28 ms·

I think copilot is a clearer copyright violation than any of the stable diffusion projects though because code has a much narrower band of expression than image

by rtkwe 4y ago

I think copilot is a clearer copyright violation than any of the stable diffusion projects though because code has a much narrower band of expression than images. It's really easy to look at the output of CoPilot and match it back to the original source and say these are the same. With stable diffusion it's much closer to someone remixing and aping the images than it is reproducing originals. I haven't been following super closely but I don't know of any claims or examples where input images were recreated to a significant degree by stable diffusion.

makeitdouble 4y ago
I think the is exacty the gap the gp is mentionning: to a trained artist it is clear as water that the original image has been lifted wholesale, even if for instance the colors are adjusted here and there. You put it as a remix, but remixes are credited and expressed as such.
- jzb 4y ago
  I haven’t seen any side by sides that seem like a lift. Any examples? I don’t see Midjourney (et al) as remixes, myself. More like “inspired by.”
  - omnimus 4y ago
    Its clear where the knowhow was lifted from it doesnt matter that if the final image is somewhat unique (almost every image is).
    - matkoniecz 4y ago
      style is not copyrightable under current rules
      - omnimus 4y ago
        But it means the models were trained on images that are under copyright. In fact many of these models were trained exclusively on such images without any permission. For example Midjourney is clearly trained on everything on artstation.com where almost all images have commercial purpose / licenses.
  - kevingadd 4y ago
    Not safe for work, but one example I saw going around: https://twitter.com/ebkim00/status/1579485164442648577 https://twitter.com/ebkim00/status/1579485164442648577 Not sure if this was fed the original image as an input or not. Also seen a couple cases where people explicitly trained a network to imitate an artist's work, like the deceased Kim Jung Gi.
    - lbotos 4y ago
      It's really interesting. I suspect the face was inpainted in, or this was a "img2img". I think over time we are going to see the following: - If you take say a star wars poster, and inpaint in a trained face over luke's, and sell that to people as a service, you will probably be approached for copyright and trademark infringement. - If you are doing the above with a satirical take, you might be able to claim fair use. - If you are using AI as a "collage generator" to smash together a ton of prompts into a "unique" piece, you may be safe from infringement but you are taking a risk as you don't know what % of source material your new work contains. I'd like to imagine if you inpaint in say 20 details with various sub-prompts that you are getting "safer".
      - numpad0 4y ago
        Features outside the face is lost/changed from original on the right, so can’t be face inpainting. Unlikely to be style transfers, because some body parts are moved. Most plausibly this was generated. So much for “generation” - it seems as if these models are just overfitting on extremely small subset of input data that it did not utterly failed to train on, almost that there could be geniuses who would be able to directly generate weight data from said images without all the gradient descent thing.
    - rtkwe 4y ago
      That's clearly lifting style, pose and general location but in each of those there are changes. Even for the original art we could find tons of examples of very similar poses and backgrounds because anime girl in a bathing suit on a beach background isn't that original of an image at the concept level. That pose also is a pretty well worn. This is the problem of applying the idea of ownership to ideas and expression like art. Art in particular is a very remix and recombination driven field.
  - matkoniecz 4y ago
    https://alexanderwales.com/wp-content/uploads/2022/08/image.png https://alexanderwales.com/wp-content/uploads/2022/08/image.... Left: “Girl with a Pearl Earring, by Johannes Vermeer” by Stable Diffusion Right: Girl with a Pearl Earring by Johannes Vermeer This specific one is not copyright violation as it is old enough for copyright to expire. But the same may happen with other images. from https://alexanderwales.com/the-ai-art-apocalypse/ https://alexanderwales.com/the-ai-art-apocalypse/ and https://alexanderwales.com/addendum-to-the-ai-art-apocalypse-theft/ https://alexanderwales.com/addendum-to-the-ai-art-apocalypse...
    - Filligree 4y ago
      If a human drew that, it would not be copyright violation.
