4 ms·

One reading of the latest US Supreme Court ruling on the Andy Warhol estate's case is that the current legal frameworks already prevent this. The ruling focuse

by sfifs 3y ago

One reading of the latest US Supreme Court ruling on the Andy Warhol estate's case is that the current legal frameworks already prevent this. The ruling focused on what was the commercial use case, not on what was the degree of transformation in Andy Warhol's artwork. The opinion appears to say that as the Prince painting was not explicitly licensed for the commercial use case of reproduction in magazine story and could conceivably have been substituted for the purpose of printing in the magazine to accompany a story by the original artist's photograph, it violated copyright. There was even an explicit distinction drawn between this work and the Campbell Soup paintings which were characterized as a social commentary and therefore fair use. There are likely class action lawsuits brewing based on this ruling because today when commercial artists in agencies create layouts, they generally license images & artwork.

judge2020 3y ago
The way I read that, it was only ruled in favor because they were directed to recreate x photo with some changes. If the Scream painting were still copyrighted, it sounds like telling an AI to "create Edvard Munch's The Scream just with lighter colors" would be a copyright violation - but something like "create a new painting in the style of Edvard Munch" would still fall under "artists can't copyright a style".
- bjt 3y ago
  The output isn't the only infringement, arguably. You could also argue (and I expect good lawyers will), that the numeric representation of the artist's works inside the model is already an infringing copy. (Just like the JPEG bytes stored on a server, even without them being blitted to a screen.)
  - zirgs 3y ago
    Models don't store any artist's works. They are way too small to do that.
    - seanhunter 3y ago
      Getty has alleged that Stable Diffusion is sometimes returning some of their copyright images[1]. Even if the model seems too small to directly store the images, it seems at least plausible to me that the parameters can act as a compression such that the model could just output an almost direct copy of an original. I have certainly seen stable diffusion emit images which look like a getty watermark has just been blurred out. [1] https://www.reuters.com/legal/getty-images-lawsuit-says-stability-ai-misused-photos-train-ai-2023-02-06/ https://www.reuters.com/legal/getty-images-lawsuit-says-stab...
      - zirgs 3y ago
        It doesn't store the original images, but it has learned how getty images watermark looks like and where it's located, because it has been repeated millions of times. So it sometimes can return that. This is why it's important to clean up the training dataset. To remove duplicates, images containing watermarks, images that are too similar to each other and so on.
        pxoe 3y ago
        [dead]
    - lewhoo 3y ago
      Models don't store any artist's works. They are way too small to do that. I have close to none knowledge on this subject but I find it very curious and I'd like to know more on that because it seems to me they don't store it but only in the traditional sense. For example, if you could procure a quote, for instance, I asked (chatgpt): - "In Game of Thrones what did Jon Snow say to Arya when he gave her the sword named 'needle' ?", - and it answers: "[...] "Stick 'em with the pointy end. [...]" Then it indicates to me that the information is there. Maybe we should consider that the model actually stores the information but the information is compressed ? Could you ask midjourney to recreate the Mona Lisa of Mickey Mouse ? The right information can't just appear out of thin air. If I recall correctly, someone had some success in identifying and modifying the right neurons or weights of some LLM which changed it "opinion" on where Rome is located ?
  - judge2020 3y ago
    Technically, the DMCA does not limit people from storing copyrighted works, only downloading, distributing and/or reproducing.
torginus 3y ago
Except this is not transformative at all - LLMs are literally trained to be able to output their training data. That's why it produced those substantially plagiarized CNET articles.
- tinco 3y ago
  It's more complex than that, they are trained to model their language and understanding of a concept to match the author. There is not enough memory in the network to remember the actual text. It's like giving a talented an artist a day with a painting, and then a day later ask them to precisely copy the painting from memory. Would they come close enough for it to be considered a forgery, or will it be a transformative reinterpretation? It will probably depend on the skill of the artist, and I feel it might go either way.
  - bugglebeetle 3y ago
    Whether a human artist can do this is largely irrelevant to the question of copyright, however. Selling a close copy of another artists’ work (unless they’re long dead) would likely be infringement.
    - underlipton 3y ago
      I think the distinction is that you don't arrest someone for merely having the capability of reproducing a work, but after they've reproduced it and attempted to use it in an infringing manner. I don't know how applicable the analogy is given that LLMs don't "think", of course.
  - torginus 3y ago
    you might say that, but the literal benchmark of LLMs (or any supervised learning algorithm for that matter) is loss, or how much 'distance' is between their output, and the validation set, when seeded with the training set. With a loss of well below 1%, which is typical, it means it can pretty much recreate the training data.
seanhunter 3y ago
For people who are unaware of this case, here[1] is a writeup. TL;DR: Andy Warhol used a photograph of Prince by Lynn Goldsmith to create a series of original silk screen artworks called "Orange Prince". The supreme court ruled (as I understand it) that Warhol's use was not sufficiently transformative to negate the right of the original copyrightholder. [1] https://www.theguardian.com/artanddesign/2023/may/18/andy-warhol-copyright-prince-paintings-lawsuit https://www.theguardian.com/artanddesign/2023/may/18/andy-wa... Edit to add: Here’s a side-by-side https://www.mondaq.com/images/article_images/1324682a.jpg https://www.mondaq.com/images/article_images/1324682a.jpg . I mean extreme right second from the bottom he hasn’t even changed the colours.

