9 ms·

Previous rulings had found that when code itself was copyrighted due to being a work of art (of human labor), the output of that code was indeed copyrightable i

by thrill 3y ago

Previous rulings had found that when code itself was copyrighted due to being a work of art (of human labor), the output of that code was indeed copyrightable if the program did the majority of the work in producing that output. This ruling seems at odds with that previous interpretation. https://mwhlawgroup.com/can-copyright-in-software-extend-to-its-output/ https://mwhlawgroup.com/can-copyright-in-software-extend-to-...

brookst 3y ago
Are weights code?
- jrajav 3y ago
  No, weights are data, just like transformation matrices used by graphics code. The data has no meaning without the code, and the code is what creates and manages it based on its inputs, as an aid to generating some output. So it's perfectly accurate to talk about the code as the main (or only) actual mechanism.
  - halfcat 3y ago
    The distinction that data has no meaning without code is unconvincing. The code has no meaning without a compiler or interpreter. The compiler and interpreter have no meaning without a CPU. If I store x86 CPU instructions in a Postgres database, is that data or code? It’s also a worth viewing this in light of different computing models. The main mechanism may look quite different if you’re looking at a CPU, GPU, or DSP model of computing.
    - somesoftdev 3y ago
      I believe code holds meaning even without a compiler or interpreter. I can read code and discern its functionality. I can even recognize existing algorithms such as quicksort. I can achieve the same understanding with pseudocode, which lacks an existing compiler. Similarly, prompts possess meaning; I can comprehend the task delegated to a machine learning model by reading the prompt. I agree that both code and data are akin in terms of meaning. Transitioning from Assembly language to Python liberates us from concerns about registers and low-level intricacies. Python directs our attention to the problems we aim to solve, in contrast to low-level languages where a significant effort is spent on specifying hardware operations. High-level languages emphasize outcomes, enabling us to focus on the result rather than meticulously instructing the hardware. AI propels us further along this trajectory. In this sense, I view a prompt and code as rather similar.
  - oneshtein 3y ago
    Transformation matrices are code. They are instructions how to transform something into something else. However, code is just data for a compiler.
  - raincole 3y ago
    > The data has no meaning without the code Data has no meaning without spec. It's a quite big difference. Code has no meaning without spec too. var a = 1 + 1 << 2; wihout a spec how do you know "+" is not overridden? Maybe it means string concat and implicit casting and 1 + 1 is "11"? How do you know the precedence of operators? << runs first or +?
- geertj 3y ago
  Why couldn’t they be? Using a Gödel like encoding every program is also a number. So it seems that numbers that are complex enough that they are isomorphic to a copyrightable program should be copyrightable. I tend to think of the neural net weights as a different instruction set, just how Intel and ARM are also different instruction sets. I fully well realize that lawmakers and the judiciary will be the final arbiter here (although I hope some logic will be part of the process).
- regularfry 3y ago
  Yes, for a rather odd VM.
yjk 3y ago
From the link: 'It was suggested that copyright protection afforded to the computer program may also extend to the output files if the program does the “lion’s share of the work” in creating the output files and the user’s input is “marginal.”' Even if you consider the user input to be marginal, the issue from the link is the copyright of the program extending to the code, not that of the user.
nerdponx 3y ago
That's a fascinating ruling. Forget AI for a minute. Let's say I write a program to generate cool images using Penrose tiles or space filling curves or some other algorithmic process, maybe deterministic, maybe (pseudo-)random. Under that ruling, it sounds like all of the outputs of my program would also fall under the copyright of the original program, unless the creator could prove that they did substantial work in addition to what the program did. Which then leads to the next question of: what is the "lion's share"? Does it matter if I randomly punch in some parameters and get a nice result on the first try, or if I spend hours trying different parameter combinations to find some thing that looks the way I want it to? Even if the outcome is still considered owned by the software author, is my particular parameter choice copyrightable by me, being my own creative output? If you think about it, the modern generative "AI" systems are more or less the same thing, but where the input parameter space is something like the entire space of written language, rather than a couple drop-down menus with algorithm settings and a seed for a PRNG. Does it matter if I slap in a simple prompt and get a nice result, or if I spent hours engineering a perfect prompt? If anything, it should be unambiguous that the prompts are copyrightable just like any other written work, being actual free-form human-readable text. So at least AI artists have that going for them. It will be really interesting to see what happens as more court rulings unfold.
- sircastor 3y ago
  Suppose I manufacture a compass or a protractor that you use to make beautiful curves in your art - surely my effort is greater than yours. You cannot create an accurate arc without my tool. Surely I deserve the lion's share of credit and profit. Incidentally, my friend the paintbrush maker would like to speak with you. I'm being facetious of course. I suspect this technology is going to rapidly push us into radically rethinking how copyright works, but it'll be delayed some 20-30 years by incumbents desperately clinging to their hordes of existing IP.
  - ChrisMarshallNY 3y ago
    Well, we have a bunch of issues, here. First, there's a bunch of (non-creative) people that just want free creative assets. They don't like it, when creatives insist on being paid for the [considerable] time, effort, education and experience that goes into the creative asset. Those people are jumping for joy. Then, we have the creatives, themselves, who may have attended the correct schools, and have the appropriate education, to produce quality assets. They don't want people to just take their work. This looks like a side-jacking. If someone sees a nice picture that was done by a creative, then they can prompt MidJourney to produce one that looks a lot like it, for the cost of the MJ subscription. They don't particularly care (yet)[0] about copyright. Then, we have the portfolio holders, who are often not creatives, but still hold the copyrights on a lot of creative work. Think Disney, or Universal. Their portfolios are an investment. They don't want to lose that investment. Then, we have the crooks, who want to make cheap, crappy copies of creative assets, regardless of who owns the copyright. At the moment, there is an element of risk, as they could get caught and punished. They are also jumping for joy. [0] I said "yet," because one of the problems with a "wild west" copyright and trademark atmosphere, is that it is extremely easy to make "almost-like" copies, and if you don't protect your IP (or can't), then it is "up for grabs." This is done now, but it's illegal, and can be stopped; especially if trademark is involved. I was at a Chinese restaurant, once, and was listening to the music. I noticed that there were many strains of hugely popular (and copyrighted) songs, like Scarborough Fair, and Moondance, all mashed together into a continuous stream. The owner confirmed what I suspected. He subscribed to a service that sends legal "muzak" to his restaurant. He had been mugged by ASCAP, once, and never wanted to go through that again. I've noticed that "free everything" people often insist that everything except their own work should be "freed."