      - matkoniecz 4y ago
        It is a clear case of derivative work (see also https://commons.wikimedia.org/wiki/Commons:Derivative_works https://commons.wikimedia.org/wiki/Commons:Derivative_works - internal docs, but their explanation of copyright status tends to be well done) This specific one would not be a problem, but doing it with a still copyrighted work would be.
      - mattkrause 4y ago
        I’m not so sure about that. The scenes à faire doctrine would certainly let you paint your own picture of a pretty girl with a large earring, even a pearl one. That, however, is definitely the same person, in the same pose/composition, in the same outfit. The colors are slightly off, but the difference feels like a technical error rather than an expressive choice.
        Thorrez 4y ago
        Even if it is an expressive choice of the new artist, if enough of the original artist's expressive choice remains, it could still be a copyright violation. Fair use can sometimes be a defense, but there are a lot of factors that go into determining whether something is fair use.
- omnimus 4y ago
  Exactly to a programmer copilot is clear violation, to a writer gpt-3 is clear violation, to an artist dalle-2 is clear violation. The artist might love copilot, the writer might love dalle, the programmer might love gpt-3. Its all the same they just dont realize this.
  - sidewndr46 4y ago
    Does dalle-2 verbatim reproduce artwork? I have never used it.
    - CapsAdmin 4y ago
      It's kind of like having millions of parameters you can tweak to get to an image. So an image does not really exist in the model. I can imagine Mona Lisa in my head, but it doesn't really "exist" verbatim in my head. It's only an approximation. I believe copilot works the same way (?)
      - heavyset_go 4y ago
        NNs can and do encode information from their training sets in the models themselves, sometimes verbatim. Sometimes the original information is there in the model, encoded/compressed/however you want to look at it, and can be reproduced.
      - hackerfromthefu 4y ago
        This is just nonsense. It's similar to saying that any digital representation of an image isn't an image just a dataset that represent it. If what you said was any sort of defense every image copyright would never apply to any digital image, because the images can be saved in different resolutions, different file formats, or encoded down. e.g. if a jpeg 'image' was only an image at an exact set of digital bits i could save it again with a different quality setting and end up with a different set of digital bits. But everyone still recognises when an image looks the same, and courts will uphold copyright claims regardless of the digital encoding of an image. So goodluck with that spurious argument that it's not copyright because 'its on the internet (oh its with AI etc).
mr_toad 4y ago
> I haven't been following super closely but I don't know of any claims or examples where input images were recreated to a significant degree by stable diffusion. I think that the argument being made by some artists is that the training process itself violates copyright just by using the training data. That’s quite different from arguing that the output violates copyright, which is what the tweet in this case was about.
- rtkwe 4y ago
  I'm dubious of that in cases where the training set isn't distributed. If we call the training copyright infringement is downloading an image infringement? is caching?
  - didibus 4y ago
    I think it's more a question of derivative work. Normally derivative work is an infringement unless it falls under fair use. Now a human can take inspiration from like 100 different sources and probably end up with something that no one would recognize as derivative to any of them. But it also wouldn't be obvious that the human did that. But with an ML model, it's clearly a derivative in that the learned function is mathematically derived from its dataset and so is all the resulting outputs. I think this brings a new question though. Because till now derivative was kind of implied that the output was recognizable as being derived. With AI, you can tweak it so the output doesn't end up being easily recognizable as derived, but we know it's still derived. Personally I think what really matters is more a question of what should be the legal framework around it. How do we balance the interests of AI companies and that of developers, artists, citizens who are the authors of the dataset that enabled the AI to exist. And what right should each party be given?
    - rtkwe 4y ago
