I'm kinda sceptical that this goes anywhere given that basically they say that whatever copilot outputs is your responsibility to vet that it doesn't break any copyright (obviously that goes against the promise of it and the PR but that's the small print that gets them out of trouble).
Well, if the trackers also hosted mixed-up blocks of data for all the torrents they tracked and their protection was "LOL make sure you don't accidentally download any of these tiny data blocks in the correct order to reconstruct the copyrighted material they may be parts of wink"
I think you're looking for consistency that the legal system just doesn't provide. The music industry is more organised and litigious than the software industry and that gives them power that you and I don't have. If you called it "Napster 2.0" specifically you'd probably be prevented from shipping by a preliminary injunction. Is that fair or consistent? No. But it's the world we live in. Programmers want laws to be irrefutable and executable logic but they just aren't.
So if you produce napster 2.0 to be the best music piracy tool, and you test it for piracy, and you promote it for piracy... you're going to have trouble.
If you produce napster 2.0 as a general purpose file sharing system, let's call it a torrent client, and you can claim no ill intent... you may have trouble but it's a lot more defensible in court.
I would find it a big stretch to say Github's intent here is to illegally distribute copyrighted code. No judgment on whether the class action has any merit, just saying I would be very surprised if discovery turns up lots of emails where Github execs are saying "this is great, it'll let people steal code."
> I would find it a big stretch to say Github's intent here is to illegally distribute copyrighted code.
Almost everything on GitHub is subject to copyright, except for some very old works (maybe something written by Ada Lovelace?), and US government works not eligible for copyright.
Now, many of the works there are also licensed under permissive licenses, but that is only a defense to copyright infringement if the terms of those licenses are being adequately fulfilled.
> Almost everything on GitHub is subject to copyright,
Agreed. Like I said, it's about intent. Can anyone say with a straight face that copilot is an elaborate scheme to profit by duplicating copyrighted work?
I don't think the defense is that it wasn't trained on copyrighted data. It obviously was.
I think the defense is that anything, including a person, that learns from a large corpus of copyrighted data will sometimes produce verbatim snippets that reflect their training data.
So when it comes to copyright infringement, are we moving the goalposts to where merely learning from copyrighted material is already infringement? I'm not sure I want to go there.
Now, IANAL, but iirc, that is all 100% okay and legal. In fact, I can even download copyrighted music and movies without issue. So, I don't even need to make sure I don't download anything under copyright.
The issue isn't downloading copyrighted stuff.
Rather, it's making available and letting others download it. That was where you got in trouble.
Knowingly downloading copyrighted material, say to get it for free, still violates the rights of the copyright holders. It's just that litigating against members of the public is bad PR and not exactly lucrative, especially when it's likely that kids downloaded the content.
People used to get busted from buying bootleg VHS and DVDs on the street before P2P filesharing was a common thing. Then, early on, people were sued for downloading copyrighted files before rightsholders decided to take a different legal strategy to go after sharers and bootleggers.
This is a bad analogy because P2P networks exist that are legal to operate, because Section 230 of the CDA prevents interactive computer services from being held responsible for user generated content.
What made Napster illegal is that the company did not create their network for fair use of content, but to explicitly violate copyright for profit.
Copilot is like Napster in this case, in that both services launder copyrighted data and distributed it to users for profit.
Copilot is not like other P2P networks that exist to share data that is either free to distribute or can be used under the fair use doctrine. Copilot explicitly takes copyrighted content and distributes it to users in violation of licenses, that's its explicit purpose.
It's entirely possible to make a Copilot-like product that was trained on data that doesn't have restrictive licensing in the same way it's entirely possible to create a P2P network for sharing files that you have the right to share legally.
Saying "it's your responsibility to not breach licenses or violate copyright" doesn't absolve your service from breaching licenses and violating copyright itself.
Yet we all use web browsers that copy copyrighted text from buffer to buffer all the time. This doesn't even include all of the copying that ISPs perform.
It might be fair to say that the read performed in training has the same character since no human is involved.
The real copyright violation would be using a derived work.
the actual copying isn't a problem, it's distribution. if i buy access to a PDF i'm not going to get in trouble for duplicating the file unless i send it to someone else.
when someone uploads their copyrighted text to a web page they are distributing it to whoever visits that page. the browser is just the medium.
A browser isn't a amalgamation of billions of pieces of other works. A browser executes and renders code it's served.
Copilot's corpus is quite literally tomes of copyrighted work that are encoded and compressed in its neural network, from which it launders that work to create similar works. Copilot itself, the neutral network, is that corpus of encoded and compressed information, you can't separate the two. Copilot stores and distributes that work without any input from rightsholders, and it does it for profit.
A better analogy would be between a browser and a file server filled with copyrighted movies whose operator charges $10/mo for access. The browser is just a browser in this analogy, where the file server is the corpus that forms Copilot itself.
You could argue that it’s the individual projects using copilot that are violating here, I guess? Like you can use curl or git to dump some AGPL code into your commercial closed software but no one would (hopefully) blame those tools.
So copilot is fine but anyone using it must abide by the collective set of licenses that it used to write code for you…?
If a license requires attribution, and you reproduce the code without attribution using your editor plugin, it seems to me the infringement is on the editor plugin.
Note that even licenses like MIT ostensibly require attribution.
How many ways are there to write many of the basic algorithms we all use though? Can I copyright "({ item }) => <li>{item.label}</li>"?
Because I sure have seen that exact code written, from scratch, in many many places.
I guess my question boils down to "What is the smallest copyrightable unit of code?". Because I'm certain suing a novelist for copyright infringement on a character that says "Hi, how are you?" would be considered absurd.
Apple Books slaps an attribution notice on the end if you copy 4 or more words from a book. The Verve got sued by The Rolling Stones for a 4 second sample on 'Bittersweet Symphony'. Post 'Blurred Lines' you can now be sued for copying "the feel" of a song.
Really what it comes down to is do you have enough resources to convince a judge or jury that X is a copy of Y? Doesn't really matter the size of X.
There were well known examples of copilot reproducing exact code snippets well before this lawsuit (e.g. the Quake's fast inverse square root function). Microsoft dealt with them by simply adding the offending function names to a blocklist.
In other words, if your open source project doesn't have such immediately recognizable code and didn't cause a shitstorm on Twitter, chances are copilot is still happily spewing out your exact code, sans the copyright and license info.
Yeah I guess so. This website reads like bullshit bingo from some weird twitter dude trying to sell you his newest product:
"AI needs to be fair & ethical for everyone. If it’s not, then it can never achieve its vaunted aims of elevating humanity. It will just become another way for the privileged few to profit from the work of the many."
Blah blah. Can we get back to the hacking on stuff mentality?
Hah, funny. I've used Pollen before and think I've had contact with him a few years ago! The blah blah about AI elevating the world is still bs imho. I still disagree with his views (https://matthewbutterick.com/chron/this-copilot-is-stupid-an...) and this law suit.
I wasn't actually talking about him specifically btw when saying "this sounds like a crypto bro from twitter". The overly enthusiastic AI talk reminded me of that, that's what I wanted to say.
If you'd ever read even a single one of the licenses to the software I'm sure you use everyday, you'd understand.
This is such an obvious and pathetic strawman.
I notice often on hackernews that people don't seem to understand anything about free or open-source software outside of the pragmatics of whether they can abuse the work for free.
You read a lot into my not so serious comment. Maybe internet comment sections aren't the right place for you.