phkahler 3y ago
>> Previous rulings had found that when code itself was copyrighted due to being a work of art (of human labor) Traditionally code would be written as a tool to achieve a specific output. If you automate the task of generating a particular piece of code, you still did all the work, or at least defined it pprecisely. While I don't see AI as an independent mind (yet) it seems independent enough from the person who prompts it to say the prompter doesn't own the output. I'm so 100 percent on what I just wrote, but it seems reasonable enough to say the precedent you mention may not apply.
- bryanrasmussen 3y ago
  I'm pretty sure if the output of the code could be seen as deterministic, that is to say the code just makes what will be produced 1000 times quicker, then the work would be copyrightable but if the output is not deterministic then we cannot say to whom the copyright should belong.
  - cma 3y ago
    What if you roll dice in making choices in a traditional work?
  - randombits0 3y ago
    Only creative expressions are copyrightable. Purely functional expressions are not.
- regularfry 3y ago
  Different area of law, but I'm pretty sure computer generated ideas have been patented before now.
TrueDuality 3y ago
The code here isn't producing a specific, designed, or planned output. It is effectively doing refined data analysis and interpretation of an intermediate format (the weights) that is exclusively dependent on its input data not on the code itself. I don't see any inconsistency in this ruling.
space_fountain 3y ago
Here the programmer was trying to make a point that AIs deserve to be marked as authors. I think this ruling is much less broad then the article is implying. Presumably an argument that the programmer deserved the copyright in deference to the creative work used to make the algorithm or that the AI is just a tool like photoshop or a camera might have worked, but the filer here wasn’t interested in that line of thinking and wanted to open up copyright to machines
- __loam 3y ago
  It's good that he failed. AI personhood is a ridiculous can of worms that we should avoid for as long as we can. Just look at the disaster that is corporate personhood.
  - Ekaros 3y ago
    Also, if he didn't wouldn't basically automated camera be able to be set as author of photographs it takes? Actually that would be interesting hack for older people. As it could give longer period of protection than their natural lifespan.
  - tadfisher 3y ago
    That disaster is the legal basis (in common law) for natural persons to sue corporate persons, and for governments to hold corporations accountable to laws separately from every individual member of the corporation, making it much easier to prosecute them.
    - yarhgoi 3y ago
      Conveniently shielding human perpetrators of white collar crime. This ephemeral fake world of contemporary society is not some sacrosanct construct, or divine mandate. There is no immutable obligation to preserve it.
dragonwriter 3y ago
It is not, the filer here claimed no role in the creation except as the employer of the AI in a work for hire, and was attempting to overturn the Registrar of Copyrights determination that AI could not be an author and that a registration listing the AI as the author was facially invalid. This case is fundamentally not about whether art in which AI is used as a tool by a human (even if the human role is entirely in programming the AI) is copyrightable as a work of the human, it is solely about whether a piece of software can be an author under copyright law, a very different question.
aaron695 3y ago
[dead]
layer8 3y ago
With LLMs, arguably the code is much less decisive for the specific output than the model, which is data. An interesting question to ask is which human labor contributed to the model. The training data probably would constitute a significant share in that.
rcxdude 3y ago
There's a subtle distinction which is probably important to seperate the two situations: The case referenced in that article (https://law.justia.com/cases/federal/district-courts/FSupp2/136/276/2579869/ https://law.justia.com/cases/federal/district-courts/FSupp2/...) is to do with the output of a program which has a specific intended input (the text of the Torah) and a completely deterministic output from that input (a bunch of matrixes of a 'bible code'), and the argument is that therefore the output is essentially already represented by the input and the program, and because the output is an inevitable result of the program (which is copyrighted), and the input (which is public domain), the writer of the program has copyright over the output as well. Compare with generative art (even pre-AI) where the output is based on a random seed and any given output is not an completely inevitable result of running the program. In this case the US copyright office at least claims that there is no human authorship, by the writer of the program or by the prompter (at least if the prompter takes no further action with the generated output): https://www.federalregister.gov/documents/2023/03/16/2023-05321/copyright-registration-guidance-works-containing-material-generated-by-artificial-intelligence https://www.federalregister.gov/documents/2023/03/16/2023-05... This is of course up for interpretation by a judge (though I think the US copyright office refusing registration may make it de-facto impossible to get it in front of a judge, as I think it's a requirement to actually bring a suit against someone for infringement), but I suspect this is the kind of distinction that they will consider regarding this particular precedent.
- jcranmer 3y ago
  > (though I think the US copyright office refusing registration may make it de-facto impossible to get it in front of a judge, as I think it's a requirement to actually bring a suit against someone for infringement) The copyright office refusing registration counts as copyright registration for the purposes of register-before-suing-for-infringement.
  - zerocrates 3y ago
    You can also just sue the Register of Copyrights for rejecting you after exhausting the internal appeals process, which is what happened here.

