If you read something that is copyrighted and rewrite it from memory, that’s still copying for purposes of the copyright act. There are cases about music where that’s unintentionally happened and the inadvertent copier was held liable. It’s basically what the machine learning is doing. Committing something to memory is no defense to copyright - I would recon a good number of people could recite any number of copyrighted works from memory (e.g., episodes of The Office, Seinfeld, or popular songs). If a person wrote that down and tried to sell it, I’m convinced a court would hold that person liable. There is a defense of “independent creation,” but basically if you saw/heard the copyrighted work before then it’s very hard to use that as a defense.

A human won't "accidentally" retype the source code to Quake, including comments, changing only the license - so no, it isn't like that.

If GitHub survives in court over this, I guarantee you someone will :3

You’ve fallen into the trap of thinking machine learning = human learning. It’s not at all equivalent. You can’t give GP3 4 years of course work at Carnegie Mellon and get autopilot. Similarly you can’t get a human to consume the complete GitHub corpus and even if they had the time they’d be unable to repeat code snippets they read verbatim with simple prompting.

If copilot produces a verbatim copy or just small changes, like changing the names of variables, for a long enough piece of code, this is definitely copyright infringement if they don't obey the license terms of the code that copilot "learned" from.

While I disagree with this sentiment, if we take this as true and consider it "the same as a human learning to code from gpl libraries" this opens up an entirely different can of worms when it comes to patent assignment.