But I'll bite: I know licensing, thank you. But what's copyrightable is not so easy. Licenses are not so easy. Copilot does not copy entire works and it's very questionable if a few lines of code are "piracy". It's a repeating discussion again and again, there's nothing novel about it except for the fact that a machine learns (and overfits for small portions of code). So please get off your high horse. I don't care for your fundamentalism.
If you know this area of IP law then you know that LOTs of licenses, copyleft or not, require attribution (which copilot never does, and can’t do by its construction), and you know that what’s problematic is when the model output is arguably not “fair use”. Examples of that abound.
You don’t need any fundamentalism to know that copilot’s output carries huge and untested legal risk. If this lawsuit clears some of this up, that’s a big win for everyone.
Yes, you're right. But my point also was that it's not so easy when it's just a few lines, isn't it? Especially since this is an international issue and the definitions of copyrightable work is not easily definable.
> You don’t need any fundamentalism to know that copilot’s output carries huge and untested legal risk. If this lawsuit clears some of this up, that’s a big win for everyone.
I agree with that! I also see this as the only proper takeaway that I think is ok. The rest is making money off this thing. But the US has a different law suit culture anyway, which I find weird.
As a non-lawyer, I am very suspicious of the claim that "Plaintiffs and the Class have suffered monetary damages as a result of Defendants’ conduct." Flagrant disregard for copyright? Sure, maybe. The output of the model is subject to copyright? Who knows! But the copyright holders being damaged in some what? Seems doubtful. The best argument I could think of would be "GitHub would have had to pay us for this, and they didn't pay us, so we lost money," but that'd presumably work out to pennies per person.
I dont think that's comparable. For starters, its not just the length of a quote that makes it fair use, but the way quotes are used i.e. to engage in commentary.
If your intent is to create a competing product for profit, chances are that won't be found as fair use, given that determining fair use depends on intent and how the content is used.
Using clips from a movie in a movie review is probably fair use.
Using clips from a movie in knock-off of that movie for profit? Probably not fair use if it's not a parody.
Copilot is not like a movie reviewer using clips to review a movie. Copilot is like a production team for a movie taking clips from another movie to make a ripoff of that movie and selling it.
Consider every repo on github to be a movie. Copilot is taking individual frames out of every movie on github and composting them into a new film.
I think most of us would agree that individually, each frame is copyrighted. But what if you take one frame from a million different movies and put them in an order that produces a new coherent movie?
The core question we need to settle in court is: does the new movie become its own copyrightable work, or is it plagiarism?
> Copilot is like a production team for a movie taking clips from another movie to make a ripoff of that movie and selling it.
I can't think of a 5 line snippet I've written or read that makes sense to claim ownership of. They don't stand on their own in the way even a 30s movie clip does.
You're mistaking the end-user's copyright infringement with Copilot's alleged infringement.
Copilot is fair use and transformative -- that is unless there is an open source Copilot that Copilot is training on, only then would it be competing and it's easy for GitHub or OpenAI to exclude those repos of copilot alternatives from the training set.
It's the license that matters, not whether the code is visible on Microsoft's website.
Code which anybody can view is called "source available". You aren't necessarily allowed to use the code, but some companies will let their customers see what is going on so they can better integrate the code, understand performance implications, debug and fix unexpected issues, etc. The customers would probably face significant legal risks if they took that code and started to sell it.
"Open source" code implies permission to re-use the code, but there is still some nuance. Some open-source licenses come with almost no restrictions, but others include limiting clauses. The GPL, for example, is "viral": anybody who uses GPL code in a project must also provide that project's source code on request.
What do you think the chances are that Microsoft would surrender the Copilot codebase upon receipt of a GPL request?
Those damages are enumerated on pages 50-52. Remember, "damages" is being used in a legal sense here -- for a non-lawyer, you can interpret it more like "a dollar value on something someone did that was wrong". This is more broad than the colloquial use of the word.
The common practice in copyright cases is to calculate damages based on the theoretical cost that the infringer would have paid if they have bought the rights in the first place. This method was used during the piratebay case to calculate damages caused by the sites founders.
They did not actually calculate damages in terms of lost movie tickets or estimates vs actually sales number of sold game copies. When it came to pre-releases where such product wouldn't have been sold legally in the first place, they simply added a multiplier to indicate that the copyright owner wouldn't have been willing to sell.
For software code, an other practice I have read is to use the man-hours that rewriting copyrighted code would cost. Using such calculations they would likely estimate the man hours based on number of lines of code and multiply that with the average salary of a programmer.
So for iseven can we go for how much a student might accept 20 an hour say and multiply that by the one minute required to create it and offer them 33 cents?
"Using such calculations they would likely estimate the man hours based on number of lines of code and multiply that with the average salary of a programmer."
The average salary of a programmer in which country?
So much programming is outsourced these days, and in some places programmers are very cheap.
This is just my guess, but I think the intention from the judges is not to actually calculate a true number. The reason they used the cost of publishing fees in the piratebay case was likely to illustrate how the court distinguished between a legal publisher vs an illegal one. The legal publisher would have bought the publishing rights, and since piratebay did not do this, the court uses those publishing fees in order to illustrate the difference.
If the court wanted to distinguish between Microsoft using their own programmers to generate code vs taking code from github users, then the salary in question would likely be that of Microsoft programmers. It would then be used to illustrate how a legal training data would look like compared to an illegal one.
The one thing we can say with complete certainty is that most programmers who had their code used without permission will not receive very much money at all if this class action lawsuit is decided in their favor.
Money likely isn't the main goal (maybe it is for the lawyers), these are open source repos. Maybe they didn't consent to have their code used as training and that seems like the kind of thing consent should be needed for. Maybe this the AI spitting out copied snippets is a violation of open source licensing without attribution.
I don't care about the money. I support this because it will establish case law that other companies can't ignore licenses as long as they throw AI somewhere in the chain.
If "I took your code and trained an AI that then generated your code" is a legal defense, the GPL and similar licenses all become moot.
"This license does not grant GitHub the right to sell Your Content. It also does not grant GitHub the right to otherwise distribute or use Your Content"
Aren't there statutory damages for copyright infringement, i.e. there is a presumption that each work infringed is worth at least a certain amount without proving actual damages?
I suspect this will be the first of many lawsuits over training data sets. Just because it is obscured by artificial neural networks doesn't mean it's an original work that is not subject to copyright restrictions.
I don't know why we're treating it as anything less than a human brain. A human can replicate a painting from memory or a picture of mickey mouse and that would likely be copyright infringement, but they could also take a drawing of Mickey Mouse sitting on the beach and given him a bloody knife & some sunglasses and it'd likely be fair use of the original art.
The AI can copy things if it wants, but it can also modify things to the point of being fair use, and it can even create new works with so little of any particular work that it's effectively creativity on the same level of humans when they draw something that popped into their heads.
Wow, this is interesting iteration in the ongoing divide between "East Coast code" vs. "West Coast code" as defined by Larry Lessig. For background, see https://lwn.net/Articles/588055/
I wonder if the plaintiffs' code would stand up to scrutiny of whether any of it was copied, even unintentionally, from other code they saw in their years of learning to program? I know that I have more-or-less transcribed from Stack Overflow/etc, and I have a strong suspicion that I have probably produced code identical to snippets I've seen in the past.
Ask HN: I want to modify the BSD 2-Clause Open Source License to explicitly prohibit the use of the licensed software in training systems like Microsoft's Copilot (and use during inference). How should the third clause be worded?
The No-AI 3-Clause Open Source Software License
Copyright (C) <YEAR> <COPYRIGHT HOLDER>
All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions
are met:
1. Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.
2. Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in
the documentation and/or other materials provided with the
distribution.
3. Use in source or binary forms for the construction or operation
of predictive software generation systems is prohibited.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
This does seem like a pretty compelling rebuttal, since the preceding comment suggests that GPL does nothing to Microsoft's ability to incorporate code into Copilot's model.
How would you ever prove the parameters of a model were generated by specific training data? Couldn't multiple sets of training data produce the same embeddings/parameters? I imagine there could be infinite possible sets of training data that would lead to the same results, depending on the type of predictive software.
Is it? A similarly casual clause in the OCB license prevented OCB from being used by the military for many years (granted, it prevented OCB from being used almost everywhere else, too).
I have no idea if this license language works or doesn't, but this is hardly the least productive subthread on this story. It's concrete and specific, and we can learn stuff from it.
IANAL, and I'm no fan of copilot, but I wonder if this kind of clause (your #3) is going to fly: you're preemptively prohibiting certain kinds of reading of the code (when code is read by the ML model in training). But is that something a license can actually do?
The legal footing that copyright gives you, on which licensing rests, certainly empowers you to limit things about how others may redistribute your work (and things derived from it), but does it empower you to limit how others may read your work? As a ridiculous example, I don't think it would be enforceable to have a license say "this code can't be used by left-handed people", since that's not what copyright is about, right?
Licenses get to set terms of redistribution. But training of the ML model -- the thing described by your #3 -- is not redistribution (imho). So maybe it's as unenforceable as saying left-handed people can't read your code.
The redistribution happens later, either when copilot blurps out some of your code, or when the copilot user then distributes something using that code (I'm curious which). At that point, whether some use of your code is infringing your license doesn't depend on the path the code took, does it? (in which case #3 is moot)
Okay; thanks for clarifying. I actually hadn't noticed that use of "use" in the BSD license before. I think I need an IP lawyer explain what that "use" means.
The legal theory for copilot is that training an ML model is fair use, not that the license allows it. If it is fair use then you can't prohibit it by license, no matter what you put in your license.
For this clause to have any positive effect, you need to 1) be willing to pursue legal action against violators and 2) actually notice that the clause has been violated.
Such language must be carefully written. What is the definition of “construction” and “operation” in a legal context? What is a “predictive software generation system”? That’s a very specific use case, you sure you covered everything you want to prohibit?
You’ve inserted your clause in such a way that this dependency cannot be used in any way to build anything similar to a “predictive software generation system”, even with attribution, as it would fail clause 3.
You have to consider that novel licenses make it difficult for any party that respects licenses to use your code. It is difficult to make one-off exceptions, especially when the text is not legally sound. So adoption of your project will be harmed.
So if you are serious about this license, you need a lawyer.
var G = 6.67e-11;
var force = G * mass1 * mass2 / distance * distance;
return force;
Second prompt:
% pointMatrixGravity - Calculates the gravitational force and torque on a
% point mass due to a matrix of point masses.
%
% [force, torque]=pointMatrixGravity(array1,array2)
%
% Inputs:
% array1 - 3xN array of point masses
% array2 - 3xN array of point masses
% Outputs:
% force - 3x1 array of gravitational force
% torque - 3x1 array of gravitational torque
%
% Author: David J. R. MacKay,
% email:
I got something a bit different from the other person who responded to you.
function force=Gmmr2Array(mass1, mass2) {
var G = 6.67408e-11;
var m1 = mass1;
var m2 = mass2;
var r = 1;
var force = G \* m1 \* m2 / (r \* r);
return force;
}
function [force, torque]=pointMatrixGravity(array1,array2) {
var G = 6.67408e-11;
var m1 = array1(1);
var m2 = array2(1);
var r = norm(array1(2:4)-array2(2:4));
var force = G \* m1 \* m2 / (r \* r);
var torque = cross(array1(2:4)-array2(2:4), force);
return [force, torque];
}
You may convey a work based on the Program, or the modifications to produce it from the Program, in the form of source code under the terms of section 4, provided that you also meet all of these conditions:
a) The work must carry prominent notices stating that you modified it, and giving a relevant date.
b) The work must carry prominent notices stating that it is released under this License and any conditions added under section 7. This requirement modifies the requirement in section 4 to “keep intact all notices”.
c) You must license the entire work, as a whole, under this License to anyone who comes into possession of a copy. This License will therefore apply, along with any applicable section 7 additional terms, to the whole of the work, and all its parts, regardless of how they are packaged. This License gives no permission to license the work in any other way, but it does not invalidate such permission if you have separately received it.
——
I don’t see how one could argue that training on GPL code is not “based on” GPL code.
A developer is a person. Copilot is software based on the GPL code. Just because you use the word “learn” does not make what a human does and software does the same thing.
I am not against this lawsuit but I'm against the implications of this because it can lead to disastrous laws.
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)
GitHub Copilot has been proven to use code without license attribution. This doesn't need to be as controversial as it is today.
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.
Maybe its time for Creative Commons License to address this. I'm curious if No-Derivative would already prohibit this? Does the ND language need tweaking? Or do they need a whole new clause.
Not for GitHub -- users who upload their code accept GitHub's license agreements which allows it to use it in many different ways, including Copilot. Kind of how when you create a Robinhood account you agree to arbitration and can't sue them.
Sure, but given the timetable for changing the law, it still seems pretty reasonable to apply the same standard to Microsoft (and by extension Github) in the meantime
Copyright laws aren't preventing you from learning cinematography by watching said Disney movies though, and using all their techniques for your own project.
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.
That sucks for little snippets of software though, doesn’t it? It’s like copyrighting individual dance moves (not allowed under the current system) and forcing dancers to never watch each other to make sure they’re never stealing.
I mean, it's not like the copyrights are keeping you from doing things. It's stopping you from looking at someone else's source. And it's not like source is easy to accidentally see like dance moves are.
I don’t quite agree. Msft took a conservative approach to copyright to protect their own business.
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.
The fact that the laws are inconsistent and expensive to defend against leads companies like Microsoft to take this conservative approach that slows down progress.
It demonstrates that it stifles copying. That may make it easier for the copier to innovate, but doesn’t dispute the main argument for having copyright protection: that, without the protection of copyright, the code wouldn’t have been written.
Most of it? I would think >50% of open source code writers find it necessary to restrict the rights to copy and use their code. In a world without copyright protections, would the GPL be legal?
(and I guess courts might, in the future, say the GPL expires when copyrights on the code expire)
That was out of abundance of caution, not based on any legal precedent.
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)
Way, way back in 1992, Unix Systems Laboratories sued BSDI for copyright infringement. Among other things, they claimed that since the BSD folks had seen the Unix source code, they were "mentally contaminated" and their code would be a copyright violation. This led to the BSD folks wearing "mentally contaminated" buttons for a while.
>A programmer can read available but not oss licensed code and learn from it. Thats fair use.
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.
I think the commenter you replied to was talking about using the functional, non-copyrightable elements of the copyrighted code. Clean-room is not even required by case law. There's precedent that explicitly calls it out as inefficient.
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)
Well they are a special case here however since they don't solve a specific problem nor build a programm per se but instead (re)build a programm after existing specs. Their explicit goal is to match the behaviour of another piece of software with a translation layer.
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.
Slippery slope? Are you familiar with judicial precedent? Being bound to precedents is central to common law legal systems, so I don't think the GP's take was so outlandish. "Slippery slopes" and "whataboutism" might be thought-terminating buzzwords online, but not in front of a judge.
> 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 ?
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.
