Also the point you're making is controversial, and definitely isn't widely agreed on. As a human, I can read public code on Github, and use my internal neural network (brain) to regurgitate sections of code, and don't need to attribute anyone (who can say which codebase I'm recalling code from? I certainly can't). So a neural network doing the same thing, but external to a human, is certainly questionanble, but it isn't a cut-and-dry case of copying without attribution.

Also, humans sometimes have limitations on what code they should have seen. It's common in the emulator community to forbid anyone that has seen proprietary code from working on black-box implementations to avoid legal issues.

Microsoft's automated violation of "no use without attribution" licenses?

You can't see the connection?

Please. Surely it's obvious to everyone but you.

Wrong. Lets say a GPL project is not hosted on GitHub officially. I can easily setup a mirror for it though on GitHub as the GPL doesn't prevent me from doing it...

Point is that anyone can put my work on GitHub, even if I don't want to.Assuming the project is under a free license though.

You can do it, but it's WRONG.

Copilot is trained on public GitHub repositories of any license: https://en.m.wikipedia.org/wiki/GitHub_Copilot#Technology

We must all stop using GitHub.

If Microsoft's theory is correct, under US law it is impossible to forbid this with a license, because a license is just an offer of additional permissions beyond what is available automatically under law, and Microsoft's theory is that ingesting code into GitHub is fair use and therefore permitted under law as an exception to copyright without any license from the copyright owner.

If Microsoft's theory is not correct, pretty much any license with an attribution requirement (among others, for other reasons) would work.

The idea that a special license is needed or of any use doesn't seem to have any justification, even in theory. (As is the idea that hosting publicly but not on GitHub changes the legal parameters.)

I just want to say that I believe the fair use argument only applies for training, but not for distribution. I make the case for that in [1].

[1]: https://gavinhoward.com/uploads/copilot.pdf

If they don't need permission from the copyright owner, either via license of GitHub T&C, because it's fair use, which is their overt legal theory, then why would it matter legally whether the code was posted to GitHub at all, much less by whom? Ingesting only code form GitHub is a practical convenience that has nothing to do with their legal theory of the right to do it.

> Yet they said nothing about that specifically

Their theory of fair use means they have the right to ingest any code, irrespective of who owns it and what conditions (if any) it is licensed under or where (or even if) it is hosted online. They don't need a separate justification for your scenario if the theory they’ve cited is correct.

> In that case, is the only option to put code somewhere else than github under a license that forbids reuploading to github, and issue DMCAs when/if people reupload your code

Nope, that doesn't help at all, legally; it may help practically as long as they are just using GitHub hosted code and not consuming code from other public hosting platforms, but it has no bearing on their legal theory of why they can ingest code without additional permissions.

What I did is say that if you feed copyrighted software to an algorithm that itself outputs software, then the license applies to the output. This covers the output of compilers and such, but it would also cover Copilot in my opinion. We'll see what a lawyer says.

However, even with a license, I wouldn't doubt that Microsoft would just put it through GitHub anyway because finding them out would be extraordinarily hard.

[1]: https://yzena.com/licenses/

There is a difference between "strong" copyleft and "weak" copyleft. An example of "weak" (non-viral) copyleft is the CDDL. In fact, the CDDL's Wikipedia page talks about strong and weak copyleft.

You can read [1] for a breakdown of copyleft by an actual lawyer. Suffice it to say that Wikipedia's Copyleft page is woefully inadequate.

[1]: https://writing.kemitchell.com/2018/10/24/How-to-Speak-Copyl...

Because this does not prevent having the code to land on GitHub at the of the day assuming it is published under a free license.

Now it depends on how you interpret the "MUST". My logic only makes sense if you consider it to a dogmatic-like prevention.

> The larger issue is that anyone using GitHub is donating their work for re-use without attribution through Copilot.

Why is this wrong?

Also, I don’t see how this is true. Code on my website is publicly accessible, but not in the public domain, nor licensed for re-use, unless I say that it is.

If you don't like that,don't release your code as open source. I'm not going to host my open source projects on my own website. That's just hard, it's much more difficult to get people to check it out if it's on Broblog.net

I personally don't believe extremely short code snippets, like the ones copilot tends to copy are problematic.

Here.

def add( x, y): return x + y

I don't want to live in some dystopia where we have dozens of lawyers deciding who owns that above code snippet. If Amazon wants to use that snippet, fine, you can use it, etc.

All Co Pilot does is optimize taking code snippets from different sources.

In these situations, whether it be originating from a human or robot, the knowledge comes from looking at public code on GitHub and it's always been a risk that your public code might not be used with proper attribution at some point. Think about how many OSS projects have core code and algorithms copied daily by companies with no public name and keep all of their source code private - it's surely caused more damage than CoPilot ever will.

The test to use is to imagine you’re writing an essay in college. If your essay contains unattributed passages from another work then the professor will not care that you memorized them rather than just copying and pasting the text. In order to be in the clear you need to properly quote and cite the original author.

