This is exactly how it should work. Any code copied from GNU GPL code should legally also be GNU GPL, no ifs, ands or buts.
Oh, your training data set contains GPL code? Guess what, all the output from that model is now GPL. It’s your burden to prove that your training data does not contain any GPL code.
That’s probably too simplistic, but they sure don’t think like humans. But also humans can infect their thought processes in a way that the code they write ends up licensed by some other terms than what they thought it should be.
Like if you have read your employer’s proprietary codebase and write the same code somewhere else, they could sue you for that. Perhaps. IANAL.
In a general, open-ended situation, sure. In this case, the model developer has all of the training data that was input for training the model. It should take all of half an hour to determine if there is any GPL code in the training data set.
Hi, I am from OpenAI/Alphabet/Meta, I have checked our training data and it contains no GPL code, or copyrighted or pirated media of any kind, it’s all perfectly totally 100% legit.
Thank you for your inquiry, hope this solves the issue and have a good day.
Social media users want to see a person walk into the bad guy’s office and punch them in the face, solving the problem. Real effective steps are usually boring and unremarkable like forcing them to make a statement on record.
Exactly, this is how you win court battles that set legal precedents which establish real-world consequences, not as one-offs but every time somebody crosses that line.
This is exactly how it should work. Any code copied from GNU GPL code should legally also be GNU GPL, no ifs, ands or buts.
Oh, your training data set contains GPL code? Guess what, all the output from that model is now GPL. It’s your burden to prove that your training data does not contain any GPL code.
@ free software foundation: your mission, should you chose to accept it ☝🏻
LLMs obviously don’t copy code. That’s not how they work.
What you’re suggesting is sort of the same as claiming that if a developer ever read GPL code, all their code is now GPL code.
Nope, humans don’t generate tokens from a statistical model. We have a fundamentally different way of thinking.
That’s probably too simplistic, but they sure don’t think like humans. But also humans can infect their thought processes in a way that the code they write ends up licensed by some other terms than what they thought it should be.
Like if you have read your employer’s proprietary codebase and write the same code somewhere else, they could sue you for that. Perhaps. IANAL.
It’s nearly impossible to prove a negative. But it would be trivial to ask it for code from a GPL project. If it can provide it, it’s in the model.
In a general, open-ended situation, sure. In this case, the model developer has all of the training data that was input for training the model. It should take all of half an hour to determine if there is any GPL code in the training data set.
Hi, I am from OpenAI/Alphabet/Meta, I have checked our training data and it contains no GPL code, or copyrighted or pirated media of any kind, it’s all perfectly totally 100% legit. Thank you for your inquiry, hope this solves the issue and have a good day.
Right, so, then we have a paper trail of fraud. Forcing them to make such a statement in writing is useful.
This guy gets it.
Social media users want to see a person walk into the bad guy’s office and punch them in the face, solving the problem. Real effective steps are usually boring and unremarkable like forcing them to make a statement on record.
Exactly, this is how you win court battles that set legal precedents which establish real-world consequences, not as one-offs but every time somebody crosses that line.
Damn that’s a great framing. Will use this talking point going forward.