      The real kink in that application of derivative work to me is the entire dataset goes into the model and is to some vanishingly small extent is used in every output how can we meaningfully assign ownership through that transition and mixing. And when we do how do we do it without exacerbating the extant problem of copyright in art? We already can't use characters and settings made during out own lifetimes in our own expression because Disney got life + 70 through Congress.
AmericanChopper 4y ago
I don’t think copilot is intrinsically a copyright violation, as you seem to be alluding to. Examples like this seem to be more controversial, but I’m not sure there’s a clear copyright violation there either. If you asked every developer on earth to implement FizzBuzz, how many actually different implementations would you get? Probably not very many. Who should own the copyright for each of them? Would the outcome be different for any other product feature? If you asked every dev on earth to write a function that checked a JWT claim, how many of them would be more or less exactly the same? Would that be a copyright violation? I hope the courts answer some of these questions one day.
- stuartd 4y ago
  > If you asked every developer on earth to implement FizzBuzz, how many actually different implementations would you get? Thousands at least. Some of which would actually work.
  - AmericanChopper 4y ago
    There’s a finite number of ways to implement a working FizzBuzz (or anything else) in any given language, that aren’t substantially similar, is my point. At least without introducing pointless code for the explicit purpose of making it look different.
- dataflow 4y ago
  > If you asked every developer on earth to implement FizzBuzz, how many actually different implementations would you get? Does it matter? If you examined every copyright lawsuit on earth over code, how many of them would actually be over FizzBuzz?
  - AmericanChopper 4y ago
    The same rationale applies to any other simple code block, as I elaborated on.
    - 4y ago
paulgb 4y ago
I don’t know of any examples of images being wholly recreated, but it’s certainly possible to use the name of some living artists to get work in their style. In those cases, it seems like not such a leap to say that the AI has obviously seen that artist’s work and that the output is a derivative work. (The obvious counterargument is that this is the same as a human looking at an artist’s work and aping the style.)
- matkoniecz 4y ago
  https://alexanderwales.com/wp-content/uploads/2022/08/image.png https://alexanderwales.com/wp-content/uploads/2022/08/image.... Left: “Girl with a Pearl Earring, by Johannes Vermeer” by Stable Diffusion Right: Girl with a Pearl Earring by Johannes Vermeer This specific one is not copyright violation as it is old enough for copyright to expire. But the same may happen with other images. from https://alexanderwales.com/the-ai-art-apocalypse/ https://alexanderwales.com/the-ai-art-apocalypse/ and https://alexanderwales.com/addendum-to-the-ai-art-apocalypse-theft/ https://alexanderwales.com/addendum-to-the-ai-art-apocalypse...
  - londons_explore 4y ago
    I think this happens a lot with famous images since that image will be in the training set hundreds of times. Even if deduplication efforts are done, that painting will still be in the background of movie shots etc.
  - rtkwe 4y ago
    > Left: “Girl with a Pearl Earring, by Johannes Vermeer” by Stable Diffusion Right: Girl with a Pearl Earring by Johannes Vermeer Even that if done by a person as far as I understand it would not constitute a copyright infringement. It's a separate work mimicking Vermeer's original. The closest real world equivalent I can think of is probably the Obama Hope case by AP vs Shepard Fairy but that settled out of court so we don't really know what the status of that kind of reproduction is legally. On top of that though the SD image isn't just a recoloring with some additions like Fairy's was so it's not quite as close to the original as that case is.
pavlov 4y ago
Stable Diffusion sometimes reproduces the large watermarks used by stock photo providers on their free sample images. That’s embarrassing at the minimum, and potentially a trademark violation.
- bigiain 4y ago
  Surely at the very least it'd be a TOS violation? I doubt any stock photo service grants you enough rights to redistribute their watermarked free image samples? Especially not in the context of a project like Stable Diffusion?
  - Filligree 4y ago
    But it's not reproducing their samples. It's just adding their watermark to newly generated pictures you can't find in the training set.
    - rovr138 4y ago
      If the watermark is their logo or name, it could copyrighted or trademarked.
      - nl 4y ago
