4 ms·

This is a pretty good summary of the complexity of the issue legally (worth reading the whole thing). The plaintiffs pretty clearly either don’t know how ML wor

This is a pretty good summary of the complexity of the issue legally (worth reading the whole thing). The plaintiffs pretty clearly either don’t know how ML works or are deliberately misconstruing the facts. But the real question is how a judge and/or jury will look at this given how tricky the debate is (as evidenced by comments in the many posts about GPT-like models right here on HN). There are a lot of emotions around this space right now. I’m very curious to see how the legal teams break things down logically and with precedents.

jrm4 4y ago
I see why you're saying what you are about the "facts," but I'm not at all convinced that it's so clear-cut that the plaintiffs are all that wrong. Which is to say, compare these models to e.g. simple image compression. If compressed images can violate, the argument that "this is merely VERY GOOD and VERSATILE image compression" probably has enough legal legs to do something with.
- AbrahamParangi 4y ago
  It is a bit like saying a random number generator is a form of image compression since it is a small program the output of which contains every copyrightable image. Compression has the property that almost all inputs are reconstructable. Generative models have the property that almost all inputs are not reconstructable.
  - pwinnski 4y ago
    Compression with extra steps is still compression. Most people don't understand the math involved in LZF-encoding, and lossy audio encoding seems even more strange, so "diffusion" just sounds like a somewhat-lossier form of compression if described in a certain way. I was very dismissive of the anti-AI-image crowd until I saw that Anne Graham Lotz pic. Only a single known source image of her exists, although many copies of it are distributed far and wide across the internet, and using her name in a prompt results in an image that seems very much like a lossy reproduction of the "original." It's an edge case, to be sure, but that demonstration looks like lossy compression, even though it's not.
    - Turing_Machine 4y ago
      > I was very dismissive of the anti-AI-image crowd until I saw that Anne Graham Lotz pic. Only a single known source image of her exists, although many copies of it are distributed far and wide across the internet https://duckduckgo.com/?q=Anne+Graham+Lotz&t=brave&iax=images&ia=images https://duckduckgo.com/?q=Anne+Graham+Lotz&t=brave&iax=image... I see dozens of images of this woman (who is apparently some kind of minister). Okay, I'm clearly missing something here. Can you explain further?
      - Turing_Machine 4y ago
        Also, besides the 20 or so copies of this specific pose I see on the first page, I also see numerous others of her either in a different pose or in a different (but similar) outfit -- she's apparently fond of that color combination. I just put my local copy of Stable Diffusion to work cranking out images based on an "Anne Graham Lotz" prompt. So far, none of them look anything like the original image. It has generated one of her wearing a similar-colored shirt, but the pose is completely different, and as noted above, she seems to like that color.
      - pwinnski 4y ago
        My mistake! I mis-remembered the original Google paper[0], which stated only that the image was in the Stable Diffusion training set. The complete set of "matches" they got were interesting, even while the number was very low. 0. https://arxiv.org/abs/2301.13188 https://arxiv.org/abs/2301.13188
- sdenton4 4y ago
  From the article: researchers were able to reproduce training images (and bad versions of them at that) 0.03 percent of the time. Not 3 percent, mind you, but 109 out of 350,000 images could be reproduced. And if you go back to the paper, these tended to be images which were heavily duplicated in the training set, so the failure modes should be easy to guard against...
  - jrm4 4y ago
    Sure, but I strongly suspect that in court the visceral impact of "Look, there's the image, it was in there," of even just ONE might just be enough to sway a judge or jury, percentages be darned.
    - sdenton4 4y ago
      Eh, there's a lot of space for argument and deliberation in a courtroom, moreso than a hacker news comment section... This is a rare issue with an obvious fix, so I wouldn't expect a ruling to hinge on it. (Additionally, the Lott image doesn't have any real harm; it's heavily reproduced already and freely enough licensed to appear on Wikipedia.) My guess is that because there's a lot of these cases coming the courts are really going to want to come up with some tests for whether a model is ok or not. So even if they say SD did a bad thing with these few images, that's not going to be enough to rule out the broader class of algorithms - next week they'll have another Getty case against the algorithm with the technical mitigation for the rare copied images.
      - jrm4 4y ago
        But a "rare issue with an obvious fix" sounds like a good Streisand effect type-target to me. Technologically I know that this is a pretty honest statement of the problem, but I'd expect A LOT OF PEOPLE to read this as "oh, look, they're trying to tweak it to 'destroy evidence.'"

jrm4 4y ago

I see why you're saying what you are about the "facts," but I'm not at all convinced that it's so clear-cut that the plaintiffs are all that wrong. Which is to say, compare these models to e.g. simple image compression. If compressed images can violate, the argument that "this is merely VERY GOOD and VERSATILE image compression" probably has enough legal legs to do something with.

