Actually, we were forbidden to look at open source code at Microsoft (circa 2009) because it might influence our coding and violate licenses.
In fact, the little precedent that exists over learning from copyrightable code is in favor of it.
More important, the rule urged by Sony would require that a software engineer, faced with two engineering solutions that each require intermediate copying of protected and unprotected material, often follow the least efficient solution (In cases in which the solution that required the fewest number of intermediate copies was also the most efficient, an engineer would pursue it, presumably, without our urging.) This is precisely the kind of “wasted effort that the proscription against the copyright of ideas and facts . . . [is] designed to prevent.” (Sony v. Connectix)
(and I guess courts might, in the future, say the GPL expires when copyrights on the code expire)
Meanwhile open source software has had an immeasurable benefit to society. My computer, tv, phone, light bulb, etc all benefit from OSS—running various licenses, and only a subset using a copyleft like license.
Like, copyright laws are also stifling my innovative business creating BluRays of Disney films and selling them on Amazon.
OpenAI did a dirty job though judging by the cases of the model just reproducing code to the comment, so I can understand why one would criticize this specific project.
https://wiki.winehq.org/Developer_FAQ#Who_can.27t_contribute...
Forbidding people who have seen the "source" programm is most likely to protect their version from going from "matching behaviour" to "behaving like", as in the same code, point. This might also be intended to build a safeguard for good intentioned developers to not break their (most likely existing) own NDAs accidently.
If you're using code and know that it will be output in some form, just stick a license attribution in the autocomplete.
In fact, did you know this is what Apple Books does by default? Say, for example, you copy and paste a code sample from The C Programming Language. 2nd Edition. What comes out? The code you copy and pasted, plus attribution.
If a human programmer reads some else's copyrighted code, OSS or otherwise, memorizes it and later reproduces it verbatim or nearly so, that is copyright infringement. If it wasn't, copyright would be meaningless.
The argument, so far as I understand it, is that Copilot is essentially a compressed copy of some or all of the repositories it was trained on. The idea that Copilot is "learning from" and transforming its training corpus seems, to me, like a fiction that has been created to excuse the copyright infringement. I guess we will have to see how it plays out in court.
As a non-lawyer it seems to me that stable diffusion is also on pretty shaky ground.
APIs are not copyrightable (in the US), so Wine is safe (in the US).
Let me tell you the story of Google Books, also known as "Authors Guild Inc. v. Google Inc"
https://en.wikipedia.org/wiki/Authors_Guild,_Inc._v._Google,....
In 2004, Google added copyrighted books to is Google Books search engine, that does search among millions of book text and shows full page results without any authors authorization. Any sane lawyer of the time would have bet on this being illegal because, well, it most certainly was. And you may be shocked to learn that it is actually not.
in 2005 the Authors Guild sues for this pretty straightforward copyright violation.
Now an important part of the story: IT TOOK 10 YEARS FOR THE JUDGEMENT TO BE DECIDED (8 years + 2 years appeal) during which, well, tech continued its little stroll. Ten year is a lot in the web world, it is even more for ML.
The judgement decided Google use of the books was fair use. Why? Not because of the law, silly. A common error we geeks do is to believe that the law is like code and that it is an invincible argument in court. No, the court was impressed by the array of people who were supporting Google, calling it an invaluable tool to find books, that actually caused many sales to increase, and therefore the harm the laws were trying to prevent was not happening while a lot of good came from it.
Now the second important part of the story: MOST OF THESE USEFUL USES HAPPENED AFTER THE LITIGATION STARTS. That's the kind of crazy world we are living in: the laws are badly designed and badly enforced, so the way to get around them is to disregard them for the greater good, and hope the tribunal won't be competent enough to be fast but not incompetent enough to fail and understand the greater picture.
Rants aside, I doubt training data use will be considered copyright infringement if the courts have a similar mindset than in 2005-2015. Copyright laws were designed to preserve the authors right to profit from copies of their work, not to give them absolute control on every possible use of every copy ever made.
Quite sure the issue at hand is about the code being copied verbatim without the license terms, not "learning" from it.
What is the difference between a neighbor watching you leave your home to visit the local grocery store and mass surveillance? Where do you draw the line?
It is pretty simple, actually.
The reason why those wouldn't apply to Copilot is because they aren't separating out APIs from implementation and just implementing what they need for the goal of compatibility or "programmer convenience". AI takes the whole work and shreds it in a blender in the hopes of creating something new. The hope of the AI community is that the fair use argument is more like Authors Guild v. Google rather than Sony v. Connectix.
You can learn from it, but if you start copying snippets or base your code on it to such an extent that its clear your work is based on it, things start to get risky.