by thrill 3y ago

brookst 3y ago

Are weights code?

jrajav 3y ago
No, weights are data, just like transformation matrices used by graphics code. The data has no meaning without the code, and the code is what creates and manages it based on its inputs, as an aid to generating some output. So it's perfectly accurate to talk about the code as the main (or only) actual mechanism.
- halfcat 3y ago
  The distinction that data has no meaning without code is unconvincing. The code has no meaning without a compiler or interpreter. The compiler and interpreter have no meaning without a CPU. If I store x86 CPU instructions in a Postgres database, is that data or code? It’s also a worth viewing this in light of different computing models. The main mechanism may look quite different if you’re looking at a CPU, GPU, or DSP model of computing.
  - somesoftdev 3y ago
    I believe code holds meaning even without a compiler or interpreter. I can read code and discern its functionality. I can even recognize existing algorithms such as quicksort. I can achieve the same understanding with pseudocode, which lacks an existing compiler. Similarly, prompts possess meaning; I can comprehend the task delegated to a machine learning model by reading the prompt. I agree that both code and data are akin in terms of meaning. Transitioning from Assembly language to Python liberates us from concerns about registers and low-level intricacies. Python directs our attention to the problems we aim to solve, in contrast to low-level languages where a significant effort is spent on specifying hardware operations. High-level languages emphasize outcomes, enabling us to focus on the result rather than meticulously instructing the hardware. AI propels us further along this trajectory. In this sense, I view a prompt and code as rather similar.
- oneshtein 3y ago
  Transformation matrices are code. They are instructions how to transform something into something else. However, code is just data for a compiler.
- raincole 3y ago
  > The data has no meaning without the code Data has no meaning without spec. It's a quite big difference. Code has no meaning without spec too. var a = 1 + 1 << 2; wihout a spec how do you know "+" is not overridden? Maybe it means string concat and implicit casting and 1 + 1 is "11"? How do you know the precedence of operators? << runs first or +?
geertj 3y ago
Why couldn’t they be? Using a Gödel like encoding every program is also a number. So it seems that numbers that are complex enough that they are isomorphic to a copyrightable program should be copyrightable. I tend to think of the neural net weights as a different instruction set, just how Intel and ARM are also different instruction sets. I fully well realize that lawmakers and the judiciary will be the final arbiter here (although I hope some logic will be part of the process).
regularfry 3y ago
Yes, for a rather odd VM.

