> What problem was it causing, other than making Microsoft look bad by showing that the real problem exists?
That's the real problem. The issue is that Microsoft looks bad. The problem was solved by making Microsoft not look as bad. Verbatim output of inputs is not a problem, it's an understood property of the model.
So wait, was the obvious copyright elephant in the room solved somehow?
> The issue is that Microsoft looks bad. The problem was solved by making Microsoft not look as bad.
Depending on who you ask, adding a hack like this in an attempt to make them not look as bad just makes it look worse. Especially when the hack is discovered.
Try getting a non-technical person to understand this issue beyond the level of "okay, so it output some copyrighted code, then they blocked that code from being output. Sounds like they fixed it". That's the obvious PR angle, and once the big, widely-publicized issue is fixed, it's hard to get another article with that much traction. Besides, if another snippet becomes popular, it can be blocked the same way. The issue is fundamentally a PR one, not a copyright one, since the responsibility for using or not using the code relies on the end-user. GitHub has the rights to use your code for training.
> The issue is fundamentally a PR one, not a copyright one, since the responsibility for using or not using the code relies on the end-user.
Copilot's FAQ says this (under heading Who owns the code GitHub Copilot helps me write?): "GitHub Copilot is a tool, like a compiler or a pen. The suggestions GitHub Copilot generates, and the code you write with its help, belong to you, and you are responsible for it."
They are essentially affirming that the output is not covered by someone else's copyright, but that is far from clear.
And I think it is precisely the copyright issue that turned this into a PR issue. Verbatim copies are just a very obvious demonstration of the copyright issue. The issue isn't gone when they filter out this specific snippet; the people who are concerned about copyright issues are going to remain concerned.
> GitHub has the rights to use your code for training.
Sure, but that's not at all the same as saying that the output produced by the AI free of copyright issues. That's kind of orthogonal.
If you use a pen to write out a copy of Harry Potter and the Philosopher's Stone, then publish copies, you are guilty of copyright infringement, not the pen.
And the sky is blue. Look, we're not talking about what you write out, we're talking about what the ML model generates for you and what Microsoft claims about the ownership of the generated output. At this point I don't feel like you're arguing in good faith. Have a good one.
I don't think there's any technical knowledge needed here; the concepts are common ones. "There was a big hoo-hah about how it used to output a certain well-known piece of code in possible violation of its licence. They fixed it by blocking any output that includes the name of that exact paragraph of code. In particular, it can still output other code from the same codebase with the same licence; just not that particular paragraph."
The direct analogy would be trying to stop it from outputting The Lord of the Rings by blocking the phrase "a long-expected party" (the title of its first chapter).
I believe the argument from GitHub is that using public code as training data for machine learning falls under fair use [1]. If something is found to be fair use then (as far as I understand) copyright does not apply at all, and hence terms in the license do not make a difference. It is not clear whether this argument will stand up if tested in court (fair use is an affirmative defense, so something cannot actually definitively be said to be fair use until someone is accused of copyright violation and a judge rules that it actually is fair use).
That said, at the very least it seems like it would be rude to include code in the training data if the developer has expressly said they don't want that.
[1] From their FAQ: "Training machine learning models on publicly available data is considered fair use across the machine learning community."
I hope it'll get tested in court soon and ruled against GitHub/Microsoft - because otherwise this will mean that Copilot and other GPT-3-like models become perfect copyright laundering machines.
Testing the fair use argument for training won't necessarily answer the question about copyright laundering.
You could easily make the case that training is fair use, but that doesn't have to imply the model's output is non-infringing.
For example, it seems reasonable to train a model by feeding copyrighted texts and images, and that model could be useful for analyzing the content, finding facts, or detecting features. But we're in murky waters when the model also starts outputting the original content (be it verbatim or "derived").
Not all that different from human learning: you can study and learn from publicly available books but that doesn't grant you the right to recite their contents and claim it as your own, original work.
Again, if the AI spits out “copyrighted code”, I’d suggest that the code in question is insufficiently creative to be copyrighted to begin with.
Put another way, copyright only applies to creative expressions, not functional expressions. It does not matter how creative the idea is. If the work of authorship is software that embodies the function (and no other expression), it is not copyrightable.
So where is the line between creative and functional expression in software? The law does not provide clear guidance. Ultimately, it’s up to a judge.
That's the real problem. The issue is that Microsoft looks bad. The problem was solved by making Microsoft not look as bad. Verbatim output of inputs is not a problem, it's an understood property of the model.