AbrahamParangi 4y ago
It is a bit like saying a random number generator is a form of image compression since it is a small program the output of which contains every copyrightable image. Compression has the property that almost all inputs are reconstructable. Generative models have the property that almost all inputs are not reconstructable.
- pwinnski 4y ago
  Compression with extra steps is still compression. Most people don't understand the math involved in LZF-encoding, and lossy audio encoding seems even more strange, so "diffusion" just sounds like a somewhat-lossier form of compression if described in a certain way. I was very dismissive of the anti-AI-image crowd until I saw that Anne Graham Lotz pic. Only a single known source image of her exists, although many copies of it are distributed far and wide across the internet, and using her name in a prompt results in an image that seems very much like a lossy reproduction of the "original." It's an edge case, to be sure, but that demonstration looks like lossy compression, even though it's not.
  - Turing_Machine 4y ago
    > I was very dismissive of the anti-AI-image crowd until I saw that Anne Graham Lotz pic. Only a single known source image of her exists, although many copies of it are distributed far and wide across the internet https://duckduckgo.com/?q=Anne+Graham+Lotz&t=brave&iax=images&ia=images https://duckduckgo.com/?q=Anne+Graham+Lotz&t=brave&iax=image... I see dozens of images of this woman (who is apparently some kind of minister). Okay, I'm clearly missing something here. Can you explain further?
    - Turing_Machine 4y ago
      Also, besides the 20 or so copies of this specific pose I see on the first page, I also see numerous others of her either in a different pose or in a different (but similar) outfit -- she's apparently fond of that color combination. I just put my local copy of Stable Diffusion to work cranking out images based on an "Anne Graham Lotz" prompt. So far, none of them look anything like the original image. It has generated one of her wearing a similar-colored shirt, but the pose is completely different, and as noted above, she seems to like that color.
    - pwinnski 4y ago
      My mistake! I mis-remembered the original Google paper[0], which stated only that the image was in the Stable Diffusion training set. The complete set of "matches" they got were interesting, even while the number was very low. 0. https://arxiv.org/abs/2301.13188 https://arxiv.org/abs/2301.13188
sdenton4 4y ago
From the article: researchers were able to reproduce training images (and bad versions of them at that) 0.03 percent of the time. Not 3 percent, mind you, but 109 out of 350,000 images could be reproduced. And if you go back to the paper, these tended to be images which were heavily duplicated in the training set, so the failure modes should be easy to guard against...
- jrm4 4y ago
  Sure, but I strongly suspect that in court the visceral impact of "Look, there's the image, it was in there," of even just ONE might just be enough to sway a judge or jury, percentages be darned.
  - sdenton4 4y ago
    Eh, there's a lot of space for argument and deliberation in a courtroom, moreso than a hacker news comment section... This is a rare issue with an obvious fix, so I wouldn't expect a ruling to hinge on it. (Additionally, the Lott image doesn't have any real harm; it's heavily reproduced already and freely enough licensed to appear on Wikipedia.) My guess is that because there's a lot of these cases coming the courts are really going to want to come up with some tests for whether a model is ok or not. So even if they say SD did a bad thing with these few images, that's not going to be enough to rule out the broader class of algorithms - next week they'll have another Getty case against the algorithm with the technical mitigation for the rare copied images.
    - jrm4 4y ago
      But a "rare issue with an obvious fix" sounds like a good Streisand effect type-target to me. Technologically I know that this is a pretty honest statement of the problem, but I'd expect A LOT OF PEOPLE to read this as "oh, look, they're trying to tweak it to 'destroy evidence.'"