yjk 3y ago

From the link: 'It was suggested that copyright protection afforded to the computer program may also extend to the output files if the program does the “lion’s share of the work” in creating the output files and the user’s input is “marginal.”' Even if you consider the user input to be marginal, the issue from the link is the copyright of the program extending to the code, not that of the user.

nerdponx 3y ago

That's a fascinating ruling. Forget AI for a minute. Let's say I write a program to generate cool images using Penrose tiles or space filling curves or some other algorithmic process, maybe deterministic, maybe (pseudo-)random. Under that ruling, it sounds like all of the outputs of my program would also fall under the copyright of the original program, unless the creator could prove that they did substantial work in addition to what the program did. Which then leads to the next question of: what is the "lion's share"? Does it matter if I randomly punch in some parameters and get a nice result on the first try, or if I spend hours trying different parameter combinations to find some thing that looks the way I want it to? Even if the outcome is still considered owned by the software author, is my particular parameter choice copyrightable by me, being my own creative output? If you think about it, the modern generative "AI" systems are more or less the same thing, but where the input parameter space is something like the entire space of written language, rather than a couple drop-down menus with algorithm settings and a seed for a PRNG. Does it matter if I slap in a simple prompt and get a nice result, or if I spent hours engineering a perfect prompt? If anything, it should be unambiguous that the prompts are copyrightable just like any other written work, being actual free-form human-readable text. So at least AI artists have that going for them. It will be really interesting to see what happens as more court rulings unfold.

sircastor 3y ago
Suppose I manufacture a compass or a protractor that you use to make beautiful curves in your art - surely my effort is greater than yours. You cannot create an accurate arc without my tool. Surely I deserve the lion's share of credit and profit. Incidentally, my friend the paintbrush maker would like to speak with you. I'm being facetious of course. I suspect this technology is going to rapidly push us into radically rethinking how copyright works, but it'll be delayed some 20-30 years by incumbents desperately clinging to their hordes of existing IP.
- ChrisMarshallNY 3y ago
  Well, we have a bunch of issues, here. First, there's a bunch of (non-creative) people that just want free creative assets. They don't like it, when creatives insist on being paid for the [considerable] time, effort, education and experience that goes into the creative asset. Those people are jumping for joy. Then, we have the creatives, themselves, who may have attended the correct schools, and have the appropriate education, to produce quality assets. They don't want people to just take their work. This looks like a side-jacking. If someone sees a nice picture that was done by a creative, then they can prompt MidJourney to produce one that looks a lot like it, for the cost of the MJ subscription. They don't particularly care (yet)[0] about copyright. Then, we have the portfolio holders, who are often not creatives, but still hold the copyrights on a lot of creative work. Think Disney, or Universal. Their portfolios are an investment. They don't want to lose that investment. Then, we have the crooks, who want to make cheap, crappy copies of creative assets, regardless of who owns the copyright. At the moment, there is an element of risk, as they could get caught and punished. They are also jumping for joy. [0] I said "yet," because one of the problems with a "wild west" copyright and trademark atmosphere, is that it is extremely easy to make "almost-like" copies, and if you don't protect your IP (or can't), then it is "up for grabs." This is done now, but it's illegal, and can be stopped; especially if trademark is involved. I was at a Chinese restaurant, once, and was listening to the music. I noticed that there were many strains of hugely popular (and copyrighted) songs, like Scarborough Fair, and Moondance, all mashed together into a continuous stream. The owner confirmed what I suspected. He subscribed to a service that sends legal "muzak" to his restaurant. He had been mugged by ASCAP, once, and never wanted to go through that again. I've noticed that "free everything" people often insist that everything except their own work should be "freed."

phkahler 3y ago

>> Previous rulings had found that when code itself was copyrighted due to being a work of art (of human labor) Traditionally code would be written as a tool to achieve a specific output. If you automate the task of generating a particular piece of code, you still did all the work, or at least defined it pprecisely. While I don't see AI as an independent mind (yet) it seems independent enough from the person who prompts it to say the prompter doesn't own the output. I'm so 100 percent on what I just wrote, but it seems reasonable enough to say the precedent you mention may not apply.

bryanrasmussen 3y ago
I'm pretty sure if the output of the code could be seen as deterministic, that is to say the code just makes what will be produced 1000 times quicker, then the work would be copyrightable but if the output is not deterministic then we cannot say to whom the copyright should belong.
- cma 3y ago
  What if you roll dice in making choices in a traditional work?
- randombits0 3y ago
  Only creative expressions are copyrightable. Purely functional expressions are not.
regularfry 3y ago
Different area of law, but I'm pretty sure computer generated ideas have been patented before now.

