3 ms·

Hypothetical Question: If someone trained a neural network to create wood textures, and 1 of N training images came from this book, would wood textures generat

by cwkoss 5y ago

Hypothetical Question: If someone trained a neural network to create wood textures, and 1 of N training images came from this book, would wood textures generated by this NN infringe on the author's copyright? To what extent does the size of N change the answer? (I'm sure 1 is no good, but is 10 enough? 100? 1000?)

tshaddox 5y ago
I imagine this would be treated by the legal system the same way any derivative work produced by a human would be.
curtisf 5y ago
(not a lawyer) My understanding of the situation in US copyright law is that the synthesized images would clearly be derivative works, and thus copyright holders have rights to the (portion of the) output which is derivative. However, such a use might be considered fair use. (Fair use is not an exception to copyright itself; it just limits what the copyright holder can stop you from doing) There is no perfect checklist of what constitutes fair use, but there are established criteria that weight in favor or against fair use. Substantiality may weigh in favor of fair use since only some of the input images appears in the output (and less so if mixing many together) Affect on the original work may weigh against fair use -- if you can use the texture synthesizer instead of buying the original texture reference book. I think there's limited precedent that machine learning models can claim fair use of their input text, but I'm guessing the specific case of reference textures => texture models may be hard to argue as fair use (unless you have rights to at least most of the inputs)
- ttt0 5y ago
  > the synthesized images would clearly be derivative works, and thus copyright holders have rights to the (portion of the) output which is derivative. Good luck proving it. So in practice it's basically legal.
  - cwkoss 5y ago
    If they find their image in the training dataset via discovery, that would prove it. I think the trickier part is figuring out that your image was used in the dataset or convincing the judge that you have standing to compel that discovery search. If you have a bunch of get requests from an AI company's corp IP that could do it.
    - ttt0 5y ago
      I'm not familiar with ML, but I assume you can simply get rid of training dataset after the fact, right? You don't have to keep it around forever? > trickier part is figuring out that your image was used in the dataset or convincing the judge that you have standing to compel that discovery search Yes, that's my point.
      - cwkoss 5y ago
        Once the model is working, the training set can be deleted - but it would be hard to modify the model in the future without. I believe there are ways to add to a training set without retraining on the whole thing, but it's trickier and the math is way over my head. I think in many cases people just retrain over the whole corpus if they want to iterate on the model - so most commercial ML models probably hold on to their training datasets and consider it a business asset.
        Mehdi2277 5y ago
        It's very doable to add without it and my experience is I've often worked with models trained online where the data has a retention of a few months so any old enough model will have been trained on data points that don't exist anymore. Large tech companies tend to not keep all data for a mix of storage cost, lower value for some domains (for recommendations freshness is quite useful), and various privacy concerns. Privacy alone means you should have ways to delete data entirely for many problems. If you train with user data and the user requests to delete all info on them through gdpr that includes your training data. I can't remember the amount of time gdpr allows for deletion (I think 30 days or maybe 60), but if you keep a lot of your data with a lower retention than that by default that makes deletion a lot easier. There is other data that has naturally very long retentions (sales orders) where even with deletion requests some aspects should be kept just for financial compliance.

by cwkoss 5y ago

tshaddox 5y ago

I imagine this would be treated by the legal system the same way any derivative work produced by a human would be.

curtisf 5y ago

(not a lawyer) My understanding of the situation in US copyright law is that the synthesized images would clearly be derivative works, and thus copyright holders have rights to the (portion of the) output which is derivative. However, such a use might be considered fair use. (Fair use is not an exception to copyright itself; it just limits what the copyright holder can stop you from doing) There is no perfect checklist of what constitutes fair use, but there are established criteria that weight in favor or against fair use. Substantiality may weigh in favor of fair use since only some of the input images appears in the output (and less so if mixing many together) Affect on the original work may weigh against fair use -- if you can use the texture synthesizer instead of buying the original texture reference book. I think there's limited precedent that machine learning models can claim fair use of their input text, but I'm guessing the specific case of reference textures => texture models may be hard to argue as fair use (unless you have rights to at least most of the inputs)

ttt0 5y ago
> the synthesized images would clearly be derivative works, and thus copyright holders have rights to the (portion of the) output which is derivative. Good luck proving it. So in practice it's basically legal.
- cwkoss 5y ago
  If they find their image in the training dataset via discovery, that would prove it. I think the trickier part is figuring out that your image was used in the dataset or convincing the judge that you have standing to compel that discovery search. If you have a bunch of get requests from an AI company's corp IP that could do it.
  - ttt0 5y ago
    I'm not familiar with ML, but I assume you can simply get rid of training dataset after the fact, right? You don't have to keep it around forever? > trickier part is figuring out that your image was used in the dataset or convincing the judge that you have standing to compel that discovery search Yes, that's my point.
    - cwkoss 5y ago
      Once the model is working, the training set can be deleted - but it would be hard to modify the model in the future without. I believe there are ways to add to a training set without retraining on the whole thing, but it's trickier and the math is way over my head. I think in many cases people just retrain over the whole corpus if they want to iterate on the model - so most commercial ML models probably hold on to their training datasets and consider it a business asset.
      - Mehdi2277 5y ago
        It's very doable to add without it and my experience is I've often worked with models trained online where the data has a retention of a few months so any old enough model will have been trained on data points that don't exist anymore. Large tech companies tend to not keep all data for a mix of storage cost, lower value for some domains (for recommendations freshness is quite useful), and various privacy concerns. Privacy alone means you should have ways to delete data entirely for many problems. If you train with user data and the user requests to delete all info on them through gdpr that includes your training data. I can't remember the amount of time gdpr allows for deletion (I think 30 days or maybe 60), but if you keep a lot of your data with a lower retention than that by default that makes deletion a lot easier. There is other data that has naturally very long retentions (sales orders) where even with deletion requests some aspects should be kept just for financial compliance.