parent is usually liable if child causes damage. And what about wages, royalties and intellectual property claims for child stars like Michael Jackson, Brittney Spears, etc?
Indeed, just as complicated as claiming ownership of a product that's 99% attributable to "a tool" or "a third party collaborator".
> When it comes to copyright infringements, the burden of proof is on the copyright holder.
sorry but your first sentence is wholly inaccurate in the USA, if I recall my layman law book from twenty years ago. TRADEMARK burden of "policing" is incumbent on the trademark holder. COPYRIGHT is very often (or was) a little author or artist in their comfy studio putting their heart and time into creation, so copyright in the USA was changed specifically such that the author is protected simply by producing the work. Legitimate marks of copyright like the @ sign with a year, are not required but encouraged.
I think you are confusing the notion of a work being copyrighted and the notion of someone complaining about any infringements on their copy righted works in a court. The first is indeed more or less automatic. You can't actually opt out of copyright. It applies to any work that is created. With very few exceptions. Hence copy left licenses, which try to undo some of the restrictions that copyright impose. Likewise the notion of putting something in the public domain is a bit tricky under international law. It's not clear what that means everywhere. An explicit license is better.
The second bit (enforcement) can be done by the copyright holders. Or, as is common in e.g. the EU by organizations like GEMA that actively charge third parties money for playing music in public. That's a weird one IMHO. The money does not actually go to the rights owner but to big owners of music rights and other members of GEMA. So they benefit from other people's work through some weird legal construction that gives them the monopoly on exploitation rights for that sort of thing. But basically, unless somebody goes to court nothing happens. That's not automatic.
Trademarks are a completely different can of worms.
On the first point, do any law geeks know whether it would be possible to craft a new CC license along the lines of "CC-BY-NOAI" although, given that companies profit from their LLMs, I'd have thought a NoCommercial CC license ought to cover it, but evidently not, under US law.
Also worth noting that Fair Use is almost unique to American law. In much of the world, the doctrine is not recognized. Some jurisdictions offer similar provisions, but with limitations against commercial use, or selective exemption in for example education.
Although enforcement seems to be fairly selective. Google and similar services that hinge their entire business model on fair use seems to be permitted to operate even outside of US jurisdiction.
Regarding 2, I wonder if we may start to see AI services injecting the occasional human-constructed answer at random to ward against unauthorized distribution; just like how some map makers in the past included the occasional fictional feature to legally guard against unauthorized copying.
The TOS for the API says that you won't use it to train your model, if someone else uses it to write articles and then you train on those there is no problem.
The money you contribute to the Wikimedia Foundation is managed (and strategically distributed) by The Tides Foundation. "Hey ChatGPT, tell me more about the Tides Foundation and their ideology."
> In January 2016, the Foundation announced the creation of an endowment to safeguard its future. The Wikimedia Endowment was established as a donor-advised fund at the Tides Foundation, with a stated goal to raise US$100 million in the next 10 years.
> The Foundation itself has provided annual grants of $5 million to its Endowment since 2016. These amounts have been recorded as part of the Foundation's "awards and grants" expenses. In September 2021, the Foundation announced that the Wikimedia Endowment had reached its initial $100 million fundraising goal in June 2021, five years ahead of its initial target.
That's 2.7% of the money donated to the Foundation, right?
Well, because Tides Foundation is a Donor Advised Fund[1], which is the Dark Money PAC equivalent in the non-profit world. That means when you donate to Wikimedia because you think you're supporting their infrastructure costs, what is actually happening is your donations are being funneled to other causes that you may or may not agree with by the Tides Foundation.
Tides Foundation - great view if you can visit in the Presidio SF; very smart people with a lot of actual experience.. stable, professional, highly recommended.
Why is fair use even an issue? The content provider provided a lawful copy to the recipient. The first sale doctrine means it’s theirs to use as they please. Since the data is distributed as a software program the copyright act permits copying it as needed for a machine to actually use it.
There’s a reason it’s called COPYright and not USEright. It’s remarkable to me that some private attorneys working for Microsoft somehow hypnotized the entire population into believing that they need to be licensed rights that they already have by law.
