6 ms·

> but surely you can't just... use stock photos without paying for the license? They aren't hosting the infringing content. Training on the data is probably co

> but surely you can't just... use stock photos without paying for the license? They aren't hosting the infringing content. Training on the data is probably covered under fair use. Generations are of _learned_ representations of the dataset, not the dataset itself. This makes it closer to outputting original works (probably owned by the person who used the model). The players involved here are known for being litigious, however. I wouldn't be surprised if OpenAI did in fact pay some hefty fee upfront to get full permission to use these images.

kej 4y ago
If they had been paying for the images upfront, wouldn't you expect them to train the model on the non-watermarked versions?
- namrog84 4y ago
  The watermarked version might be more prolific with better metatags and descriptions around them. The non watermarked versions are likely internal only and have far less diverse descriptions.
- ShamelessC 4y ago
  Good point! They certainly had no obligation to pay, either. Perhaps they just scraped it all.
gricardo99 4y ago
if they paid for access, or permission, why train on the watermark versions? I’m guessing they assumed fair use and there will be lawsuits.
- criddell 4y ago
  Is that representation of the watermark a trademark? If so, then copyright infringement might not matter, but use of the trademark may.
BeefWellington 4y ago
> Training on the data is probably covered under fair use. Generations are of _learned_ representations of the dataset, not the dataset itself. This makes it closer to outputting original works (probably owned by the person who used the model). "Probably" is doing a lot of heavy lifting in that sentence. As for "_learned_", that's pretty debatable considering it's reproducing recognizable trademark infringement. > The players involved here are known for being litigious, however. I wouldn't be surprised if OpenAI did in fact pay some hefty fee upfront to get full permission to use these images. I have no idea why anyone would assume the "move fast and break things" disruption mindset that pervades tech companies these days, especially in spaces like ML/"AI", would mean they considered the legality, ethics, or good business sense of their training dataset. As with Copilot, I suspect the DALL-E terms of use puts the onus on the user to avoid using infringing items.
- yreg 4y ago
  Unlike Copilot, DALL-E et al. don't produce verbatim copies of trained data. Copying ideas and styles has always been a fundamental part of art history, so an artwork right holder might have a hard time successfuly sueing a user for the user's generated image looking similar to the right holder's artwork.
  - BeefWellington 4y ago
    "Verbatim" is an interesting term since I'm not certain it matters. In this case OP here demonstrated DALL-E generating a trademarked watermark on top of an image. I doubt the courts, looking at that, would believe that that's not close enough to their trademark to infringe. The art world's copyright suits are all over the place in terms of what's sufficient to meet the threshold of "fair use" or "not a copy". It's hard for me as a layperson to see works by Richard Prince[1] as substantially transformative (clearly one work is derived from the other) and even the different courts couldn't agree on this as it was initially found in favor of the plaintiffs but then Prince won his appeal. My approach to this kind of thing is simply this: Does this technology inherently open me up to lawsuits in undecided or highly unreliable legal territory? If yes, steer well clear of using it in any capacity. [1]: https://www.artnews.com/art-in-america/features/richard-prince-wins-major-victory-in-landmark-copyright-suit-59404/ https://www.artnews.com/art-in-america/features/richard-prin...
chrismorgan 4y ago
I would be very surprised if OpenAI paid anything for these, because it would set precedent that copyright infringement was applicable, which would be fatal down the road. (The only argument they could possibly mount in their defence would be that they wanted to train on the original images without watermarks.)
bhedgeoser 4y ago
What if I write a machine learning algorithm that only generates images that it has seen in the training dataset, with one pixel slightly different.
- can16358p 4y ago
  It won't be transformative enough and you'd probably lose the case. (IANAL)
  - sh4rks 4y ago
    What about two pixels?
    - can16358p 4y ago
      Not enough... three... four... I think at some point there's a blurry gray area where a human judge would decide it is infringement or not. Of course not with a few pixels but at whole image level.
whywhywhywhy 4y ago
What if my dataset is just the one Getty image I don’t want to pay for.

kej 4y ago

If they had been paying for the images upfront, wouldn't you expect them to train the model on the non-watermarked versions?

namrog84 4y ago
The watermarked version might be more prolific with better metatags and descriptions around them. The non watermarked versions are likely internal only and have far less diverse descriptions.
ShamelessC 4y ago
Good point! They certainly had no obligation to pay, either. Perhaps they just scraped it all.

gricardo99 4y ago

if they paid for access, or permission, why train on the watermark versions? I’m guessing they assumed fair use and there will be lawsuits.

criddell 4y ago
Is that representation of the watermark a trademark? If so, then copyright infringement might not matter, but use of the trademark may.

BeefWellington 4y ago

> Training on the data is probably covered under fair use. Generations are of _learned_ representations of the dataset, not the dataset itself. This makes it closer to outputting original works (probably owned by the person who used the model). "Probably" is doing a lot of heavy lifting in that sentence. As for "_learned_", that's pretty debatable considering it's reproducing recognizable trademark infringement. > The players involved here are known for being litigious, however. I wouldn't be surprised if OpenAI did in fact pay some hefty fee upfront to get full permission to use these images. I have no idea why anyone would assume the "move fast and break things" disruption mindset that pervades tech companies these days, especially in spaces like ML/"AI", would mean they considered the legality, ethics, or good business sense of their training dataset. As with Copilot, I suspect the DALL-E terms of use puts the onus on the user to avoid using infringing items.

yreg 4y ago
Unlike Copilot, DALL-E et al. don't produce verbatim copies of trained data. Copying ideas and styles has always been a fundamental part of art history, so an artwork right holder might have a hard time successfuly sueing a user for the user's generated image looking similar to the right holder's artwork.
- BeefWellington 4y ago
  "Verbatim" is an interesting term since I'm not certain it matters. In this case OP here demonstrated DALL-E generating a trademarked watermark on top of an image. I doubt the courts, looking at that, would believe that that's not close enough to their trademark to infringe. The art world's copyright suits are all over the place in terms of what's sufficient to meet the threshold of "fair use" or "not a copy". It's hard for me as a layperson to see works by Richard Prince[1] as substantially transformative (clearly one work is derived from the other) and even the different courts couldn't agree on this as it was initially found in favor of the plaintiffs but then Prince won his appeal. My approach to this kind of thing is simply this: Does this technology inherently open me up to lawsuits in undecided or highly unreliable legal territory? If yes, steer well clear of using it in any capacity. [1]: https://www.artnews.com/art-in-america/features/richard-prince-wins-major-victory-in-landmark-copyright-suit-59404/ https://www.artnews.com/art-in-america/features/richard-prin...

chrismorgan 4y ago

I would be very surprised if OpenAI paid anything for these, because it would set precedent that copyright infringement was applicable, which would be fatal down the road. (The only argument they could possibly mount in their defence would be that they wanted to train on the original images without watermarks.)

bhedgeoser 4y ago

What if I write a machine learning algorithm that only generates images that it has seen in the training dataset, with one pixel slightly different.

can16358p 4y ago
It won't be transformative enough and you'd probably lose the case. (IANAL)
- sh4rks 4y ago
  What about two pixels?
  - can16358p 4y ago
    Not enough... three... four... I think at some point there's a blurry gray area where a human judge would decide it is infringement or not. Of course not with a few pixels but at whole image level.

whywhywhywhy 4y ago

What if my dataset is just the one Getty image I don’t want to pay for.