GitHub copilot does not provide attribution. All it does is obfuscate the original source to make attribution impossible. Copyright lawyers ought to have a field day with this one.

>In that B scenario, would the human be infringing on the original's copyright? //

Depends on jurisdiction but in general I'd say no as quoting, particularly an insubstantial portion, is allowed (certainly under USA Fair Use, possibly under UK Fair Dealing).

However, and again depends on jurisdiction, copying a whole work to use for AI training or searching can be restricted (though it's on the GitHub license in this case). The snippet might be allowed but copying into a training corpus may be excluded (I think UK have an exception for AI training, but I may have misremembered, it might be a suggestion??).

There are quite a few companies where copilot and copoilot-like technologies seem radioactive at least for the moment.

Even if Microsoft currently only uses GitHub-hosted code as an input to Copilot, their theory of the legality does not depend on code being hosted there and applies to any code they can get their hands on. The idea that hosting code publicly at some non-GitHub location is going to keep it out of Copilot is not well justified.

How would self hosting your code prevent Microsoft/GitHub from using it in the Copilot training dataset? If using content from GitHub irrespective of their license to train Copilot is legal, so is training from code available on your website.

It might be time for popular licenses to update themselves for this new-found threat.

No, Microsoft's claim is fair use, not GitHub terms and conditions. You don't have to speculate on what you think their possible claims are, when they've publicly stated their actual claims.

I’d be interested to know if code subject to DCMA takedown would be removed from the copilot training set…

Did you click the link for this article? Did you see that MS removed the author’s name in the license header and replaced it with their own?

[1]: https://github.com/bminor/glibc

Copilot is trained on public GitHub repositories of any license: https://en.m.wikipedia.org/wiki/GitHub_Copilot#Technology

If they scrape your website, that's different.

0: https://docs.github.com/en/github/site-policy/github-terms-o...

1: > This license does not grant GitHub the right to .. otherwise distribute or use Your Content outside of our provision of the Service

Not a very good one - clicking the link produces a download dialog on Firefox (I'm guessing because the website neglects to indicate a Content-Type).

Ironically this is an argument against trying to do everything yourself - you might waste time chasing the long tail of thousands of little details that had been solved many times over elsewhere.

Also hosting it yourself doesn't mean that you do everything yourself.

It would definitely spawn build plugins that do most of what you need if this practice actually establishes itself.

Works fine for me (v90/Linux).

I don't know what to tell you. It's not there.

https://dump.cy.md/9b2a31fa5184397159fe42a85c197242/16404476...

It is there if I check with cURL, though.

Edit: looks like this bug is triggered by the absence of Accept-Encoding in the request. When the server compresses the response, it neglects to include the Content-Type of the compressed content.

> You defend Microsoft, here and in other parts of this thread, as they sell the work of open source authors without attribution and in violation of explicit license statements in the code.

I did no such thing. Please stop.

Please review https://news.ycombinator.com/newsguidelines.html and don't do anything like this on HN again.

You can search online for my name / username to find my open-source contribution history, if you would like.

I hope that sufficiently answers your query.

This is my last reply in this subthread; feel free to assume the same answer for all questions along the same line.

I also invite you to review the HN comment guidelines: https://news.ycombinator.com/newsguidelines.html

I think this can only be valid of Github's terms of use clearly specify that or the chosen license allows it.

I actually see copilot benefiting GPL projects: suppose a programmer uses copilot to develop a proprietary software and copilot regurgitates GPL'ed code: now the proprietary software is a derivative work and must be GPL'ed too.

Another potential outcome is that so many projects could now end up unknowingly using gpl code that enforcement becomes an impractical whack-a-mole, far more so than today. Being told to rework or relicense your project because of copiloted gpl code could easily end up with hobbyists wholesale begrudging the gpl license rather than copilot itself.

It's quite interesting how it seems to have come full circle, at least according to this stackoverflow comment about Stallman releasing a paper in 1992 on how to effectively launder AT&T code via textual changes (I have no idea if he did, I just remembered the comment): https://unix.stackexchange.com/a/591221

You may be practically limiting copilots use of your code but I don't see any licensing difference if Microsoft hosts Copyright code or scrapes copyright code.

So yeah, you are donating your work to Google when you put it on a publicly accessible web site.

[1] https://www.google.com/search?q=how+old+is+queen+elizabeth

[1] https://developer.mozilla.org/en-US/docs/Web/HTML/Microdata

I think that’s on purpose, not to “steal” code from GPL’ed softwares’ authors, but to get GPL’ed code into commercial projects. Then Microsoft can say “oops, we were right all along! The GPL is so dangerously viral that you can’t even host your software on the same server!”

I know that sounds stupid, unless you were there for the Halloween Documents. After all these years, it seems to me that Microsoft hasn’t done the 180 they portray.

I might put this in my public ones as well.

I am a happy Sourcehut customer. Roughly the same cost as a VPS, but comes with tech support.