> The court of law would probably have you unveil and tear apart your process and find that you were trying to plagiarize in a roundabout way.

This is the whole idea. Copilot is spitting out considerable chunks of code that is licensed under GPL and it will be up to GitHub to prove that Copilot is not trying to plagiarize this code in a roundabout way.

In the very least, Copilot should have separate data stores for different groups of licenses: public domain, attribution-only, copyleft, etc.. That would already make it much more usable than the current "here's some code, it came from I don't know where, don't ask me" that literally looks like black market deals except they are GitHub-branded.

> Copilot is spitting out considerable chunks of code that is licensed under GPL

One regurgitation in 10 weeks per user. Not considerable. Could be just skipped with a simple search by Github.

> You do understand that's not how laws work in general, right?

The only reason the law doesn't work this way for Microsoft Copilot is because the copyright holders are individuals who do not have the capital or expertise to file suit.

If Microsoft instead released a video editor addon that was trained on Disney movies and which would sometimes insert scenes of _any_ Disney movie you can bet your ass we wouldn't be having the same discussion.

Comparing code to movies - in code even a single char difference can change the meaning of everything, in movies - you can skip whole scenes and still get the meaning. I don't think the two are compatible, they are judged by different standards.

> The court of law would probably have you unveil and tear apart your process and find that you were trying to plagiarize in a roundabout way.

Well, yes, that's the point.

I’m not sure what the point is. That the argument of copying Disney movies is not a good analogy, or that courts will find plagiarism in software copyright cases involving Copilot?

The point is that Copilot is effectively doing the Disney movie thing, just with code, and yet this article argues this is all fine. As it is, the article turns Copilot into an universal copyright laundering machine.

> you were trying to plagiarize in a roundabout way

So using machine learning with one movie is illegal, but using it with a million isn't?

Generally speaking - probably yes. That’s because the output would be (a) transformative and (b) not likely to affect the profitability of the original work (at least not directly).

Note: I’m not really trying to comment specifically about the code/movie examples - just the general notion that the more input there is (from different sources), the likelihood that the use will be considered “Fair Use” increases.

https://en.wikipedia.org/wiki/Fair_use

Do you think this Copilot is some sort of advanced AGI who just became a genius programmer? Almost every piece of code that it "generates" you will find it almost verbatim in one or several public repos.

If you ever tried AI Dungeon before it imploded, you'd agree GPT-3 is pretty impressive and is generally capable of coming up with original stuff. Obviously with a bias towards the kind of stuff that's in the training dataset, which is why you try to use a very broad dataset. But still original.