TrueDuality 3y ago

The code here isn't producing a specific, designed, or planned output. It is effectively doing refined data analysis and interpretation of an intermediate format (the weights) that is exclusively dependent on its input data not on the code itself. I don't see any inconsistency in this ruling.

space_fountain 3y ago

Here the programmer was trying to make a point that AIs deserve to be marked as authors. I think this ruling is much less broad then the article is implying. Presumably an argument that the programmer deserved the copyright in deference to the creative work used to make the algorithm or that the AI is just a tool like photoshop or a camera might have worked, but the filer here wasn’t interested in that line of thinking and wanted to open up copyright to machines

__loam 3y ago
It's good that he failed. AI personhood is a ridiculous can of worms that we should avoid for as long as we can. Just look at the disaster that is corporate personhood.
- Ekaros 3y ago
  Also, if he didn't wouldn't basically automated camera be able to be set as author of photographs it takes? Actually that would be interesting hack for older people. As it could give longer period of protection than their natural lifespan.
- tadfisher 3y ago
  That disaster is the legal basis (in common law) for natural persons to sue corporate persons, and for governments to hold corporations accountable to laws separately from every individual member of the corporation, making it much easier to prosecute them.
  - yarhgoi 3y ago
    Conveniently shielding human perpetrators of white collar crime. This ephemeral fake world of contemporary society is not some sacrosanct construct, or divine mandate. There is no immutable obligation to preserve it.

dragonwriter 3y ago

It is not, the filer here claimed no role in the creation except as the employer of the AI in a work for hire, and was attempting to overturn the Registrar of Copyrights determination that AI could not be an author and that a registration listing the AI as the author was facially invalid. This case is fundamentally not about whether art in which AI is used as a tool by a human (even if the human role is entirely in programming the AI) is copyrightable as a work of the human, it is solely about whether a piece of software can be an author under copyright law, a very different question.

aaron695 3y ago

[dead]

layer8 3y ago

With LLMs, arguably the code is much less decisive for the specific output than the model, which is data. An interesting question to ask is which human labor contributed to the model. The training data probably would constitute a significant share in that.

rcxdude 3y ago

There's a subtle distinction which is probably important to seperate the two situations: The case referenced in that article (https://law.justia.com/cases/federal/district-courts/FSupp2/136/276/2579869/ https://law.justia.com/cases/federal/district-courts/FSupp2/...) is to do with the output of a program which has a specific intended input (the text of the Torah) and a completely deterministic output from that input (a bunch of matrixes of a 'bible code'), and the argument is that therefore the output is essentially already represented by the input and the program, and because the output is an inevitable result of the program (which is copyrighted), and the input (which is public domain), the writer of the program has copyright over the output as well. Compare with generative art (even pre-AI) where the output is based on a random seed and any given output is not an completely inevitable result of running the program. In this case the US copyright office at least claims that there is no human authorship, by the writer of the program or by the prompter (at least if the prompter takes no further action with the generated output): https://www.federalregister.gov/documents/2023/03/16/2023-05321/copyright-registration-guidance-works-containing-material-generated-by-artificial-intelligence https://www.federalregister.gov/documents/2023/03/16/2023-05... This is of course up for interpretation by a judge (though I think the US copyright office refusing registration may make it de-facto impossible to get it in front of a judge, as I think it's a requirement to actually bring a suit against someone for infringement), but I suspect this is the kind of distinction that they will consider regarding this particular precedent.

jcranmer 3y ago
> (though I think the US copyright office refusing registration may make it de-facto impossible to get it in front of a judge, as I think it's a requirement to actually bring a suit against someone for infringement) The copyright office refusing registration counts as copyright registration for the purposes of register-before-suing-for-infringement.
- zerocrates 3y ago
  You can also just sue the Register of Copyrights for rejecting you after exhausting the internal appeals process, which is what happened here.