by sfifs 3y ago

judge2020 3y ago

The way I read that, it was only ruled in favor because they were directed to recreate x photo with some changes. If the Scream painting were still copyrighted, it sounds like telling an AI to "create Edvard Munch's The Scream just with lighter colors" would be a copyright violation - but something like "create a new painting in the style of Edvard Munch" would still fall under "artists can't copyright a style".

bjt 3y ago
The output isn't the only infringement, arguably. You could also argue (and I expect good lawyers will), that the numeric representation of the artist's works inside the model is already an infringing copy. (Just like the JPEG bytes stored on a server, even without them being blitted to a screen.)
- zirgs 3y ago
  Models don't store any artist's works. They are way too small to do that.
  - seanhunter 3y ago
    Getty has alleged that Stable Diffusion is sometimes returning some of their copyright images[1]. Even if the model seems too small to directly store the images, it seems at least plausible to me that the parameters can act as a compression such that the model could just output an almost direct copy of an original. I have certainly seen stable diffusion emit images which look like a getty watermark has just been blurred out. [1] https://www.reuters.com/legal/getty-images-lawsuit-says-stability-ai-misused-photos-train-ai-2023-02-06/ https://www.reuters.com/legal/getty-images-lawsuit-says-stab...
    - zirgs 3y ago
      It doesn't store the original images, but it has learned how getty images watermark looks like and where it's located, because it has been repeated millions of times. So it sometimes can return that. This is why it's important to clean up the training dataset. To remove duplicates, images containing watermarks, images that are too similar to each other and so on.
      - pxoe 3y ago
        [dead]
  - lewhoo 3y ago
    Models don't store any artist's works. They are way too small to do that. I have close to none knowledge on this subject but I find it very curious and I'd like to know more on that because it seems to me they don't store it but only in the traditional sense. For example, if you could procure a quote, for instance, I asked (chatgpt): - "In Game of Thrones what did Jon Snow say to Arya when he gave her the sword named 'needle' ?", - and it answers: "[...] "Stick 'em with the pointy end. [...]" Then it indicates to me that the information is there. Maybe we should consider that the model actually stores the information but the information is compressed ? Could you ask midjourney to recreate the Mona Lisa of Mickey Mouse ? The right information can't just appear out of thin air. If I recall correctly, someone had some success in identifying and modifying the right neurons or weights of some LLM which changed it "opinion" on where Rome is located ?
- judge2020 3y ago
  Technically, the DMCA does not limit people from storing copyrighted works, only downloading, distributing and/or reproducing.

torginus 3y ago

Except this is not transformative at all - LLMs are literally trained to be able to output their training data. That's why it produced those substantially plagiarized CNET articles.

tinco 3y ago
It's more complex than that, they are trained to model their language and understanding of a concept to match the author. There is not enough memory in the network to remember the actual text. It's like giving a talented an artist a day with a painting, and then a day later ask them to precisely copy the painting from memory. Would they come close enough for it to be considered a forgery, or will it be a transformative reinterpretation? It will probably depend on the skill of the artist, and I feel it might go either way.
- bugglebeetle 3y ago
  Whether a human artist can do this is largely irrelevant to the question of copyright, however. Selling a close copy of another artists’ work (unless they’re long dead) would likely be infringement.
  - underlipton 3y ago
    I think the distinction is that you don't arrest someone for merely having the capability of reproducing a work, but after they've reproduced it and attempted to use it in an infringing manner. I don't know how applicable the analogy is given that LLMs don't "think", of course.
- torginus 3y ago
  you might say that, but the literal benchmark of LLMs (or any supervised learning algorithm for that matter) is loss, or how much 'distance' is between their output, and the validation set, when seeded with the training set. With a loss of well below 1%, which is typical, it means it can pretty much recreate the training data.

seanhunter 3y ago

For people who are unaware of this case, here[1] is a writeup. TL;DR: Andy Warhol used a photograph of Prince by Lynn Goldsmith to create a series of original silk screen artworks called "Orange Prince". The supreme court ruled (as I understand it) that Warhol's use was not sufficiently transformative to negate the right of the original copyrightholder. [1] https://www.theguardian.com/artanddesign/2023/may/18/andy-warhol-copyright-prince-paintings-lawsuit https://www.theguardian.com/artanddesign/2023/may/18/andy-wa... Edit to add: Here’s a side-by-side https://www.mondaq.com/images/article_images/1324682a.jpg https://www.mondaq.com/images/article_images/1324682a.jpg . I mean extreme right second from the bottom he hasn’t even changed the colours.