It might or might not be depending on the situation. Some of it might come down to intent.
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.
Wine/Proton are safe because there is controlling 9th and SCOTUS precedent in favor of reimplementation of APIs.
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.
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.
> If a machine does it, is it wrong ? What is the line between copying and machine learning ?
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?
> A programmer can read available but not oss licensed code and learn from it. Thats fair use.
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).
Agreed. But it could go the other way as well. Let's say MS / HB wins and the decision establishes and even less healthy / profitable (?) outcome over the long term.
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.
It would be good to have a definitive and simple line for fair use that could be applied to all forms of copyright. Right now fair use is defined by four guidelines:
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.
> 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 ?
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.
TabNine has absolutely improved my life as a programmer. There's something really rewarding about having a robot read your mind for entire blocks of code.
It's not just functions either, one of the most common things that it helps me with daily is simple stuff like this:
Typing
const x = {
a: 'one',
b: 'two',
...
}
And later I'll be typing
y = [
a['one'],
b[' <-- it auto-completes the rest here
]
It's really amazing the amount of busy-work typing in programming that a smart pattern matching algo could help with.
I don't think this is a good example of the value of these things. You can just as easily do that same thing with advanced text editor features. Sublime for example supports multi-cursor editing. Just hold alt+shift+arrow keys to add a cursor, then type in the brackets you want. Ctrl+D can be used to select the next occurrence of the current selection with multiple cursors, built-in commands from the command pallete can do anything to your current selection (e.g. convert case), etc.
All of that efficiency without having to pay a monthly subscription, wasting electricity on some AI model, and worrying about the legal/moral implications.
Why? You can copy and paste the entire section, and use multiple cursors to add in the brackets.
going from
a: 'one',
to
a['one'],
just requires you to add two brackets and remove the colon. With multiple cursors you can do that exact same operation for all lines in a few keystrokes.
It's having to go find the other block you want, copy and paste it, and then set up the multiple cursors and type, versus it just happening automatically without any of that.
In my use cases I’ve long moved on from writing the original hash. Having it autocomplete without having to open a file and tab back/forth (or find then copy/paste a block to the other file to temporary work on it) etc.
But what's lost in my over simplified example is the contetxt is usually way more involved. I'm usually passing those as arguments to some function or other unique syntax situation that a glorified find and replace can solve. It's all about doing it in the times you would never think even bother writing a custom command because typing is faster given the unique syntactical context... The only thing faster then is autocomplete.
I'm not actually recreating a new hash with the convienient same format.
That autocomplete was sort of ok in tabnine, but Copilot completely blows it out of the water. Resource consumption for Copilot is also much more restrained.
Which reminds me I have to cancel my tabnine subscription. Been paying them for a year without using it.
I haven't tried Copilot personally but thanks I'll try it. I did try TabNine over a year ago and found it's improved dramatically since then so maybe it's gotten better within training.
Does anyone have a problem with it, so long as the material it trained on was with explicit permission/license and not potentially in violation of copyright?
That's where the line is for it to be suspect IMO.
I guess I'm just afraid that it might not be as good as it is that way.
It's a bit like how GPT-3, Stable Diffusion and all those generative models use extensive amounts of copyrighted material in training to get as good as they do.
In those cases however the output space is so vast that plagiarism is very unlikely.
The interesting thing is that the names get explicitly attached to these styles. It isn't exactly a copyright issue, but I'm sure it will get litigated regardless.
I think the prompt "GPT-3, tell me what the lyrics for the song Stan by Eminem is" is very likely to output copyrighted material. The same copyrighted material is, of course, already republished without permission on google.com.
there are literally thousands of years of artwork that fall under public domain, the idea that the dataset isn't big enough to make good images without copyright infringement and attribution laundering is frankly laughable.
My guess is that is not as much about the amount of available data but how accessible it is. Scraping the internet seems to be one of the preferred ways of gathering vast amounts of, in particular, text and images.
Telling apart what's public domain or not is not a trivially automatable task.
If one just relies on curated libraries of vetted public domain content you don't get, by far, the expected amout of variability and diversity.
It being permissively licensed is virtually irrelevant because only a minority of code is so permissively licensed you can just do what you like under any license. Far more is do what you like within the scope of the license. For example GPL do with it what you like so long as any derivative work is also GPL.
This is what I hope comes out of the lawsuit. If a company wants to sell an AI model, they need to own all of the training data. It can't be "fair use" to take other peoples' works at zero cost, and use it to build a commercial product without compensation.
And maybe models trained on public data should be in the public domain, so that AI research can happen without requiring massive investments to obtain the training data.
Ok you got me, that wording was lazy on my part. But that's a really bad take on yours:
> It was trained on OSS which is explicitly licensed for free use.
That's not what the lawsuit is about. It's not about money, it's about licensing. OSS licenses have specific requirements and restrictions for using them, and Copilot explicitly ignores those requirements, thus violating the license agreement.
The GPL, for example, requires you to release your own source code if you use it in a publicly-released product. If you don't do that, you're committing copyright infringement, since you're copying someone's work without permission.
Yeah, and I think that's fair re: licensing. Curious to see how it pans out.
Also, re: your edit, not quite. They require you to release modified source under certain conditions if you make modifications to it. If everybody had to release code using GPL to the world, every companies code would currently be released to the world. There's more nuance than that. The gnu site covers a lot of that nuance (https://www.gnu.org/licenses/gpl-faq.en.html#UnreleasedMods)
LGPL is the one that enterprises won't touch with a 10 foot pole, due to more restrictive licensing, and more conditions under which you'd have to open source your own code.
Most companies building commercial products on top of FOSS are obeying the license requirements. (I have been through due diligence reviews where we had to demonstrate that, for each library/tool/package.)
The same cannot be said for Copilot: there have been prior examples here on HN showing that it can emit large chunks of copyrighted code (without the license).
There has to be a reasonable context here. Even if it’s trained on proprietary code it rarely ever is inserting that code directly in a way that is at all relevant to how it was used in the past.
Obvious licensing needs to be respected and it shouldn’t be hard to solve that problem. But 99.9% of code isn’t some unique algorithm, it’s gluing libraries and setting up basic structures.
Most of the examples I’ve seen done line up with the reality of code completion tools. Code is rarely valuable when broken up into its small parts.
Even copying a full codebase is rarely enough to draw value from… there’s way more to a software business than the raw code. But that’s a different problem.
why not just train it on your own code or an opt-in data base of voluntarily contributed code? why does everyone else have to make your life easier [and generate enormous wealth for a third party with zero compensation for their work] involuntarily?
I feel like Charlie Gordon from Algernon with and without Copilot.
Literally 10x faster development.
Case in point: had an unexpected project and no time to complete it.
Within an hour Copilot helped me:
* Write a couple of tricky matplotlib plots
* Do some extensive analysis with Pandas
* Write a couple of SQL queries
* Write a Flask back-end and deploy it
* Write a bit of a front-end
* This all with extra comments , links to documentation and pretty reasonable style
I have experience with all of the above mentioned but the speed increase was considerable.
This would a a good day's work without Copilot and there would be less commenting and hackier code.
Before Copilot I would be cursing a lot more reading various docs...
The key thing that Copilot does it reduces latency for your thoughts-action-results loop.
Does the open source really suffer if less people read documentation directly? Would you really be less likely to create an open source library if you knew someone can now use your library at 10x speed?
The inference ability has crossed uncanny valley so many times.
I find myself wondering whether there is a speech recognition component at times.
When teaching a lecture I will start saying something and write a prompt at the same time and the sentence produced by Copilot will be spot on what I've just said.
