5 ms·

The core issue is the content these models get trained on. If you train your model on copyrighted work are you guilty of copyright infringement when the output

The core issue is the content these models get trained on. If you train your model on copyrighted work are you guilty of copyright infringement when the output is used commercially? We are going to see a lot of these cases over the next few years.

alecfreudenberg 3y ago
does Bill Hader spending countless hours watching celebrity scenes to work on his impressions count as training on copyrighted material
- mensetmanusman 3y ago
  No, because you can’t point to the bits where the information is stored.
  - prepend 3y ago
    Certainly you can. You can audit Bill Header’s streaming history to see exactly what he watched and when. But I think that’s immaterial as I think the answer is that it doesn’t matter if you train on copyrighted work. Or maybe better that you don’t need a special license. If I steal books and train on them, then I think that’s copyright infringement not because of the training but because I made an infringing copy. However, if I have a license already to read those books (ie, I bought a copy at a book store) then it’s not infringement to train an AI, or loan it to a million people or whatever I like with it as I bought a copy.
    - Riverheart 3y ago
      “loan it to a million people” Yes but you’re not allowed to copy and distribute so you’re never going to actually do that. Also you’ll pass away with the knowledge you gained unlike the AI you’re training it on which will contain remnants of the work in all copies of itself, however abstracted away you consider it.
      - prepend 3y ago
        I’m allied to make notes and create derived products that are sufficiently distinctive. And I can pass that on to my descendants. I can very legally loan my book or movie to one person a week for all perpetuity so it could be a million people. Does the actual number matter?
interestica 3y ago
Musicians train on copyrighted content and then play music that is similar to that originating content. ...And that's what a musical genre is.
astrange 3y ago
"Using" copyrighted material is not a copyright violation, "copying" it is. So the question is if "training" is copying, but it doesn't really seem like it is.
- bdowling 3y ago
  If training requires making a copy in memory, then that may be copyright infringement.
  - python999 3y ago
    One way to achieve full defensibility could therefore be to train on a human mimic of the corpus?
  - prepend 3y ago
    Me viewing it with my eyes involved making a copy in my neurons. Or viewing it on my television requires multiple copies as the data moves through wires and tuners and memory and storage. Obviously a video just be copied into RAM to process and display for normal viewing. Why would making a copy in memory to run through a training algorithm be different than an algorithm to encode or decode for display on a screen? I don’t think transmission should count as making copies.
    - ChatGTP 3y ago
      A distinction should be made here, you’re not a hard drive which an store perfect copies of information. So it’s not the same thing.
      - prepend 3y ago
        My machine learning system isn’t storing the material. It’s just reading it in and passing it through RAM. It’s different to store material vs stream and process.
        3y ago
asdfman123 3y ago
We’d better go after all the people who read a lot of Vonnegut or listened to a lot of The Beatles and tried to imitate them too, then.
- prepend 3y ago
  I suppose Oasis is really doomed then. I used to think “well this dumb law can’t be so bad if it kills Facebook that’s a net positive” but I don’t think having directions that allow Beatles cover bands like Oasis but disallow an AI Oasis will be good for society. Comically, I think this because one day we’ll be digital beings and it will be funny if these old timey 21st century laws only allow creativity from meat-based consciousnesses because my digital consciousness is basically just an advanced AI that I legally treat as me.
- gumballindie 3y ago
  Those are people not databases. But yes, if you imitate too closely and generate profit it's called ip theft, copyright infringement, or plagiarism.
  - asdfman123 3y ago
    I get your point, but creative people are in fact like databases. They gather as much input as possible, and synthesize output based on that and a few mutations.
  - bsaul 3y ago
    original melody, chords and lyrics never generated copyright infringement no matter what the composer listened to. The problem with the music industry is the industry part : it's been generating standardized, and easily mass produced music for the past 20 years. You can't expect the process to not be fully automated eventually ( i've already felt like some pop or rap song could have been for a long time). You can't expect it not to backfire at some point. And IMHO this is well deserved.
    - gumballindie 3y ago
grumbel 3y ago
I think that depends on intend. If you train the model specifically to clone an artists artstyle or characters, maybe even use their name for marketing (in the style of <famous artist>), than that might be a copyright violation and/or trademark violation. On the other side if you just dump tons of artist into a model to teach it how to draw, it'll be a much harder argument to make when the result doesn't really look like any artist in particular. However a big problem with current models is that their vocabulary is still quite limited, so the only way to get some interesting art style with just a text prompt is to have a "by <artist name>" in the prompt. That's rather fishy. Meanwhile if you custom train a model, the amount of images you put into it will be drastically lower and more focused than what went into into initial training, so that's a bit fishy too. Long term I expect art models to filter trademark terms and just get better language understanding or other means of input (e.g. sketches in ControlNet), so that the model ends up being unable to reproduce something close to an existing work out of the box. It wouldn't stop the user from using that model to perform some copyright violation, but it would require intend and effort on the users part.

