8 ms·

Wikimedia Foundation: Copyright Analysis of ChatGPT

61 points by tomodachi94 4y ago

isaacfrond 4y ago
tl;dr Two big questions: 1. 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
- Dalewyn 4y ago
  >1. is the use of copyright-protected works to train AI models fair use. >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.
  - pxoe 4y ago
    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.
  - crote 4y ago
    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.
- jillesvangurp 4y ago
  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.
- mikro2nd 4y ago
  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. How does any of this help me, a non-US person?
- marginalia_nu 4y ago
  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.
- jhanschoo 4y ago
  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.
seydor 4y ago
What about the most important Q: Can you use chatGPT output to train another model?
- pipo234 4y ago
  And if "no", what can you do to make sure your model isn't using C hatGPT output by accident.
  - sebzim4500 4y ago
    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.
    - nickthegreek 4y ago
      Or you just use the web interface, not the API.
- diggan 4y ago
  You can, as in it's possible. But against OpenAI's policies so if they find out that's what you're doing, you'll get kicked out of the platform.
  - seydor 4y ago
    But it's not illegal. So you could ddos it
yucky 4y ago
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."
- fathyb 4y ago
  > 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? https://en.wikipedia.org/wiki/Wikimedia_Foundation#Wikimedia_Endowment https://en.wikipedia.org/wiki/Wikimedia_Foundation#Wikimedia...
  - swalling 4y ago
    Yes. And the endowment is also moving off Tides now that it has its own approved 501(c)(3) status. https://diff.wikimedia.org/2022/10/26/governance-updates-for-the-wikimedia-endowment/ https://diff.wikimedia.org/2022/10/26/governance-updates-for...
- consumer451 4y ago
  Can you please specify what it is that concerns you about the foundation?
  - yucky 4y ago
    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. Why would anybody support that? [1] https://en.wikipedia.org//wiki/Donor-advised_fund https://en.wikipedia.org//wiki/Donor-advised_fund
User23 4y ago
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.
mirages 4y ago
" it will sometimes provide a wrong answer to a question or “hallucinate” material that does not exist " 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
anileated 4y ago
> 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. Isn’t it clear that it is?
- cyanydeez 4y ago
  Attribution means "cc license A was used in this work" not "we used a bunch of cc license"
  - anileated 4y ago
    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.

61 points by tomodachi94 4y ago

tl;dr Two big questions: 1. 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

Dalewyn 4y ago
>1. is the use of copyright-protected works to train AI models fair use. >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.
- pxoe 4y ago
  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.
- crote 4y ago
  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.
jillesvangurp 4y ago
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.
mikro2nd 4y ago
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. How does any of this help me, a non-US person?
marginalia_nu 4y ago
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.
jhanschoo 4y ago
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.

seydor 4y ago

What about the most important Q: Can you use chatGPT output to train another model?

pipo234 4y ago
And if "no", what can you do to make sure your model isn't using C hatGPT output by accident.
- sebzim4500 4y ago
  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.
  - nickthegreek 4y ago
    Or you just use the web interface, not the API.
diggan 4y ago
You can, as in it's possible. But against OpenAI's policies so if they find out that's what you're doing, you'll get kicked out of the platform.
- seydor 4y ago
  But it's not illegal. So you could ddos it

yucky 4y ago

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."

fathyb 4y ago
> 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? https://en.wikipedia.org/wiki/Wikimedia_Foundation#Wikimedia_Endowment https://en.wikipedia.org/wiki/Wikimedia_Foundation#Wikimedia...
- swalling 4y ago
  Yes. And the endowment is also moving off Tides now that it has its own approved 501(c)(3) status. https://diff.wikimedia.org/2022/10/26/governance-updates-for-the-wikimedia-endowment/ https://diff.wikimedia.org/2022/10/26/governance-updates-for...
consumer451 4y ago
Can you please specify what it is that concerns you about the foundation?
- yucky 4y ago
  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. Why would anybody support that? [1] https://en.wikipedia.org//wiki/Donor-advised_fund https://en.wikipedia.org//wiki/Donor-advised_fund

User23 4y ago

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.

mirages 4y ago

" it will sometimes provide a wrong answer to a question or “hallucinate” material that does not exist " 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

anileated 4y ago

> 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. Isn’t it clear that it is?

cyanydeez 4y ago
Attribution means "cc license A was used in this work" not "we used a bunch of cc license"
- anileated 4y ago
  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.