Ideally there would an open source version of Copilot that respects everyone's wishes. I fear that is impossible.
If you had legal expertise and a strong opinion on the matter I suppose you could write a persuasive brief for the consideration of the court. If you have a strong opinion but aren't a legal eagle you could write to your legislators in support of legislation explicitly supporting this use case or organize the support of people more capable in that arena.
If you are opinionated but lazy, no judgement here as I sit here watching TV, you could add a notation at the top of your repos explicitly supporting the usage of your code in such tools as fair use.
Notably if your code is derivative of other works you have no power to grant permission for such use for code you don't own so best include some weasel words to that effect. Say.
I SUPPORT AND EXPLICITLY GRANT PERMISSION FOR THE USAGE OF THE BELOW CODE TO TRAIN ML SYSTEMS TO PRODUCE USEFUL HIGH QUALITY AUTOCOMPLETE FOR THE BETTERMENT AND UTILITY OF MY FELLOW PROGRAMMERS TO THE EXTENT ALLOWABLE BY LICENSE AND LAW. NOTHING ABOUT THIS GRANT SHALL BE CONSTRUED TO GRANT PERMISSION TO ANY CODE I DO NOT OWN THE RIGHTS TO NOR ENCOURAGE ANY INFRINGING USE OF SAID CODE.
Years from now when such cases are being heard and appealed ad nauseam a large portion of repos bearing such notices may persuade a judge that such use is a desired and normal use.
You could even make a GPLesque modification if you were so included where you said. SO LONG AS THE RESULTING TOOLING AND DATA IS MADE AVAILABLE TO ALL
Note not only am I not your lawyer, I am not a lawyer of any sort so if you think you'll end up in court best buy the time of an actual lawyer instead of a smart ass from the internet.
This. I've seen so many class action lass suits where at the end of the day the highest gain per Capita always ends up going to the lawyers. Fuck this guy and everyone trying to make money from this.
I have quite a bit of respect for Matthew Butterick. I don't think he's just a lawyer looking to earn a quick buck. He cares about software and wants to make the world a better place.
> But neither Matthew Butterick nor anyone at the Joseph Saveri Law Firm is your lawyer
This is curious. None of them are my lawyers, but surely at least some of them are someone's lawyers? Isn't it wrong to put such a blanket disclaimer on a website which might well be read by their clients?
On page 18, they show Copilot produces the following code:
>function isEven(n) {
> return n % 2 === 0;
>}
They then say, "Copilot’s Output, like Codex’s, is derived from existing code. Namely, sample code that appears in the online book Mastering JS, written by Valeri Karpov."
Surely everyone reading this has written that code verbatim at some point in their lives. How can they assert that this code is derived specifically from Mastering JS, or that Karpov has any copyright to that code?
There is no way in hell that isEven is covered by copyright.
"In computer programs, concerns for efficiency may limit the possible ways to achieve a particular function, making a particular expression necessary to achieving the idea. In this case, the expression is not protected by copyright."
True, but trademark and copyright are pretty different on a fundamental level both in the purpose behind the law and how its implemented. I suspect if it weren't for the term "intellectual property" tying trademark, copyright and patents together, we wouldn't really think of them in such a unified way since they are all really different from each other.
> There is no way in hell that isEven is covered by copyright.
Hey, I said the same thing about APIs, but here we are.
Edit: Actually, the Supreme Court declined ruling whether APIs are copyrightable, but they did say that if they are, reusing them like google reused the java apis in android would fall under fair use. Given that lower courts did think that APIs should be copyrightable, we don't know if they are anymore.
What's interesting in that case is I would argue the interface code is MORE important than the implementing code. You can hire any SWE to re-implement the entire API and it being pretty straightforward. The interface code was what actually mattered with developers and took creative expression, design, sequence, organization to put together and feedback and iteration over the years. Google knew this too, taking the interface code was WAY more important for them than if they had done the reverse.
They determined the other `isEven()` function was cribbed from Eloquent Javascript because of matching comments. I wonder if the complaint just left off telltale comments from that Mastering JS one?
Yeah, the other one I found much more persuasive. The extra comments were unequivocally reproduced from the claimed source. (although that output was from Codex, rather than Copilot).
Yep, same. Not in JS, but in Haskell, for the Even Fib project Euler problem. Something like a million people have submitted right answers for that problem and assuming half wrote their own filter rather than importing a isEven library then that's half a million people there.
They changed it, I'm 100 % sure. The profile picture was Saul from Breaking Bad. I assume they read the comments here and changed it in a matter of one or two minutes.
That seems like a really bad choice of an example for this, but as I haven't read the document I don't have any other context beyond what you've posted here, I have to take your word for it that that's the purpose of this snippet.
However, if you are looking to understand the reasoning behind this lawsuit, there are lots of better examples online where Copilot blatantly ripped off open source code.
It's possible the complaint is using a trivial example to illustrate the type of argument plaintiffs want to make during any trial. A 200-line example is too unwieldy for non-programmers to digest, especially given the formatting constraints of a legal brief.
Look at paragraphs 90 and 91 on page 27 of the complaint[1]:
"90. GitHub concedes that in ordinary use, Copilot will reproduce passages of code verbatim: “Our latest internal research shows that about 1% of the time, a suggestion [Output] may contain some code snippets longer than ~150 characters that matches” code from the training data. This standard is more limited than is necessary for copyright infringement. But even using GitHub’s own metric and the most conservative possible criteria, Copilot has violated the DMCA at least tens of thousands of times."
Does distributing licensed code without attribution on a mass scale count as fair use?
If Copilot is inadvertently providing a programmer with copyrighted code, is that programmer and/or their employer responsible for copyright infringement?
There's a lot of interesting legal complications I think the courts will want to adjudicate.
It seems obvious that AI models are derivative works of the works they are trained on but it also seems obvious that it is totally legally untested whether they are derivative works in the formal legal sense of copyright law. So it should be a good case assuming we have wise and enlightened judges who understand all nuances and can guide us into the future.
copilot is great, and ignorance is bliss, isn't it
The situation that this lawsuit is trying to save you from is this: (1) copilot blurps out some code X that you use, and then redistribute in some form (monetized or not); (2) it turns out company C owns copyright on something Y that copilot was trained on, and then (3) C makes a strong case that X is part of Y, and that your use of X does not fall under "fair use", i.e. you infringed on the licensing terms that C set for Y.
You are now in legal trouble, and copilot put you there, because it never warned that you X is part of Y, and that Y comes with such and such licensing terms.
Whether we like copilot or not, we should be grateful that this case is seeking to clarify some things are currently legally untested. Microsoft's assertions may muddy the waters, but that doesn't make law.
The title of the submitted PDF document: "Microsoft Word - 2022-11-02 Copilot Complaint (near final)"[0]
I've noticed this a lot and it's quite funny seeing what the actual filename of the document was. Does this just get included as metadata by default when you export to PDF?
If a software developer learns how to code better by reading GPL software and then later uses the skills they developed to build closed source for profit software should they be sued?
If a software developer writes a program to remember a million lines of GPL code, then uses that dataset to "generate" some of that code, then they are essentially violating that license with extra steps.
The extra steps aren't enough to exhonorate them. It's just a convoluted copy operation.
Is just like how a lossy encoding of a song is still - with respect to copyright - a copy of that song. The data is totally different, and some of the original is missing. It's still a derivative work. So is a remix. So is a reperformance.
823 comments
[ 2.6 ms ] story [ 310 ms ] thread> behalf of a proposed class of possibly millions of GitHub users...