        And it's the responsibility of the person using the tool to generate that image not to violate copyright by redistributing it.
        MereInterest 4y ago
        Just like it's the person's responsibility to only recombine jpeg basis states when they don't correspond to a copyrighted image? It seems more and more to be the case that the trained model is, in large part, a very compact representation of the training data. I'm not seeing a difference between distributing a model that can be used to reconstruct the input images, as opposed to distributing jpeg basis states that can be used to reconstruct the original image.
DannyBee 4y ago
Well no. Code is only protected to the degree it is creative and not functionally driven anyway. So the reduced band of possible expression often directly reduces the protectability-through-copyright.
bigiain 4y ago
> With stable diffusion it's much closer to someone remixing and aping the images than it is reproducing originals. So very similar to how the music industry treats sampling then? Everybody using CoPilot needs to get "code sample clearance" from the original copyright holder before publishing their remix or new program that uses snippets of somebody else's code... Try explaining _that_ to your boss and legal department. "To: <all software dev> Effective immediately, any use of Github is forbidden without prior written approval from both the CTO and General Councel."
- kmeisthax 4y ago
  This is already a problem with anyone who ever copypastes from Stack Overflow. You're all violating CC-BY-SA[0] and nobody really cares about this. [0] https://stackoverflow.com/help/licensing https://stackoverflow.com/help/licensing
  - bscphil 4y ago
    If I ever take any code from SO, I include a comment with a link to it. Surely that's standard practice for anything longer than a line or two?
    - fourthark 4y ago
      I do the same. I think it satisfies BY (attribution) but not SA (Share Alike). As GP says, no one really cares, but it seems hard to satisfy SA... even if you are pasting into open source, is your license compatible with CC? Perhaps I'm over-thinking this.
kmeisthax 4y ago
The reason why it's easy to match Copilot results back to the original source is that the users are starting with prompts that match their public code, deliberately to cause prompt regurgitation. Stable Diffusion actually has a similar problem. Certain terms that directly call up a particular famous painting by name - say, the Mona Lisa[0] - will just produce that painting, possibly tiled on top of itself, and it won't bother with any of the other keywords or phrases you throw at it. The underlying problem is that the AI just outright forgets that it's supposed to create novel works when you give it anything resembling the training set data. If it was just that the AI could spit out training set data when you ask for it, I wouldn't be concerned[1], but this could also happen inadvertently. This would mean that anyone using Copilot to write production code would be risking copyright liability. Through the AI they have access to the entire training set, and the AI has a habit of accidentally producing output that's substantially similar to it. Those are the two prongs of a copyright infringement claim right there. [0] For the record I was trying to get it to draw a picture of the Mona Lisa slapping Yoshikage Kira across the cheek [1] Anyone using an AI system to "launder" creative works is still infringing copyright. AI does not carve a shiny new loophole in the GPL.
- xani_ 4y ago
  > The reason why it's easy to match Copilot results back to the original source is that the users are starting with prompts that match their public code, deliberately to cause prompt regurgitation. The reason doesn't really matter...
  - lofatdairy 4y ago
    GP is just highlighting why this is so common and often a challenging edge case. If you ask it for something that's exactly in its dataset, the "best" solution that minimizes loss will be that existing code. Thus, it's somewhat intrinsic to applying statistical learning to text completion. This means MS really shouldn't have used copyleft code at all, and really shouldn't be selling copilot in this state, but "luckily" for them, short of a class action suit I don't really see any recourse for the programmers who's work they're reselling.
dv_dt 4y ago
I suspect it’s going to be a discussion similar to the introduction of music sampling, followed by a lot of litigation, followed by a settling of law on the matter. The interesting part is if AI will be considered a tooling mechanism much like the tooling used to record and manipulate a music sample into a new composition.