> With this in mind, it is important to note that Creative Commons licenses allow for free reproduction and reuse,
It’s only free unless NonCommercial or NoDerivatives modifiers are present.
> However, it is not clear yet whether massively copying content from these sources may result in a violation of the Creative Commons license if attribution is not granted.
If “CC-A” is used (no NC, no ND) then even though commercial use and derivatives (aka LLM training) are allowed the author of the original should still be attributed wherever their work is used.
31 comments
[ 7.0 ms ] story [ 83.5 ms ] thread1. is the use of copyright-protected works to train AI models fair use.
Answer: probably, we don't know for sure.
2. can you have copyright on the output of an AI?
Answer: probably not, unless there is sufficient human creativity
>Answer: probably, we don't know for sure.
It's worth remembering Wikipedia has everything to gain by being allowed to handwave legalities away under "Fair Use".
While their thoughts might be interesting to hear, we shouldn't forget they are strongly biased concerning the subject matter.
AI services truly do not have a strong foothold in trying to pass their stuff as "fair use", let alone trying to argue that their "use" is "fairer" than, say, wikipedia's.
Wikipedia is most likely a substantial source of ChatGPT's training data. Considering Wikipedia primarily falls under the CC-BY-SA license, if it isn't fair use it could potentially make parts of ChatGPT - and potentially even its output - fall under CC-BY-SA too.
When it comes to copyright infringements, the burden of proof is on the copyright holder. I might claim the work of an AI as my own without informing anyone. And then when somebody copies my work I might sue them. How would they prove my work is not my own but created by an AI? If you fail to do that, a judge is going to rule in my favor. They'd have to.
Or when an AI supposedly infringes on a copyrighted work, who do you drag into court for that (the AI, it's developers, or the user?) and what's your argumentation going to be? You have to prove that the work is clearly yours and not fair use (e.g. a snippet, fragment, or a loose resemblance). And then you have to prove illegal distribution of that work that goes beyond fair use.
This is where the whole thing will fall apart IMHO. I'm sure the legal debates will drag out a long time but you can kind of see where this is going. The cat is out of the bag, and the volume of cases will soon be such that it will be hard to stay on top of them. While there are going to be a multitude of people who feel vaguely cheated, they might not have much of a solid basis for starting legal cases. A few might try but that might just nudge judges to put a stop to those cases with some tactical rulings. Enough people have attempted to abuse copyright for all sorts of reasons ever since the whole notion was formalized that any drastic changes in the way copyright works are both impractical and unlikely to survive a lot of scrutiny if/when that gets escalated to the supreme court.
A lot of amateur lawyers with very fuzzy notions about copyright, the whole notion of property, ownership, intellectual property etc. are not going to really going to be able to change that. And of course copyright law is not just a US thing. Other countries have copyright laws too and there are treaties and trade agreements between countries. All kind of hint at a likely outcome: unless an AI is producing verbatim copies of something copyrighted, it likely is simply fair use, whether people like that or not. And when people copy work produced by an AI, it's up to them to ensure that the work is not copyrighted. And up to the copyright holder to go after them. The AI itself cannot be sued. And a neural network bears no resemblance to any copyrighted work whatsoever; so it's hard to argue that the creators of those can be a party in any such cases. I'm sure people will try. But I'm skeptical of those people getting very far.
Creating some kind of exception for AIs in copyright law is going to be very hard and impractical. Getting international agreement on that even harder. At best it will take very long for anything to happen. By which time there will be so many AIs that outlawing them is not going to be practical or popular.
Copyright and "intellectual property" have always been fuzzy subjects, and the legal frameworks have always struggled to catch up with changes in society and technology. The origins of copyright are (I think) about publishers re-printing stuff printed by other publishers. It has come a long way with photo copiers, audio tape and with "copy" being the essential primitive of any digital computer. And now AIs can recycle creative thoughts in some sense.
I suppose at some point, it makes more sense to look at what kind of law we need in practice, rather than how existing law applies to reality. Should AI models be allowed to learn from publicly available resources, without permission? Probably yes. Should AI output be owned by the AI owner, or anyone? Probably no.
