3 ms·

I don't know for certain, but I suspect that Google utilized images from the web in training this system. Even if they didn't, suppose they had. I think this ca

by jfoster 11y ago

I don't know for certain, but I suspect that Google utilized images from the web in training this system. Even if they didn't, suppose they had. I think this can raise an interesting question around copyright. In training an AI system with hundreds/thousands of bits of data, no single piece of training data makes much of a difference. If one of my images on the web that I had captioned with the keyword 'dog' was used to train this system about what a dog looks like, is the model they end up with a derivative work of my captioned image? Yes, but my data would make up an infinitesimally small part of that model. Yet, in aggregate, the trained model might almost wholly rely on lots of copyrighted, rights-reserved images. Would the resulting model be a copyright infringement? It would seem as though no rights owner would have a substantial enough claim. Yet, without all of the copyrighted works, perhaps the model would be ineffective.

lallysingh 11y ago
I donno. If I'm composing music, and I heard your music before that, is my music implicitly derived? Your music certainly had some infinitesimal affect on me.
- jfoster 11y ago
  Absolutely, but any music you compose probably also has an element of originality in it, or at least draws inspiration from something that is non-musical in nature. For a trained AI, the entire model is just a derivative work of all the training data. It's not as though the AI had a lover who left them and drew inspiration from that experience to become a more effective AI.
  - lallysingh 11y ago
    So to avoid the copyright situation, I just mix in some random nonce data, infinitesimally small, to make it equivalent?
    - jfoster 11y ago
      Maybe that fixes it. Maybe you don't even have to if the copyright question only ever remains theoretical. I have a feeling it might only be a matter of time until this question arises in a case, though.
  - squeaky-clean 11y ago
    > It's not as though the AI had a lover who left them and drew inspiration from that experience to become a more effective AI. Well then your ex-lover clearly deserves co-songwriting credits. As well as their parents. And anybody who has influenced them personally, or even anyone who made the food that they've eaten. There's gotta be a point at which the original is just too far removed from the end result for it to be infringing, otherwise you could just keep going. Also is a model that classifies things isn't the same as those things themselves, or images of those things, and would most I certainly hope it would be considered transformative enough to not be infringing (I am not a lawyer though). I could give it an image of a dog, and it will tell me what it thinks it is. But there doesn't seem to be any way for me to say "show me a dog" and get back any sort of image, infringing or not. Some thought experiments: Let's say you have a copyrighted photo, and I design an API that allows anyone to upload a photo and get a true/false of whether or not it's the same file as your photo. Is this copyright infringement if I never release the original photo?
  - bduerst 11y ago
    >but any music you compose probably also has an element of originality in it Is deep dream original in your opinion?
brianwawok 11y ago
So people trained google images that Comcast = nazi flag. I guess we need to send in an OCR of a nazi flag and see if it comes back Comcast?
jaxondu 11y ago
By using this API, we're effectively training Google's system to be more and more accurate. Shouldn't Google pay us for using it? Just saying :-)
- bduerst 11y ago
  Not quite. The terms and conditions say that you own the IP and that it isn't used beyond providing the immediate service, meaning they aren't using it to train future models. Besides, it's a call without any feedback, so it's not that valuable as far as training goes.
- ecthiender 11y ago
  Good point!
- skeletonjelly 11y ago
  They're probably already used Google customers as training unknowingly and now it's time to productise it and charge for it.
cromwellian 11y ago
This would be like saying Google's Search Index is a copyright infringement. Or indeed, anything that utilizes any public copyrighted information to infer any details. To me this is taking copyright protection way way too far and the kind of system that could severely gum up progress and innovation.

by jfoster 11y ago

lallysingh 11y ago

I donno. If I'm composing music, and I heard your music before that, is my music implicitly derived? Your music certainly had some infinitesimal affect on me.

jfoster 11y ago
Absolutely, but any music you compose probably also has an element of originality in it, or at least draws inspiration from something that is non-musical in nature. For a trained AI, the entire model is just a derivative work of all the training data. It's not as though the AI had a lover who left them and drew inspiration from that experience to become a more effective AI.
- lallysingh 11y ago
  So to avoid the copyright situation, I just mix in some random nonce data, infinitesimally small, to make it equivalent?
  - jfoster 11y ago
    Maybe that fixes it. Maybe you don't even have to if the copyright question only ever remains theoretical. I have a feeling it might only be a matter of time until this question arises in a case, though.
- squeaky-clean 11y ago
  > It's not as though the AI had a lover who left them and drew inspiration from that experience to become a more effective AI. Well then your ex-lover clearly deserves co-songwriting credits. As well as their parents. And anybody who has influenced them personally, or even anyone who made the food that they've eaten. There's gotta be a point at which the original is just too far removed from the end result for it to be infringing, otherwise you could just keep going. Also is a model that classifies things isn't the same as those things themselves, or images of those things, and would most I certainly hope it would be considered transformative enough to not be infringing (I am not a lawyer though). I could give it an image of a dog, and it will tell me what it thinks it is. But there doesn't seem to be any way for me to say "show me a dog" and get back any sort of image, infringing or not. Some thought experiments: Let's say you have a copyrighted photo, and I design an API that allows anyone to upload a photo and get a true/false of whether or not it's the same file as your photo. Is this copyright infringement if I never release the original photo?
- bduerst 11y ago
  >but any music you compose probably also has an element of originality in it Is deep dream original in your opinion?

brianwawok 11y ago

So people trained google images that Comcast = nazi flag. I guess we need to send in an OCR of a nazi flag and see if it comes back Comcast?

jaxondu 11y ago

By using this API, we're effectively training Google's system to be more and more accurate. Shouldn't Google pay us for using it? Just saying :-)

bduerst 11y ago
Not quite. The terms and conditions say that you own the IP and that it isn't used beyond providing the immediate service, meaning they aren't using it to train future models. Besides, it's a call without any feedback, so it's not that valuable as far as training goes.
ecthiender 11y ago
Good point!
skeletonjelly 11y ago
They're probably already used Google customers as training unknowingly and now it's time to productise it and charge for it.

cromwellian 11y ago

This would be like saying Google's Search Index is a copyright infringement. Or indeed, anything that utilizes any public copyrighted information to infer any details. To me this is taking copyright protection way way too far and the kind of system that could severely gum up progress and innovation.