The appendix includes the 11 licenses that the plaintiffs say GitHub Copilot violates: https://githubcopilotlitigation.com/pdf/1-1-github_complaint...
So if you produce napster 2.0 to be the best music piracy tool, and you test it for piracy, and you promote it for piracy... you're going to have trouble.
If you produce napster 2.0 as a general purpose file sharing system, let's call it a torrent client, and you can claim no ill intent... you may have trouble but it's a lot more defensible in court.
I would find it a big stretch to say Github's intent here is to illegally distribute copyrighted code. No judgment on whether the class action has any merit, just saying I would be very surprised if discovery turns up lots of emails where Github execs are saying "this is great, it'll let people steal code."
Almost everything on GitHub is subject to copyright, except for some very old works (maybe something written by Ada Lovelace?), and US government works not eligible for copyright.
Now, many of the works there are also licensed under permissive licenses, but that is only a defense to copyright infringement if the terms of those licenses are being adequately fulfilled.
Agreed. Like I said, it's about intent. Can anyone say with a straight face that copilot is an elaborate scheme to profit by duplicating copyrighted work?
I don't think the defense is that it wasn't trained on copyrighted data. It obviously was.
I think the defense is that anything, including a person, that learns from a large corpus of copyrighted data will sometimes produce verbatim snippets that reflect their training data.
So when it comes to copyright infringement, are we moving the goalposts to where merely learning from copyrighted material is already infringement? I'm not sure I want to go there.
The issue isn't downloading copyrighted stuff.
Rather, it's making available and letting others download it. That was where you got in trouble.
People used to get busted from buying bootleg VHS and DVDs on the street before P2P filesharing was a common thing. Then, early on, people were sued for downloading copyrighted files before rightsholders decided to take a different legal strategy to go after sharers and bootleggers.
What made Napster illegal is that the company did not create their network for fair use of content, but to explicitly violate copyright for profit.
Copilot is like Napster in this case, in that both services launder copyrighted data and distributed it to users for profit.
Copilot is not like other P2P networks that exist to share data that is either free to distribute or can be used under the fair use doctrine. Copilot explicitly takes copyrighted content and distributes it to users in violation of licenses, that's its explicit purpose.
It's entirely possible to make a Copilot-like product that was trained on data that doesn't have restrictive licensing in the same way it's entirely possible to create a P2P network for sharing files that you have the right to share legally.
It might be fair to say that the read performed in training has the same character since no human is involved.
The real copyright violation would be using a derived work.
when someone uploads their copyrighted text to a web page they are distributing it to whoever visits that page. the browser is just the medium.
Copilot's corpus is quite literally tomes of copyrighted work that are encoded and compressed in its neural network, from which it launders that work to create similar works. Copilot itself, the neutral network, is that corpus of encoded and compressed information, you can't separate the two. Copilot stores and distributes that work without any input from rightsholders, and it does it for profit.
A better analogy would be between a browser and a file server filled with copyrighted movies whose operator charges $10/mo for access. The browser is just a browser in this analogy, where the file server is the corpus that forms Copilot itself.
If you think this way, hashing is a copyright violation.
So copilot is fine but anyone using it must abide by the collective set of licenses that it used to write code for you…?
Note that even licenses like MIT ostensibly require attribution.
Can the generated code be traced back to the code used for training and the original copyrights and licenses for that code?
If so, what attribution(s) and license(s) should apply to the generated code?
Because I sure have seen that exact code written, from scratch, in many many places.
I guess my question boils down to "What is the smallest copyrightable unit of code?". Because I'm certain suing a novelist for copyright infringement on a character that says "Hi, how are you?" would be considered absurd.
Really what it comes down to is do you have enough resources to convince a judge or jury that X is a copy of Y? Doesn't really matter the size of X.
In other words, if your open source project doesn't have such immediately recognizable code and didn't cause a shitstorm on Twitter, chances are copilot is still happily spewing out your exact code, sans the copyright and license info.
If I'm being honest I'm a bit annoyed at this. What's the problem and what's the point of this?
"AI needs to be fair & ethical for everyone. If it’s not, then it can never achieve its vaunted aims of elevating humanity. It will just become another way for the privileged few to profit from the work of the many."
Blah blah. Can we get back to the hacking on stuff mentality?
Not exactly the curriculum of a twitter weirdo.
I wasn't actually talking about him specifically btw when saying "this sounds like a crypto bro from twitter". The overly enthusiastic AI talk reminded me of that, that's what I wanted to say.
I notice often on hackernews that people don't seem to understand anything about free or open-source software outside of the pragmatics of whether they can abuse the work for free.
But I'll bite: I know licensing, thank you. But what's copyrightable is not so easy. Licenses are not so easy. Copilot does not copy entire works and it's very questionable if a few lines of code are "piracy". It's a repeating discussion again and again, there's nothing novel about it except for the fact that a machine learns (and overfits for small portions of code). So please get off your high horse. I don't care for your fundamentalism.
You don’t need any fundamentalism to know that copilot’s output carries huge and untested legal risk. If this lawsuit clears some of this up, that’s a big win for everyone.
> You don’t need any fundamentalism to know that copilot’s output carries huge and untested legal risk. If this lawsuit clears some of this up, that’s a big win for everyone.
I agree with that! I also see this as the only proper takeaway that I think is ok. The rest is making money off this thing. But the US has a different law suit culture anyway, which I find weird.
Unrelated, how is it that Mechanical Turk was never truely integrated w/ AWS?
[0] https://www.mturk.com/
[1] https://waxy.org/2008/09/girl_turk/
If you have co-pilot trained on my code base (which was private), that then reproduces near replica's of my code then they sell it for $5/year...
Well, I'm eligible for damages.
(If it was, please tell me how, since that would save me $5/year across multiple libraries..!)
Copying a few lines is not the same as copying the whole thing. Sharing quotes from a book is not copyright infringement.
Using clips from a movie in a movie review is probably fair use.
Using clips from a movie in knock-off of that movie for profit? Probably not fair use if it's not a parody.
Copilot is not like a movie reviewer using clips to review a movie. Copilot is like a production team for a movie taking clips from another movie to make a ripoff of that movie and selling it.
Consider every repo on github to be a movie. Copilot is taking individual frames out of every movie on github and composting them into a new film.
I think most of us would agree that individually, each frame is copyrighted. But what if you take one frame from a million different movies and put them in an order that produces a new coherent movie?
The core question we need to settle in court is: does the new movie become its own copyrightable work, or is it plagiarism?
https://en.wikipedia.org/wiki/Sampling_(music)#Legal_and_eth...
I.e. any use without permission is illegal.
I can't think of a 5 line snippet I've written or read that makes sense to claim ownership of. They don't stand on their own in the way even a 30s movie clip does.
Copilot is fair use and transformative -- that is unless there is an open source Copilot that Copilot is training on, only then would it be competing and it's easy for GitHub or OpenAI to exclude those repos of copilot alternatives from the training set.
It is if I take those quotes and publish them as my own in my own book.
Code which anybody can view is called "source available". You aren't necessarily allowed to use the code, but some companies will let their customers see what is going on so they can better integrate the code, understand performance implications, debug and fix unexpected issues, etc. The customers would probably face significant legal risks if they took that code and started to sell it.
"Open source" code implies permission to re-use the code, but there is still some nuance. Some open-source licenses come with almost no restrictions, but others include limiting clauses. The GPL, for example, is "viral": anybody who uses GPL code in a project must also provide that project's source code on request.
