So then I guess I can learn the song, write down the notes from memory, and play that cover. It’s all gone through a neural network and so it’s all fine, right?
I’m not an AI fanboy, but this is kinda a lame take. If the AI produced the same song it heard it would be a cover, sure, and subject to copyright, yes. But most of the time the AIs produce something that is similar to but different from its input.
So yeah, if you listened to a bunch of AC/DC and then wrote a song that sounded like it could be an AC/DC song but isn’t an AC/DC song, that wouldn’t be copyright infringement.
7 notes (2 different ones) were enough for copyright infringement.
The bar for copyright infringement my humans is incredibly low. All you need to infringe on copyright is that your work is “derived” from a copyrighted source work. If you take an original song and change it so that in the end every single note is different, it’s technically still a derived work and still copyright infringement (though it becomes hard to prove that at that point).
If you use the same rule for AI, everything an AI ever outputs is derived works. If you removed all original works from the AI training sets, the AI would do nothing at all.
But for some reason even if the AI outputs whole chapters of books word for word (which most good LLMs can), it’s for some reason not a copyright infringement.
The only reason that’s the case is that the involved judges have no technical understanding and let themselves be bamboozled by fancy new tech. Or because of money exchanged between AI corporations and judges.
Okay hang on. Yes, Ice Ice Baby and Under Pressure don’t have a lot of notes in common, in terms of absolute note count, but when the songs come on, the layperson doesn’t know which is which. Any normal person would listen to 10 seconds of Ice Ice Baby and go “oh yeah, that’s Under Pressure by Queen”.
So yeah, if there’s a prompt that people can use to trick an AI into spitting out a chapter verbatim that’s interesting, but I would say minor infringment. No one is going to read a Ton Clancy novel by systematically tricking the AI to spit out each entire chapter one after the other, and it’s presented to essentially an audience of one, the promoter.
But if I was to take that chapter, the one it spit out verbatim, and put it as a chapter of my book that I published, then yeah, definitely I could be sued for copyright, even if I didn’t do it willingly. Because people would read it and go “oh totally, that’s Pelican Brief”
You are arguing that people not being able to differentiate two songs by a 10 second snippet is bad enough for copyright infringement, but spitting out full chapters is ok because nobody is going to read a full book that way?
Nobody’s going to listen to a Queen song in 10 second snippets either. The intro is similar, but after the intro the song is entirely different. Nobody would mistake e.g. the verse or the refrain of Ice Ice Baby with Under Pressure. It’s two entirely separate genres.
So if the intro sounding kinda similar is in issue, then outputting the intro of a book should be just as bad.
But if I was to take that chapter, the one it spit out verbatim, and put it as a chapter of my book that I published, then yeah, definitely I could be sued for copyright, even if I didn’t do it willingly. Because people would read it and go “oh totally, that’s Pelican Brief”
That’s what I don’t get. If I download a book illegally off piratedbooks.to (made up URL), then not only I but also piratedbooks.to will be liable for the copyright infringement, and generally piratedbooks.to will be much more liable because they are commercially distributing pirated goods.
Now if I ask ChatGPT to output copyrighted content (an action that isn’t much different than “asking” a piracy platform’s website for the content by clicking the “download” button), then OpenAI is for some reason not liable at all for that because it was all me doing this. Even though their platform provided the pirated content.
So, I memorized all the binary digits in the waveform of a Metallica song and wrote it down from memory.
Unfortunately I’ve got a bad memory so I had to do them one digit at a time, but technically I didn’t copy anything, I only wrote down my own memories, right? Also, my handwriting is worse than my doctor’s so I chose to write it on a computer.
The funniest thing happened when I accidentally opened this text file in Winamp. It almost sounded like music, except that the drums were played like shit.
No, but you can learn it
So then I guess I can learn the song, write down the notes from memory, and play that cover. It’s all gone through a neural network and so it’s all fine, right?
I’m not an AI fanboy, but this is kinda a lame take. If the AI produced the same song it heard it would be a cover, sure, and subject to copyright, yes. But most of the time the AIs produce something that is similar to but different from its input.
So yeah, if you listened to a bunch of AC/DC and then wrote a song that sounded like it could be an AC/DC song but isn’t an AC/DC song, that wouldn’t be copyright infringement.
Can I bring to mind Vanilla Ice vs Queen?
7 notes (2 different ones) were enough for copyright infringement.
The bar for copyright infringement my humans is incredibly low. All you need to infringe on copyright is that your work is “derived” from a copyrighted source work. If you take an original song and change it so that in the end every single note is different, it’s technically still a derived work and still copyright infringement (though it becomes hard to prove that at that point).
If you use the same rule for AI, everything an AI ever outputs is derived works. If you removed all original works from the AI training sets, the AI would do nothing at all.
But for some reason even if the AI outputs whole chapters of books word for word (which most good LLMs can), it’s for some reason not a copyright infringement.
The only reason that’s the case is that the involved judges have no technical understanding and let themselves be bamboozled by fancy new tech. Or because of money exchanged between AI corporations and judges.
Okay hang on. Yes, Ice Ice Baby and Under Pressure don’t have a lot of notes in common, in terms of absolute note count, but when the songs come on, the layperson doesn’t know which is which. Any normal person would listen to 10 seconds of Ice Ice Baby and go “oh yeah, that’s Under Pressure by Queen”.
So yeah, if there’s a prompt that people can use to trick an AI into spitting out a chapter verbatim that’s interesting, but I would say minor infringment. No one is going to read a Ton Clancy novel by systematically tricking the AI to spit out each entire chapter one after the other, and it’s presented to essentially an audience of one, the promoter.
But if I was to take that chapter, the one it spit out verbatim, and put it as a chapter of my book that I published, then yeah, definitely I could be sued for copyright, even if I didn’t do it willingly. Because people would read it and go “oh totally, that’s Pelican Brief”
You are arguing that people not being able to differentiate two songs by a 10 second snippet is bad enough for copyright infringement, but spitting out full chapters is ok because nobody is going to read a full book that way?
Nobody’s going to listen to a Queen song in 10 second snippets either. The intro is similar, but after the intro the song is entirely different. Nobody would mistake e.g. the verse or the refrain of Ice Ice Baby with Under Pressure. It’s two entirely separate genres.
So if the intro sounding kinda similar is in issue, then outputting the intro of a book should be just as bad.
That’s what I don’t get. If I download a book illegally off piratedbooks.to (made up URL), then not only I but also piratedbooks.to will be liable for the copyright infringement, and generally piratedbooks.to will be much more liable because they are commercially distributing pirated goods.
Now if I ask ChatGPT to output copyrighted content (an action that isn’t much different than “asking” a piracy platform’s website for the content by clicking the “download” button), then OpenAI is for some reason not liable at all for that because it was all me doing this. Even though their platform provided the pirated content.
So, I memorized all the binary digits in the waveform of a Metallica song and wrote it down from memory.
Unfortunately I’ve got a bad memory so I had to do them one digit at a time, but technically I didn’t copy anything, I only wrote down my own memories, right? Also, my handwriting is worse than my doctor’s so I chose to write it on a computer.
The funniest thing happened when I accidentally opened this text file in Winamp. It almost sounded like music, except that the drums were played like shit.
So, don’t release albums, release tutorials?