DeVault: GitHub Copilot and open source laundering [LWN.net]

DeVault: GitHub Copilot and open source laundering

Posted Jun 23, 2022 22:26 UTC (Thu) by jmspeex (guest, #51639)
Parent article: DeVault: GitHub Copilot and open source laundering

Actually, this is not just about the GPL and copyleft licenses. Even a license as permissive as the BSD still has a few requirements (including the copyright notice) that are easy to break with ML code generation. So if the code your ML model spits out is BSD-licensed, you could still get in trouble over it.

to post comments

DeVault: GitHub Copilot and open source laundering

Posted Jun 24, 2022 1:07 UTC (Fri) by developer122 (guest, #152928) [Link] (15 responses)

This is an important point.

There were a raft of people arguing that all code from copilot should be GPL due to ingested GPL code, to which I reply: What gives the GPL priority?

There is code on github under lots of licences, eg the CDDL which is a copyleft free software licence but incompatible with the GPL. As there is significant CDDL code on github, under the same reasoning perhaps all the output should be CDDL?

An then there's all the code on github which specifies no licence at all. Just because it is publicly available does not mean any and all rights have been granted to you. Plenty of this code is proprietary.

The end result is an unlicencable pile of legal mush that no sane lawyer should be going anywhere near.

DeVault: GitHub Copilot and open source laundering

Posted Jun 24, 2022 8:16 UTC (Fri) by bluca (subscriber, #118303) [Link] (14 responses)

Training of the model is not done under the terms of any license. It's done under the copyright directive laws, which give an exception on copyright rules when doing text and data mining, for any purpose. I don't see how there's any 'legal mush'.

DeVault: GitHub Copilot and open source laundering

Posted Jun 24, 2022 9:19 UTC (Fri) by Vipketsh (guest, #134480) [Link] (9 responses)

I think the "legal mush" comes from what constitutes data/text mining and if the entirety of copilot wholly fits under it.

I can see it being difficult to dispel an argument that Copilot does something more than just mining when it is able to regurgitate numerous lines of code verbatim, including comments. Is it text/data mining if the representation of some code is changed to a set of statistical weights ? Is that any different to, say, compressing the code ? Is it different to encrypting the code requiring a password to gain back the original representation ?

DeVault: GitHub Copilot and open source laundering

Posted Jun 24, 2022 11:09 UTC (Fri) by bluca (subscriber, #118303) [Link] (8 responses)

I'm not a lawyer, but me thinks you'll have a pretty hard time trying to prove getting text bodies to train an AI model does not constitute Text and Data Mining. I mean, it's pretty much what it is for, and AI/machine learning are mentioned plenty of times by the legislators as examples, as far as I can see.

DeVault: GitHub Copilot and open source laundering

Posted Jun 24, 2022 11:52 UTC (Fri) by Vipketsh (guest, #134480) [Link] (7 responses)

It's the same thing, just you changed the question to "what is an AI model?" and what constitutes "training". If it is allowed to produce a black box where you feed in some code and then get part of that code out the other end verbatim what limits are placed on the inside of that black box ? Can I call an HDD an "AI model" and "train" it by storing some code on it ?

The major considering factor in this case for me (IANAL) is that if "AI models" are allowed to reproduce code verbatim then there is a legally sound method to de-license any and all code: Take something similar to copilot train it with the code you would like to de-license and then get your device to reproduce it. That is clearly not acceptable, yet it is what copilot does in some cases.

In the end this is a new situation which touches on various concepts and courts and/or law makers need to figure out where this and similar situations stand.

DeVault: GitHub Copilot and open source laundering

Posted Jun 24, 2022 13:01 UTC (Fri) by bluca (subscriber, #118303) [Link]

Again, not a lawyer, but if you think you can convince a judge that an hard drive is an AI model, I wouldn't put my money on it, but knock yourself out.

DeVault: GitHub Copilot and open source laundering

Posted Jun 24, 2022 18:37 UTC (Fri) by NYKevin (subscriber, #129325) [Link]

IMHO we don't need to figure this out, unless somebody wants to allege that GitHub is violating the AGPL by not providing the model's source code, parameters, etc. Such a lawsuit would be complicated, messy, and bring in all sorts of difficult legal issues.

For literally any other legal claim under the sun, you can just point directly at the output and say "Regardless of how it was created, that output is substantially similar to some portion of the XYZ codebase, and therefore it is a derivative work of XYZ and is in violation of [whatever license applies]." Judges have been dealing with the "we can't directly prove that the defendant copied the plaintiff's work" problem for a very long time, and are perfectly capable of finding for the plaintiff anyway.

DeVault: GitHub Copilot and open source laundering

Posted Jun 24, 2022 19:05 UTC (Fri) by samth (guest, #1290) [Link] (2 responses)

No, you can't "trick" copyright law. A judge would say "obviously those are different". The fact that from a computing perspective, there is a sense in which they're the same thing, or that there is not way to draw a bright line distinction between a hard drive and Copilot, doesn't matter at all. The law does not work the way computers do.

DeVault: GitHub Copilot and open source laundering

Posted Jun 24, 2022 21:16 UTC (Fri) by Wol (subscriber, #4433) [Link] (1 responses)

> The law does not work the way computers do.

The problem is when the law says "black is white", or "true is false". What was that about the ?Pennsylvania legislature trying to define "pi = 3"?