alecfreudenberg 3y ago

does Bill Hader spending countless hours watching celebrity scenes to work on his impressions count as training on copyrighted material

mensetmanusman 3y ago
No, because you can’t point to the bits where the information is stored.
- prepend 3y ago
  Certainly you can. You can audit Bill Header’s streaming history to see exactly what he watched and when. But I think that’s immaterial as I think the answer is that it doesn’t matter if you train on copyrighted work. Or maybe better that you don’t need a special license. If I steal books and train on them, then I think that’s copyright infringement not because of the training but because I made an infringing copy. However, if I have a license already to read those books (ie, I bought a copy at a book store) then it’s not infringement to train an AI, or loan it to a million people or whatever I like with it as I bought a copy.
  - Riverheart 3y ago
    “loan it to a million people” Yes but you’re not allowed to copy and distribute so you’re never going to actually do that. Also you’ll pass away with the knowledge you gained unlike the AI you’re training it on which will contain remnants of the work in all copies of itself, however abstracted away you consider it.
    - prepend 3y ago
      I’m allied to make notes and create derived products that are sufficiently distinctive. And I can pass that on to my descendants. I can very legally loan my book or movie to one person a week for all perpetuity so it could be a million people. Does the actual number matter?

interestica 3y ago

Musicians train on copyrighted content and then play music that is similar to that originating content. ...And that's what a musical genre is.

astrange 3y ago

"Using" copyrighted material is not a copyright violation, "copying" it is. So the question is if "training" is copying, but it doesn't really seem like it is.

bdowling 3y ago
If training requires making a copy in memory, then that may be copyright infringement.
- python999 3y ago
  One way to achieve full defensibility could therefore be to train on a human mimic of the corpus?
- prepend 3y ago
  Me viewing it with my eyes involved making a copy in my neurons. Or viewing it on my television requires multiple copies as the data moves through wires and tuners and memory and storage. Obviously a video just be copied into RAM to process and display for normal viewing. Why would making a copy in memory to run through a training algorithm be different than an algorithm to encode or decode for display on a screen? I don’t think transmission should count as making copies.
  - ChatGTP 3y ago
    A distinction should be made here, you’re not a hard drive which an store perfect copies of information. So it’s not the same thing.
    - prepend 3y ago
      My machine learning system isn’t storing the material. It’s just reading it in and passing it through RAM. It’s different to store material vs stream and process.
      - 3y ago

asdfman123 3y ago

We’d better go after all the people who read a lot of Vonnegut or listened to a lot of The Beatles and tried to imitate them too, then.

prepend 3y ago
I suppose Oasis is really doomed then. I used to think “well this dumb law can’t be so bad if it kills Facebook that’s a net positive” but I don’t think having directions that allow Beatles cover bands like Oasis but disallow an AI Oasis will be good for society. Comically, I think this because one day we’ll be digital beings and it will be funny if these old timey 21st century laws only allow creativity from meat-based consciousnesses because my digital consciousness is basically just an advanced AI that I legally treat as me.
gumballindie 3y ago
Those are people not databases. But yes, if you imitate too closely and generate profit it's called ip theft, copyright infringement, or plagiarism.
- asdfman123 3y ago
  I get your point, but creative people are in fact like databases. They gather as much input as possible, and synthesize output based on that and a few mutations.
- bsaul 3y ago
  original melody, chords and lyrics never generated copyright infringement no matter what the composer listened to. The problem with the music industry is the industry part : it's been generating standardized, and easily mass produced music for the past 20 years. You can't expect the process to not be fully automated eventually ( i've already felt like some pop or rap song could have been for a long time). You can't expect it not to backfire at some point. And IMHO this is well deserved.
  - gumballindie 3y ago

grumbel 3y ago

I think that depends on intend. If you train the model specifically to clone an artists artstyle or characters, maybe even use their name for marketing (in the style of <famous artist>), than that might be a copyright violation and/or trademark violation. On the other side if you just dump tons of artist into a model to teach it how to draw, it'll be a much harder argument to make when the result doesn't really look like any artist in particular. However a big problem with current models is that their vocabulary is still quite limited, so the only way to get some interesting art style with just a text prompt is to have a "by <artist name>" in the prompt. That's rather fishy. Meanwhile if you custom train a model, the amount of images you put into it will be drastically lower and more focused than what went into into initial training, so that's a bit fishy too. Long term I expect art models to filter trademark terms and just get better language understanding or other means of input (e.g. sketches in ControlNet), so that the model ends up being unable to reproduce something close to an existing work out of the box. It wouldn't stop the user from using that model to perform some copyright violation, but it would require intend and effort on the users part.