What do you think the chances are that Microsoft would surrender the Copilot codebase upon receipt of a GPL request?
Sometimes damages are statutory, i.e. they have a fixed dollar amount written right into the law. This lawsuit references one such law: https://www.law.cornell.edu/uscode/text/17/1203
They did not actually calculate damages in terms of lost movie tickets or estimates vs actually sales number of sold game copies. When it came to pre-releases where such product wouldn't have been sold legally in the first place, they simply added a multiplier to indicate that the copyright owner wouldn't have been willing to sell.
For software code, an other practice I have read is to use the man-hours that rewriting copyrighted code would cost. Using such calculations they would likely estimate the man hours based on number of lines of code and multiply that with the average salary of a programmer.
The average salary of a programmer in which country?
So much programming is outsourced these days, and in some places programmers are very cheap.
If the court wanted to distinguish between Microsoft using their own programmers to generate code vs taking code from github users, then the salary in question would likely be that of Microsoft programmers. It would then be used to illustrate how a legal training data would look like compared to an illegal one.
If "I took your code and trained an AI that then generated your code" is a legal defense, the GPL and similar licenses all become moot.
https://docs.github.com/en/site-policy/github-terms/github-t...
"You grant us and our legal successors the right to store, archive, parse, and display Your Content"
Copilot displays content. Case closed.
"This license does not grant GitHub the right to sell Your Content. It also does not grant GitHub the right to otherwise distribute or use Your Content"
If someone wants to use it commercially without complying with the GPL, I have no problem with allowing that, for a price.
Either use the code freely and openly, or pay me so you can make money on my code.
Copilot could conceivably allow someone to use my code commercially (and in a closed manner) without negotiating with me, the copyright holder.
The AI can copy things if it wants, but it can also modify things to the point of being fair use, and it can even create new works with so little of any particular work that it's effectively creativity on the same level of humans when they draw something that popped into their heads.
Did you know before airplanes were invented common law said you owned the air above your land all the way to the heavens.
Learning off code isn't the same as using the code as-is.
https://en.m.wikipedia.org/wiki/BSD_licenses
Remember the lawsuit of HiQ labs vs LinkedIn? Scraping, or viewing public data on a public webpage is legal.
https://gizmodo.com/linkedin-scraping-data-legal-court-case-...
I have no idea if this license language works or doesn't, but this is hardly the least productive subthread on this story. It's concrete and specific, and we can learn stuff from it.
The legal footing that copyright gives you, on which licensing rests, certainly empowers you to limit things about how others may redistribute your work (and things derived from it), but does it empower you to limit how others may read your work? As a ridiculous example, I don't think it would be enforceable to have a license say "this code can't be used by left-handed people", since that's not what copyright is about, right?
I think we can constrain use with the third clause.
My question is, how should we word that clause?
The redistribution happens later, either when copilot blurps out some of your code, or when the copilot user then distributes something using that code (I'm curious which). At that point, whether some use of your code is infringing your license doesn't depend on the path the code took, does it? (in which case #3 is moot)
The only change I made is adding clause 3:
Such language must be carefully written. What is the definition of “construction” and “operation” in a legal context? What is a “predictive software generation system”? That’s a very specific use case, you sure you covered everything you want to prohibit?
You’ve inserted your clause in such a way that this dependency cannot be used in any way to build anything similar to a “predictive software generation system”, even with attribution, as it would fail clause 3.
You have to consider that novel licenses make it difficult for any party that respects licenses to use your code. It is difficult to make one-off exceptions, especially when the text is not legally sound. So adoption of your project will be harmed.
So if you are serious about this license, you need a lawyer.
I'd love to know if some of my GPL v3 code [1, 2] has landed in the training set
[1] https://github.com/4kbt/NewtonianEotWashToolkit/blob/master/...
[2] https://github.com/4kbt/NewtonianEotWashToolkit/blob/master/...
Intereting that my results were different than yours!
For pointMatrixGravity: https://gist.github.com/ridiculousfish/af05137a4090e92de3a97...
function force=Gmmr2Array(mass1, mass2) {
}function [force, torque]=pointMatrixGravity(array1,array2) {
a) The work must carry prominent notices stating that you modified it, and giving a relevant date. b) The work must carry prominent notices stating that it is released under this License and any conditions added under section 7. This requirement modifies the requirement in section 4 to “keep intact all notices”. c) You must license the entire work, as a whole, under this License to anyone who comes into possession of a copy. This License will therefore apply, along with any applicable section 7 additional terms, to the whole of the work, and all its parts, regardless of how they are packaged. This License gives no permission to license the work in any other way, but it does not invalidate such permission if you have separately received it.
——
I don’t see how one could argue that training on GPL code is not “based on” GPL code.
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)
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.
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...
Actually, we were forbidden to look at open source code at Microsoft (circa 2009) because it might influence our coding and violate licenses.
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.
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.
(and I guess courts might, in the future, say the GPL expires when copyrights on the code expire)
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)
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)
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.
Just because both activities are calling "learning" does not mean they are the same thing. They are fundamentally, physically different activities.
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.
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.
Quite sure the issue at hand is about the code being copied verbatim without the license terms, not "learning" from it.
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.
> 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.
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.
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).
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.
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.
I find the pattern matching and repetitive code generation really helpful. And the library autocomplete on steroids, too.
Meh. Tricky subject.
It's not just functions either, one of the most common things that it helps me with daily is simple stuff like this:
Typing
And later I'll be typing It's really amazing the amount of busy-work typing in programming that a smart pattern matching algo could help with.All of that efficiency without having to pay a monthly subscription, wasting electricity on some AI model, and worrying about the legal/moral implications.
going from
to just requires you to add two brackets and remove the colon. With multiple cursors you can do that exact same operation for all lines in a few keystrokes.But what's lost in my over simplified example is the contetxt is usually way more involved. I'm usually passing those as arguments to some function or other unique syntax situation that a glorified find and replace can solve. It's all about doing it in the times you would never think even bother writing a custom command because typing is faster given the unique syntactical context... The only thing faster then is autocomplete.
I'm not actually recreating a new hash with the convienient same format.
This is automated and happens immediately without you even thinking about it.
You only ever pull out the complicated Vim editing when you have a particular hard task, I’m talking about the small stuff many times a day.
Which reminds me I have to cancel my tabnine subscription. Been paying them for a year without using it.
That's where the line is for it to be suspect IMO.
It's a bit like how GPT-3, Stable Diffusion and all those generative models use extensive amounts of copyrighted material in training to get as good as they do.
In those cases however the output space is so vast that plagiarism is very unlikely.
With code, not so much.
https://hyperallergic.com/766241/hes-bigger-than-picasso-on-...
The interesting thing is that the names get explicitly attached to these styles. It isn't exactly a copyright issue, but I'm sure it will get litigated regardless.
Telling apart what's public domain or not is not a trivially automatable task.
If one just relies on curated libraries of vetted public domain content you don't get, by far, the expected amout of variability and diversity.
And maybe models trained on public data should be in the public domain, so that AI research can happen without requiring massive investments to obtain the training data.
You just described open source software.
That's the whole heart of this lawsuit, and equally Copilot. It was trained on OSS which is explicitly licensed for free use.
> It was trained on OSS which is explicitly licensed for free use.
That's not what the lawsuit is about. It's not about money, it's about licensing. OSS licenses have specific requirements and restrictions for using them, and Copilot explicitly ignores those requirements, thus violating the license agreement.
The GPL, for example, requires you to release your own source code if you use it in a publicly-released product. If you don't do that, you're committing copyright infringement, since you're copying someone's work without permission.
