[flagged]
If I had a photographic memory and I used it to replicate parts of GPLed software verbatim while erasing the license, I could not excuse it in court that I simply "learned from" the examples.
Some companies outright bar their employees from reading GPLed code because they see it as too high of a liability. But if a computer does it, then suddenly it is a-ok. Apparently according to the courts too.
If you're going to allow copyright laundering, at least allow it for both humans and computers. It's only fair.
Right, because you would have done more than learning, you would have then gone past learning and used that learning to reproduce the work.
It works exactly the same for a LLM. Training the model on content you have legal access to is fine. Aftwards, somone using that model to produce a replica of that content is engaged in copyright enfringement.
You seem set on conflating the act of learning with the act of reproduction. You are allowed to learn from copyrighted works you have legal access to, you just aren't allowed to duplicate those works.
If someone hires me to write some code, and I give them GPLed code (without telling them it is GPLed), I'm the one who broke the license, not them.
I don't think this is legally true. The law isn't fully settled here, but things seem to be moving towards the LLM user being the holder of the copyright of any work produced by that user prompting the LLM. It seems like this would also place the enfringement onus on the user, not the provider.
> If someone hires me to write some code, and I give them GPLed code (without telling them it is GPLed), I'm the one who broke the license, not them.
If you produce code using a LLM, you (probably) own the copyright. If that code is already GPL'd, you would be the one engaged in enfringement.
"Learning" is an established word for this, happy to stick with "training" if that helps your comprehension.
> LLMs don't "learn" but they _do_ in some cases, faithfully regurgitate what they have been trained on.
> Legally, we call that "making a copy."
Yes, when you use a LLM to make a copy .. that is making a copy.
When you train a LLM... That isn't making a copy, that is training. No copy is created until output is generated that contains a copy.
If we are saying that the LLM learns things and then made the copy, then the LLM made the crime and should receive the legal punishment and be sent to jail, banning it from society until it is deemed safe to return. It is not like the installed copy is some child spawn from digital DNA and thus the parent continue to roam while the child get sent to jail. If we are to treat it like a living being that learns things, then every copy and every version is part of the same individual and thus the whole individual get sent to jail. No copy is created when installed on a new device.
Right, because then we have to decide at what point our use of AI becomes slavery.
It's also an interesting double standard, wherein if I were to steal OpenAI's models, no AI worshippers would have any issue condemning my action, but when a large company clearly violates the license terms of free software, you give them a pass.
If GPT-5 were "open sourced", I don't think the vast majority of AI users would seriously object.
Which is funny since that's a much clearer case of "learning from" than outright compressing all open source code into a giant pile of weights by learning a low-dimensional probability distribution of token sequences.
Information wants to be free.
That is not nearly the extent of AI training data (e.g. OpenAI training its image models on Studio Ghibli art). But if by "gave their work away for free" you mean "allowed others to make [proprietary] derivative works", then that is in many cases simply not true (e.g. GPL software, or artists who publish work protected by copyright).