Can we (the world) come to an agreement? From there, can we change laws to make this work the way it should?
I doubt there is any consensus for this. Or even a consensus on this being a problem to begin with. The easiest thing for legislators is to do nothing. Predicting that they will do exactly that is almost a foregone conclusion. After all that is what historically happens any time somebody has some issue with the way copyright works. In light of that, all judges can do is rule consistently with previous rulings. And there are a lot of those. Disregarding those is not something they can do. There's going to be lots of haggling on the details but in the end, they can't just wipe out centuries of previous cases and change e.g. the very loosely defined notion of fair use just because it is convenient for some subset of copyright holders.
Replace the word AI with "child" in the your comment and you kind of have the legal situation in a nutshell. If you stop thinking about AIs as special in any way and just focus on the outcome: a brain like thing producing content that may or may not infringe on some copyrighted work when it is distributed by someone. The distributing part is going to be done by people typically and it's what creates the infringement.
The simplest thing for a judge is to simply look at the content, disregard the way in which it was produced, and compare it to the original work. Either that's going to be a clear infringement under pre-existing notions of that or it isn't. The way in which that copy was created exactly is not really material to the case. There's no need to make any exceptions for there being an AI involved in some way or not. Either the content is an infringement or it is not.
Of course that would require judges and lawyers to be rational and logical, which is not an assumption I'd make lightly. But usually these things have a way of coming to some kind of conclusion that big stakeholders in society can live with.
Very insightful. I had not thought of it that way.
In a legal way, a minor is also very similar: the parent is usually liable if child causes damage. And what about wages, royalties and intellectual property claims for child stars like Michael Jackson, Brittney Spears, etc?
Indeed, just as complicated as claiming ownership of a product that's 99% attributable to "a tool" or "a third party collaborator".
sorry but your first sentence is wholly inaccurate in the USA, if I recall my layman law book from twenty years ago. TRADEMARK burden of "policing" is incumbent on the trademark holder. COPYRIGHT is very often (or was) a little author or artist in their comfy studio putting their heart and time into creation, so copyright in the USA was changed specifically such that the author is protected simply by producing the work. Legitimate marks of copyright like the @ sign with a year, are not required but encouraged.
The second bit (enforcement) can be done by the copyright holders. Or, as is common in e.g. the EU by organizations like GEMA that actively charge third parties money for playing music in public. That's a weird one IMHO. The money does not actually go to the rights owner but to big owners of music rights and other members of GEMA. So they benefit from other people's work through some weird legal construction that gives them the monopoly on exploitation rights for that sort of thing. But basically, unless somebody goes to court nothing happens. That's not automatic.
Trademarks are a completely different can of worms.
How does any of this help me, a non-US person?
Although enforcement seems to be fairly selective. Google and similar services that hinge their entire business model on fair use seems to be permitted to operate even outside of US jurisdiction.
> The Foundation itself has provided annual grants of $5 million to its Endowment since 2016. These amounts have been recorded as part of the Foundation's "awards and grants" expenses. In September 2021, the Foundation announced that the Wikimedia Endowment had reached its initial $100 million fundraising goal in June 2021, five years ahead of its initial target.
That's 2.7% of the money donated to the Foundation, right?
https://en.wikipedia.org/wiki/Wikimedia_Foundation#Wikimedia...
https://diff.wikimedia.org/2022/10/26/governance-updates-for...
Why would anybody support that?
[1] https://en.wikipedia.org//wiki/Donor-advised_fund
There’s a reason it’s called COPYright and not USEright. It’s remarkable to me that some private attorneys working for Microsoft somehow hypnotized the entire population into believing that they need to be licensed rights that they already have by law.
That makes sense why when I use it as a worldbuilding tool for my RP campaigns it's not bad because there is no "bad" answer in that case technically
It’s only free unless NonCommercial or NoDerivatives modifiers are present.
> However, it is not clear yet whether massively copying content from these sources may result in a violation of the Creative Commons license if attribution is not granted.
Isn’t it clear that it is?