5 ms·
> AI consuming copyrighted data and producing an output has to be considered a derivative work (or indeed, the model itself will be considered a derivative work
by malcolmgreaves 3y ago
> AI consuming copyrighted data and producing an output has to be considered a derivative work (or indeed, the model itself will be considered a derivative work) or IP protections are effectively broken.
It’s not derivative work though. First, a human didn’t create it, so copyright protections don’t exist on its output. Machines don’t enjoy copyright protections, people do.
It’s mechanically copying and reproducing part of its input data set. Making a tool that regurgitates others’ copyrighted IP is going to been seen as aiding mass copyright violations. Exactly like how Napster got sued: they’re holding a bunch of material they shouldn’t be. The only new twist to this case is the data is encoded in a transformer’s weights. This should be correctly seen as the same as having encrypted copyrighted data using a lossy algorithm.
- dilap 3y agoSeems unfair as we converge on AGI. If I memorize the lyrics to a song, is that a copyright violation? The lyrics are encoded in the arrangement of my neurons, after all.
- endisneigh 3y agoI don't know why people use these analogies. No person can memorize terabytes worth of lyrics.
- arthurcolle 3y agoWhat if someone could? What if rainman could
- justinclift 3y agoWhat if rainman could fly too? I guess rainman wouldn't need an airport for short haul trips. So, "good on rainman". Now, for the rest of us...
- dilap 3y agoDoes that invalidate the analogy? The point I'm trying to make is if you rule these behaviors illegal, then you're necessarily making intelligent AI illegal, because humans are capable of the same behaviors.
- endisneigh 3y agoPlagiarism is already a copyright violation so I’m not sure what your point is, and what you’re alluding to is literally what the post is about…
- paulmd 3y agoPlagiarism is not a copyright violation, it’s a violation of the social rules of private (academic) institutions and imposed by those institutions upon their members. If plagiarism was a copyright violation then citing the source wouldn’t make the copyright violation go away. Putting the artists name in the title of the YouTube doesn’t make it not a copyright violation.
- hnaccount141 3y agoThat's not what GP is saying. The (potentially) copyright infringing part is the reproduction of copyrighted material, not the encoding itself. In the same way that learning the lyrics to a song isn't copyright infringement, but performing that song live without permission would be.
- justinclift 3y agoPerhaps the difference is that you're a human doing it, and the other isn't?
- brrrrrm 3y agoI think attempts to tease out this distinction is going to make these laws unwieldy to use in practice
- tticvs 3y ago> First, a human didn’t create it, so copyright protections don’t exist on its output. Machines don’t enjoy copyright protections, people do. This is not correct. AI models are tools that humans use. This is like saying "it was typed on a computer therefore it doesn't enjoy copyright protections"
- 8note 3y agoThis has a ruling, see https://www.theartnewspaper.com/2023/05/04/us-copyright-office-artificial-intelligence-art-regulation https://www.theartnewspaper.com/2023/05/04/us-copyright-offi... When you're using ai models to generate the art, it's not considered human enough. If you then make a bunch of modifications to it, sure, but giving the initial prompt is currently insufficient
- tticvs 3y ago> The Copyright Office granted copyright to the book as a whole but not to the individual images in the book Interesting and makes sense. I guess the best comparison I can think of is a random number generated with Math.random() is not copyrightable. But including that random number is a program is copyrightable. I'd say that a corollary to this is that if the images/words generated by the machine are not copyrightable then they cannot also be _copyright violations_