by rtkwe 4y ago

makeitdouble 4y ago

I think the is exacty the gap the gp is mentionning: to a trained artist it is clear as water that the original image has been lifted wholesale, even if for instance the colors are adjusted here and there. You put it as a remix, but remixes are credited and expressed as such.

jzb 4y ago
I haven’t seen any side by sides that seem like a lift. Any examples? I don’t see Midjourney (et al) as remixes, myself. More like “inspired by.”
- omnimus 4y ago
  Its clear where the knowhow was lifted from it doesnt matter that if the final image is somewhat unique (almost every image is).
  - matkoniecz 4y ago
    style is not copyrightable under current rules
    - omnimus 4y ago
      But it means the models were trained on images that are under copyright. In fact many of these models were trained exclusively on such images without any permission. For example Midjourney is clearly trained on everything on artstation.com where almost all images have commercial purpose / licenses.
- kevingadd 4y ago
  Not safe for work, but one example I saw going around: https://twitter.com/ebkim00/status/1579485164442648577 https://twitter.com/ebkim00/status/1579485164442648577 Not sure if this was fed the original image as an input or not. Also seen a couple cases where people explicitly trained a network to imitate an artist's work, like the deceased Kim Jung Gi.
  - lbotos 4y ago
    It's really interesting. I suspect the face was inpainted in, or this was a "img2img". I think over time we are going to see the following: - If you take say a star wars poster, and inpaint in a trained face over luke's, and sell that to people as a service, you will probably be approached for copyright and trademark infringement. - If you are doing the above with a satirical take, you might be able to claim fair use. - If you are using AI as a "collage generator" to smash together a ton of prompts into a "unique" piece, you may be safe from infringement but you are taking a risk as you don't know what % of source material your new work contains. I'd like to imagine if you inpaint in say 20 details with various sub-prompts that you are getting "safer".
    - numpad0 4y ago
      Features outside the face is lost/changed from original on the right, so can’t be face inpainting. Unlikely to be style transfers, because some body parts are moved. Most plausibly this was generated. So much for “generation” - it seems as if these models are just overfitting on extremely small subset of input data that it did not utterly failed to train on, almost that there could be geniuses who would be able to directly generate weight data from said images without all the gradient descent thing.
  - rtkwe 4y ago
    That's clearly lifting style, pose and general location but in each of those there are changes. Even for the original art we could find tons of examples of very similar poses and backgrounds because anime girl in a bathing suit on a beach background isn't that original of an image at the concept level. That pose also is a pretty well worn. This is the problem of applying the idea of ownership to ideas and expression like art. Art in particular is a very remix and recombination driven field.
- matkoniecz 4y ago
  https://alexanderwales.com/wp-content/uploads/2022/08/image.png https://alexanderwales.com/wp-content/uploads/2022/08/image.... Left: “Girl with a Pearl Earring, by Johannes Vermeer” by Stable Diffusion Right: Girl with a Pearl Earring by Johannes Vermeer This specific one is not copyright violation as it is old enough for copyright to expire. But the same may happen with other images. from https://alexanderwales.com/the-ai-art-apocalypse/ https://alexanderwales.com/the-ai-art-apocalypse/ and https://alexanderwales.com/addendum-to-the-ai-art-apocalypse-theft/ https://alexanderwales.com/addendum-to-the-ai-art-apocalypse...
  - Filligree 4y ago
    If a human drew that, it would not be copyright violation.
    - matkoniecz 4y ago
      It is a clear case of derivative work (see also https://commons.wikimedia.org/wiki/Commons:Derivative_works https://commons.wikimedia.org/wiki/Commons:Derivative_works - internal docs, but their explanation of copyright status tends to be well done) This specific one would not be a problem, but doing it with a still copyrighted work would be.
    - mattkrause 4y ago
      I’m not so sure about that. The scenes à faire doctrine would certainly let you paint your own picture of a pretty girl with a large earring, even a pearl one. That, however, is definitely the same person, in the same pose/composition, in the same outfit. The colors are slightly off, but the difference feels like a technical error rather than an expressive choice.
      - Thorrez 4y ago
        Even if it is an expressive choice of the new artist, if enough of the original artist's expressive choice remains, it could still be a copyright violation. Fair use can sometimes be a defense, but there are a lot of factors that go into determining whether something is fair use.
omnimus 4y ago
Exactly to a programmer copilot is clear violation, to a writer gpt-3 is clear violation, to an artist dalle-2 is clear violation. The artist might love copilot, the writer might love dalle, the programmer might love gpt-3. Its all the same they just dont realize this.
- sidewndr46 4y ago
  Does dalle-2 verbatim reproduce artwork? I have never used it.
  - CapsAdmin 4y ago
    It's kind of like having millions of parameters you can tweak to get to an image. So an image does not really exist in the model. I can imagine Mona Lisa in my head, but it doesn't really "exist" verbatim in my head. It's only an approximation. I believe copilot works the same way (?)
    - heavyset_go 4y ago
      NNs can and do encode information from their training sets in the models themselves, sometimes verbatim. Sometimes the original information is there in the model, encoded/compressed/however you want to look at it, and can be reproduced.
    - hackerfromthefu 4y ago
      This is just nonsense. It's similar to saying that any digital representation of an image isn't an image just a dataset that represent it. If what you said was any sort of defense every image copyright would never apply to any digital image, because the images can be saved in different resolutions, different file formats, or encoded down. e.g. if a jpeg 'image' was only an image at an exact set of digital bits i could save it again with a different quality setting and end up with a different set of digital bits. But everyone still recognises when an image looks the same, and courts will uphold copyright claims regardless of the digital encoding of an image. So goodluck with that spurious argument that it's not copyright because 'its on the internet (oh its with AI etc).

mr_toad 4y ago

> I haven't been following super closely but I don't know of any claims or examples where input images were recreated to a significant degree by stable diffusion. I think that the argument being made by some artists is that the training process itself violates copyright just by using the training data. That’s quite different from arguing that the output violates copyright, which is what the tweet in this case was about.