Also, re: your edit, not quite. They require you to release modified source under certain conditions if you make modifications to it. If everybody had to release code using GPL to the world, every companies code would currently be released to the world. There's more nuance than that. The gnu site covers a lot of that nuance (https://www.gnu.org/licenses/gpl-faq.en.html#UnreleasedMods)
LGPL is the one that enterprises won't touch with a 10 foot pole, due to more restrictive licensing, and more conditions under which you'd have to open source your own code.
The same cannot be said for Copilot: there have been prior examples here on HN showing that it can emit large chunks of copyrighted code (without the license).
Most open-source software is not licensed for free use. MIT and GPL, the two most common licenses, both require attribution.
Obvious licensing needs to be respected and it shouldn’t be hard to solve that problem. But 99.9% of code isn’t some unique algorithm, it’s gluing libraries and setting up basic structures.
Most of the examples I’ve seen done line up with the reality of code completion tools. Code is rarely valuable when broken up into its small parts.
Even copying a full codebase is rarely enough to draw value from… there’s way more to a software business than the raw code. But that’s a different problem.
Literally 10x faster development.
Case in point: had an unexpected project and no time to complete it. Within an hour Copilot helped me:
* Write a couple of tricky matplotlib plots
* Do some extensive analysis with Pandas
* Write a couple of SQL queries
* Write a Flask back-end and deploy it
* Write a bit of a front-end
* This all with extra comments , links to documentation and pretty reasonable style
I have experience with all of the above mentioned but the speed increase was considerable.
This would a a good day's work without Copilot and there would be less commenting and hackier code.
Before Copilot I would be cursing a lot more reading various docs...
The key thing that Copilot does it reduces latency for your thoughts-action-results loop.
Does the open source really suffer if less people read documentation directly? Would you really be less likely to create an open source library if you knew someone can now use your library at 10x speed?
The inference ability has crossed uncanny valley so many times.
I find myself wondering whether there is a speech recognition component at times.
When teaching a lecture I will start saying something and write a prompt at the same time and the sentence produced by Copilot will be spot on what I've just said.
Ideally there would an open source version of Copilot that respects everyone's wishes. I fear that is impossible.
Obviously not financially as Microsoft has basically YES amounts of money.
If you are opinionated but lazy, no judgement here as I sit here watching TV, you could add a notation at the top of your repos explicitly supporting the usage of your code in such tools as fair use.
Notably if your code is derivative of other works you have no power to grant permission for such use for code you don't own so best include some weasel words to that effect. Say.
I SUPPORT AND EXPLICITLY GRANT PERMISSION FOR THE USAGE OF THE BELOW CODE TO TRAIN ML SYSTEMS TO PRODUCE USEFUL HIGH QUALITY AUTOCOMPLETE FOR THE BETTERMENT AND UTILITY OF MY FELLOW PROGRAMMERS TO THE EXTENT ALLOWABLE BY LICENSE AND LAW. NOTHING ABOUT THIS GRANT SHALL BE CONSTRUED TO GRANT PERMISSION TO ANY CODE I DO NOT OWN THE RIGHTS TO NOR ENCOURAGE ANY INFRINGING USE OF SAID CODE.
Years from now when such cases are being heard and appealed ad nauseam a large portion of repos bearing such notices may persuade a judge that such use is a desired and normal use.
You could even make a GPLesque modification if you were so included where you said. SO LONG AS THE RESULTING TOOLING AND DATA IS MADE AVAILABLE TO ALL
Note not only am I not your lawyer, I am not a lawyer of any sort so if you think you'll end up in court best buy the time of an actual lawyer instead of a smart ass from the internet.
This person (a lawyer) saw an opportunity to make money and jumped on it like a hungry tiger on fresh meat.
> But neither Matthew Butterick nor anyone at the Joseph Saveri Law Firm is your lawyer
This is curious. None of them are my lawyers, but surely at least some of them are someone's lawyers? Isn't it wrong to put such a blanket disclaimer on a website which might well be read by their clients?
>function isEven(n) {
> return n % 2 === 0;
>}
They then say, "Copilot’s Output, like Codex’s, is derived from existing code. Namely, sample code that appears in the online book Mastering JS, written by Valeri Karpov."
Surely everyone reading this has written that code verbatim at some point in their lives. How can they assert that this code is derived specifically from Mastering JS, or that Karpov has any copyright to that code?
"In computer programs, concerns for efficiency may limit the possible ways to achieve a particular function, making a particular expression necessary to achieving the idea. In this case, the expression is not protected by copyright."
https://en.wikipedia.org/wiki/Abstraction-Filtration-Compari...
Think about how absurd this is. So if Microsoft was the first company to write and publish an isEven function then no one else can legally use it?
As a note the same applies to logos. Very simple logos that are only some lines and shapes, do not have copyright (in usa)
Hey, I said the same thing about APIs, but here we are.
Edit: Actually, the Supreme Court declined ruling whether APIs are copyrightable, but they did say that if they are, reusing them like google reused the java apis in android would fall under fair use. Given that lower courts did think that APIs should be copyrightable, we don't know if they are anymore.
Ironically their Twitter account uses a screenshot from a TV series as profile picture. I wonder how legal that is, even if meant as a joke.
https://twitter.com/saverlawfirm
Edit: It's been changed 2 minutes after I wrote this comment
Or is your comment itself the joke?
"Joined November 2022", following one account and no followers. It's generous to consider it a genuine account, no?
However, if you are looking to understand the reasoning behind this lawsuit, there are lots of better examples online where Copilot blatantly ripped off open source code.
Look at paragraphs 90 and 91 on page 27 of the complaint[1]:
"90. GitHub concedes that in ordinary use, Copilot will reproduce passages of code verbatim: “Our latest internal research shows that about 1% of the time, a suggestion [Output] may contain some code snippets longer than ~150 characters that matches” code from the training data. This standard is more limited than is necessary for copyright infringement. But even using GitHub’s own metric and the most conservative possible criteria, Copilot has violated the DMCA at least tens of thousands of times."
Does distributing licensed code without attribution on a mass scale count as fair use?
If Copilot is inadvertently providing a programmer with copyrighted code, is that programmer and/or their employer responsible for copyright infringement?
There's a lot of interesting legal complications I think the courts will want to adjudicate.
[1] https://githubcopilotlitigation.com/pdf/1-0-github_complaint...
The situation that this lawsuit is trying to save you from is this: (1) copilot blurps out some code X that you use, and then redistribute in some form (monetized or not); (2) it turns out company C owns copyright on something Y that copilot was trained on, and then (3) C makes a strong case that X is part of Y, and that your use of X does not fall under "fair use", i.e. you infringed on the licensing terms that C set for Y.
You are now in legal trouble, and copilot put you there, because it never warned that you X is part of Y, and that Y comes with such and such licensing terms.
Whether we like copilot or not, we should be grateful that this case is seeking to clarify some things are currently legally untested. Microsoft's assertions may muddy the waters, but that doesn't make law.
If not, it's a pretty clear sign they consider it radioactive.
I've noticed this a lot and it's quite funny seeing what the actual filename of the document was. Does this just get included as metadata by default when you export to PDF?
[0] https://githubcopilotlitigation.com/pdf/1-0-github_complaint...
The extra steps aren't enough to exhonorate them. It's just a convoluted copy operation.
Is just like how a lossy encoding of a song is still - with respect to copyright - a copy of that song. The data is totally different, and some of the original is missing. It's still a derivative work. So is a remix. So is a reperformance.