Lawyers are supposed to understand logic. Maths is a subset of logic. Yet how many legal screw-ups do we have because lawyers want to legislate maths out of existence?

And then we get people here who think law is "black and white". PJ was quite clear - "law is squishy". BECAUSE lawyers don't understand black and white! Or because life mostly ISN'T black and white!

It would be lovely if things were that simple. But what answer do you get when you ask the question "what's 6 times 9?". When you ask the question "What is reality?" Because even a Physicist will turn round and say "the Universe doesn't know"!. YOUR reality is different from MINE, and we have no way of telling if there even IS a correct version of events.

For the most part, law tends to fix (or forget) its worst mistakes, but isn't that true of Computer Scientists too?

You need to try and twist reality to your view, or you'll find other people WILL twist it to theirs. Quite possibly without even trying or intending to ...

Cheers,
Wol

DeVault: GitHub Copilot and open source laundering

Posted Jun 24, 2022 22:52 UTC (Fri) by mathstuf (subscriber, #69389) [Link]

> What was that about the ?Pennsylvania legislature trying to define "pi = 3"?

Hey, PA has its problems. But this one came from Indiana. Luckily there was a school teacher that was there to help teach some math.

https://www.straightdope.com/21341975/did-a-state-legisla...

DeVault: GitHub Copilot and open source laundering

Posted Jun 24, 2022 21:12 UTC (Fri) by rgmoore (✭ supporter ✭, #75) [Link] (1 responses)

The major considering factor in this case for me (IANAL) is that if "AI models" are allowed to reproduce code verbatim then there is a legally sound method to de-license any and all code: Take something similar to copilot train it with the code you would like to de-license and then get your device to reproduce it. That is clearly not acceptable, yet it is what copilot does in some cases.

No, because copyright is fundamentally path independent. All that is needed is to show that an identifiable part of copyrighted work A shows up in work B. Once the author of work A has shown that, a copyright violation is assumed, and it's up to the author of B to provide a specific defense for why it isn't a copyright violation. There are valid defenses, but laundering the code through an AI is not one of them.

DeVault: GitHub Copilot and open source laundering

Posted Jun 25, 2022 6:00 UTC (Sat) by NYKevin (subscriber, #129325) [Link]

Clarification: Copyright is path-independent, provided that a path actually exists. In theory, you do have to show that the defendant had access to the copyrighted work, and that this somehow affected the contents of the allegedly infringing work. But in practice, this is a very weak requirement. You don't have to make specific claims about exactly what the defendant did or the path they took from A to B. You just have to show that the path probably exists. OTOH, if the defendant can plausibly argue lack of access (e.g. "Your code was proprietary, and I couldn't read it even if I wanted to"), then that is a valid defense to copyright infringement (provided, of course, that the finder of fact believes this argument).

DeVault: GitHub Copilot and open source laundering

Posted Jun 24, 2022 16:56 UTC (Fri) by Lennie (subscriber, #49641) [Link]

The comment above had a very nice example how it's NOT that simple:

https://lwn.net/Articles/898818/

DeVault: GitHub Copilot and open source laundering

Posted Jun 25, 2022 14:23 UTC (Sat) by eduperez (guest, #11232) [Link] (2 responses)

> Training of the model is not done under the terms of any license. It's done under the copyright directive laws, which give an exception on copyright rules when doing text and data mining, for any purpose. I don't see how there's any 'legal mush'.

I do not think anybody is arguing about the model training, but the output from the model once it has been trained.

If the model produces an output that is a verbatim copy of some GPL'd code (see https://twitter.com/mitsuhiko/status/1410886329924194309 for an example), is that code free now, just because it was produced by some AI? Or is it still protected by the GPL, because it is a derivative work? When can we consider that the output has been produced by the AI, and when can we consider it is still a derived work? This is the legal mush.

DeVault: GitHub Copilot and open source laundering

Posted Jun 25, 2022 21:27 UTC (Sat) by Vipketsh (guest, #134480) [Link] (1 responses)

I think rgmoore and NYKevin answered that question pretty convincingly in their replies to my comment above: if you believe that your code exists in something it shouldn't you argue that and courts don't care how it got to where it shouldn't have, only if it is there or not. Perhaps I can put it another way: what the AI model produces doesn't matter, only where that output ends up in.

Even so, I still think that the legal status of the "AI model" warrants further introspection. When talking about open source code the "AI model" isn't much of a consideration because at worst it stores code in some cryptic way that is available in a much more easier digested form. But that changes a whole lot if the training material is some propriety code stolen from somewhere. If your "AI model" is able to reproduce the stolen code verbatim (or sufficient parts for copyright to apply) and training of the "AI model" is "give an exception on copyright rules when doing text and data mining, for any purpose" (bluca's words) that should mean that this trained "AI model" is fully legal and thus a legal distribution mechanism for the stolen code. Surely there is a legal principal to prevent things to work out this way ? At what point does an "AI model" turn into a "distribution mechanism" ?

DeVault: GitHub Copilot and open source laundering

Posted Jun 26, 2022 9:46 UTC (Sun) by bluca (subscriber, #118303) [Link]

The TDM provisions in the copyright directive apply only for legally accessible corpora. A stolen body of code cannot be data mined legally just because it's available, it needs to be legally available.