For comparison, people have tried to get around copyright of photos by hiring an illustrator to "draw" the photo, which doesn't work legally. This situation seems similar.
Like if the drawing was meant to be an artistic rendering with independent artistic value, much more likely to be fair use. If the drawing was meant to be a loop-hole to avoid paying the licensing fee on the original, its much less likely. Fair use has a bunch of criteria - a lot of it depends on intention and how the usage would affect the original copyright holder.
I would add that fair use lets you use a copyrighted work, it doesn't make the copyright go away, just adds some cases where you can use the work notwithstanding the original copyright, but the original copyright is still there.
Note: IANAL, this all could be wrong. I dont have any cases, i do know that people propose this sort of thing at wikipedia from time to time - i.e. hiring someone to draw copyrighted photos - and it usually gets shot down as not solving the problem, although im not familiar with the legal basis.
Edit: I guess they do address it in their faq and I'd summarize it "Depends if copyright law applies and depends if it's considered derivative". https://creativecommons.org/faq/#artificial-intelligence-and...
> Tomorrow it will be against stable diffusion or (dall-e, gpt-3 whatever)
> And then eventually against Wine/Proton and emulators (are APIs copyrightable)
Textbook definition of F.U.D.
No it isn't, at least not automatically which is why infringement of licenses exists at all, the fact that you have a brain doesn't change that and never has. If you reproduce someone's code you can be in hot water, and that should be the case for an operator of a machine.
It's also why the concept of a clean room implementation exists at all.
More important, the rule urged by Sony would require that a software engineer, faced with two engineering solutions that each require intermediate copying of protected and unprotected material, often follow the least efficient solution (In cases in which the solution that required the fewest number of intermediate copies was also the most efficient, an engineer would pursue it, presumably, without our urging.) This is precisely the kind of “wasted effort that the proscription against the copyright of ideas and facts . . . [is] designed to prevent.” (Sony v. Connectix)
Remember when Napster was all the rage. And then Jobs and Apple stepped in and set an expectation for the value of a song (at 99 cents)? And that made music into the razor and the iPod the much more profitable blades. Sure it pushed back Napster but artists - as the creator of the goods - have yet to recover.
I'm not saying this is the same thing. It's not. Only noting that today's "win" is tomorrow's loss. This very well could be a case of be careful what you wish for.
Just because both activities are calling "learning" does not mean they are the same thing. They are fundamentally, physically different activities.
The purpose and character of the use, including whether such use is of a commercial nature or is for nonprofit educational purposes
The nature of the copyrighted work
The amount and substantiality of the portion used in relation to the copyrighted work as a whole
The effect of the use upon the potential market for or value of the copyrighted work.
A programmer who studied in school and learned to code did so clearly for and educational purpose. The nature of the work is primarily facts and ideas, while expression and fixation is generally not what the school is focusing on (obviously some copying of style and implementation could occur). The amount and substantiality of the original works is likely to be so minor as to be unrecognized, and the effect of the use upon the potential market when student learn from existing works would be very hard to measure (if it could be detected).
When a machine do this, are we going to give the same answers? Their purpose is explicitly commercial. Machines operate on expression and fixation, and the operators can't extract the idea that a model should have learned in order to explain how a given output is generated. Machines makes no distinction of the amount and substantiality of the original works, with no ability to argue for how they intentionally limited their use of the original work. And finally, GitHub Copilot and other tools like them do not consider the potential market of the infringed work.
API's are generally covered by the interoperability exception. I am unsure how that is related copilot or dall-e (and the likes). In the Oracle v. Google case the court also found that the API in question was neither an expression or fixation of an idea. A co-pilot that only generated header code could in theory be more likely to fall within fair use, but then the scope of the project would be tiny compared to what exist now.
My (extremely amateur) understanding is that what is meant by "learn from it" is one of the hinge points of the legal question.
If a programmer reads licensed code and reproduces it verbatim or near-verbatim in a project with a conflicting license, that becomes a legal problem in certain circumstances.
If a programmer reads the same code and gets an idea to implement something different, that's less troublesome (or at least, if it is troublesome it's in a different area; if the idea was related to a patentable process, then other questions arise, but I'm even less qualified to speak to that area of law).
There's nothing special about copy/paste buttons that make them the only way you can infringe copyright.
Fair use doesn't automatically kick in just because someone uses what they took/copied as part of a larger artifact; it's a really complicated legal line.
A programmer can read available but not oss licensed code and learn from it. Thats fair use. If a machine does it, is it wrong ? What is the line between copying and machine learning ? Where does overfitting come in ?
Today they're filing a lawsuit against copilot.
Tomorrow it will be against stable diffusion or (dall-e, gpt-3 whatever)
And then eventually against Wine/Proton and emulators (are APIs copyrightable)