rtkwe 4y ago
I'm dubious of that in cases where the training set isn't distributed. If we call the training copyright infringement is downloading an image infringement? is caching?
- didibus 4y ago
  I think it's more a question of derivative work. Normally derivative work is an infringement unless it falls under fair use. Now a human can take inspiration from like 100 different sources and probably end up with something that no one would recognize as derivative to any of them. But it also wouldn't be obvious that the human did that. But with an ML model, it's clearly a derivative in that the learned function is mathematically derived from its dataset and so is all the resulting outputs. I think this brings a new question though. Because till now derivative was kind of implied that the output was recognizable as being derived. With AI, you can tweak it so the output doesn't end up being easily recognizable as derived, but we know it's still derived. Personally I think what really matters is more a question of what should be the legal framework around it. How do we balance the interests of AI companies and that of developers, artists, citizens who are the authors of the dataset that enabled the AI to exist. And what right should each party be given?
  - rtkwe 4y ago
    The real kink in that application of derivative work to me is the entire dataset goes into the model and is to some vanishingly small extent is used in every output how can we meaningfully assign ownership through that transition and mixing. And when we do how do we do it without exacerbating the extant problem of copyright in art? We already can't use characters and settings made during out own lifetimes in our own expression because Disney got life + 70 through Congress.

AmericanChopper 4y ago

I don’t think copilot is intrinsically a copyright violation, as you seem to be alluding to. Examples like this seem to be more controversial, but I’m not sure there’s a clear copyright violation there either. If you asked every developer on earth to implement FizzBuzz, how many actually different implementations would you get? Probably not very many. Who should own the copyright for each of them? Would the outcome be different for any other product feature? If you asked every dev on earth to write a function that checked a JWT claim, how many of them would be more or less exactly the same? Would that be a copyright violation? I hope the courts answer some of these questions one day.

stuartd 4y ago
> If you asked every developer on earth to implement FizzBuzz, how many actually different implementations would you get? Thousands at least. Some of which would actually work.
- AmericanChopper 4y ago
  There’s a finite number of ways to implement a working FizzBuzz (or anything else) in any given language, that aren’t substantially similar, is my point. At least without introducing pointless code for the explicit purpose of making it look different.
dataflow 4y ago
> If you asked every developer on earth to implement FizzBuzz, how many actually different implementations would you get? Does it matter? If you examined every copyright lawsuit on earth over code, how many of them would actually be over FizzBuzz?
- AmericanChopper 4y ago
  The same rationale applies to any other simple code block, as I elaborated on.
  - 4y ago

paulgb 4y ago

I don’t know of any examples of images being wholly recreated, but it’s certainly possible to use the name of some living artists to get work in their style. In those cases, it seems like not such a leap to say that the AI has obviously seen that artist’s work and that the output is a derivative work. (The obvious counterargument is that this is the same as a human looking at an artist’s work and aping the style.)

matkoniecz 4y ago
https://alexanderwales.com/wp-content/uploads/2022/08/image.png https://alexanderwales.com/wp-content/uploads/2022/08/image.... Left: “Girl with a Pearl Earring, by Johannes Vermeer” by Stable Diffusion Right: Girl with a Pearl Earring by Johannes Vermeer This specific one is not copyright violation as it is old enough for copyright to expire. But the same may happen with other images. from https://alexanderwales.com/the-ai-art-apocalypse/ https://alexanderwales.com/the-ai-art-apocalypse/ and https://alexanderwales.com/addendum-to-the-ai-art-apocalypse-theft/ https://alexanderwales.com/addendum-to-the-ai-art-apocalypse...
- londons_explore 4y ago
  I think this happens a lot with famous images since that image will be in the training set hundreds of times. Even if deduplication efforts are done, that painting will still be in the background of movie shots etc.
- rtkwe 4y ago
  > Left: “Girl with a Pearl Earring, by Johannes Vermeer” by Stable Diffusion Right: Girl with a Pearl Earring by Johannes Vermeer Even that if done by a person as far as I understand it would not constitute a copyright infringement. It's a separate work mimicking Vermeer's original. The closest real world equivalent I can think of is probably the Obama Hope case by AP vs Shepard Fairy but that settled out of court so we don't really know what the status of that kind of reproduction is legally. On top of that though the SD image isn't just a recoloring with some additions like Fairy's was so it's not quite as close to the original as that case is.

pavlov 4y ago

Stable Diffusion sometimes reproduces the large watermarks used by stock photo providers on their free sample images. That’s embarrassing at the minimum, and potentially a trademark violation.

bigiain 4y ago
Surely at the very least it'd be a TOS violation? I doubt any stock photo service grants you enough rights to redistribute their watermarked free image samples? Especially not in the context of a project like Stable Diffusion?
- Filligree 4y ago
  But it's not reproducing their samples. It's just adding their watermark to newly generated pictures you can't find in the training set.
  - rovr138 4y ago
    If the watermark is their logo or name, it could copyrighted or trademarked.
    - nl 4y ago
      And it's the responsibility of the person using the tool to generate that image not to violate copyright by redistributing it.
      - MereInterest 4y ago
        Just like it's the person's responsibility to only recombine jpeg basis states when they don't correspond to a copyrighted image? It seems more and more to be the case that the trained model is, in large part, a very compact representation of the training data. I'm not seeing a difference between distributing a model that can be used to reconstruct the input images, as opposed to distributing jpeg basis states that can be used to reconstruct the original image.

DannyBee 4y ago

Well no. Code is only protected to the degree it is creative and not functionally driven anyway. So the reduced band of possible expression often directly reduces the protectability-through-copyright.

bigiain 4y ago

> With stable diffusion it's much closer to someone remixing and aping the images than it is reproducing originals. So very similar to how the music industry treats sampling then? Everybody using CoPilot needs to get "code sample clearance" from the original copyright holder before publishing their remix or new program that uses snippets of somebody else's code... Try explaining _that_ to your boss and legal department. "To: <all software dev> Effective immediately, any use of Github is forbidden without prior written approval from both the CTO and General Councel."

kmeisthax 4y ago
This is already a problem with anyone who ever copypastes from Stack Overflow. You're all violating CC-BY-SA[0] and nobody really cares about this. [0] https://stackoverflow.com/help/licensing https://stackoverflow.com/help/licensing
- bscphil 4y ago
  If I ever take any code from SO, I include a comment with a link to it. Surely that's standard practice for anything longer than a line or two?
  - fourthark 4y ago
    I do the same. I think it satisfies BY (attribution) but not SA (Share Alike). As GP says, no one really cares, but it seems hard to satisfy SA... even if you are pasting into open source, is your license compatible with CC? Perhaps I'm over-thinking this.

kmeisthax 4y ago

The reason why it's easy to match Copilot results back to the original source is that the users are starting with prompts that match their public code, deliberately to cause prompt regurgitation. Stable Diffusion actually has a similar problem. Certain terms that directly call up a particular famous painting by name - say, the Mona Lisa[0] - will just produce that painting, possibly tiled on top of itself, and it won't bother with any of the other keywords or phrases you throw at it. The underlying problem is that the AI just outright forgets that it's supposed to create novel works when you give it anything resembling the training set data. If it was just that the AI could spit out training set data when you ask for it, I wouldn't be concerned[1], but this could also happen inadvertently. This would mean that anyone using Copilot to write production code would be risking copyright liability. Through the AI they have access to the entire training set, and the AI has a habit of accidentally producing output that's substantially similar to it. Those are the two prongs of a copyright infringement claim right there. [0] For the record I was trying to get it to draw a picture of the Mona Lisa slapping Yoshikage Kira across the cheek [1] Anyone using an AI system to "launder" creative works is still infringing copyright. AI does not carve a shiny new loophole in the GPL.

xani_ 4y ago
> The reason why it's easy to match Copilot results back to the original source is that the users are starting with prompts that match their public code, deliberately to cause prompt regurgitation. The reason doesn't really matter...
- lofatdairy 4y ago
  GP is just highlighting why this is so common and often a challenging edge case. If you ask it for something that's exactly in its dataset, the "best" solution that minimizes loss will be that existing code. Thus, it's somewhat intrinsic to applying statistical learning to text completion. This means MS really shouldn't have used copyleft code at all, and really shouldn't be selling copilot in this state, but "luckily" for them, short of a class action suit I don't really see any recourse for the programmers who's work they're reselling.

dv_dt 4y ago

I suspect it’s going to be a discussion similar to the introduction of music sampling, followed by a lot of litigation, followed by a settling of law on the matter. The interesting part is if AI will be considered a tooling mechanism much like the tooling used to record and manipulate a music sample into a new composition.