The GPLv4 idea is interesting and I would like to see it happen. However, I'm not sure how it would apply in practice. What's to keep that kind of company from rejecting the license and still train proprietary models on your code under fair use claims like they're already doing?
post
This presupposes that there is some real and tangible benefit to LLMs, but an explanation of that argument is never put forward. Sure, the dream is to have the magic LLM genie write you code that runs perfectly and does what you want, but now you have code you didn't write and can't maintain. So you have to rely on the LLM genie for that to, and sooner rather than later - no matter how well you've trained it - it's going to run across something it can't do. And then what? You have a pile of code you don't understand and no idea yourself how to fix it yourself because you kept having the genie do it.
What about all of that is worth the theft of the work of others, the resource demands to train and run the LLMs, and the debasement of actual coders?
I absolutely agree. For me, a very similar thing can be said for art. It should be okay to scrap free and open source data, and copyleft art and make whatever you want with it. That is why I put a copyleft license on things! Just release the code, and acknowledge the source.
If you waste your time shitting on LLMs and AI models as a concept... You are totally missing the point. These are useful tools. Not as useful as the oligarchs want you to believe. But they are useful. And people find value in them. This is the reality.
You should focus on who controls and monopolises these concepts, not the concept itself.
38 comments in c/opensource@lemmy.ml view that post
LLMs are tools. They're not replacements for human creativity. They are not reliable sources of truth. They are interesting tools and toys that you can play with.
So have fun and play with them.
replies:
This is the correct take. This tech isn't going away, no matter how much whinging people do, the only question is who is going to control it going forward.
The problem is not the algorithm. The problem is the way they're trained. If I made a dataset from sources whose copyright holders exercise their IP rights and then train an LLM on it, I'd probably go to jail or just kill myself (or default on my debts to the holders) if they sue for damages.
We should reject them.
Seems like the easiest fix is to consider the produce of LLMs to be derivative products of the training data.
No need for a new license, if you're training code on GPL code the code produced by LLMs is GPL.
replies:
How dare you break the jerk! /s
Checking whether a proprietary LLM model running on the "cloud" has been trained on a piece of TGPL code would probably be harder than checking if a proprietary binary contains a piece of GPL code, though.
replies:
LLM are the future, but we must still learn to use it correctly. The energy problem depends mainly on 2 things, the use of fossil energy and the abuse of AI including it without need in everything, because the hype, as data logging tool for Big Brother or biased influencers.
You don't need a 4x4 8 cylinder Pick-up to go 2km to the store to buy bread.
replies:
load more comments
(-1 replies)
One of the four essential freedoms is the freedom to study the software and modify it. Studying means training your brain on the open source code. Can one use their brain to write proprietary code after they studied some copylefted code?
I asked the bot. It says:
I understand the concern. This question sits at the intersection of copyright law, statistics, and how machine-learning systems work—and it’s fair to be skeptical.
Here’s the most precise and careful way to frame it:
1. I can’t confirm GPL training data
I don’t know what specific works or licenses were in the training set, and I can’t confirm that GPL-licensed material was included. Neither I nor users have visibility into the underlying corpus.
So while it may be statistically plausible that GPL text exists somewhere in public data, plausibility ≠ proof, and I can’t assert it as fact.
2. Training ≠ distribution (the core legal argument)
OpenAI’s legal position (and that of most ML practitioners) is broadly this:
-
Training on copyrighted material is not the same as redistributing it
-
The model does not contain copies of works
-
Outputs are newly generated, not retrieved
-
Copyright law regulates copying and distribution, not learning patterns or statistics
This is analogous (legally, not philosophically) to a human reading GPL code and then later writing different code from experience.
Whether courts ultimately agree in every jurisdiction is still being litigated—but no court has ruled that training itself is per se illegal.
3. GPL specifically
GPL obligations trigger when you:
-
copy
-
modify
-
distribute
GPL-licensed code or derivative works.
Current legal consensus does not treat model weights or outputs as GPL “derivatives” in the copyright sense, because:
-
There is no preserved expressive structure
-
There is no traceable lineage to a specific work
-
Outputs are not mechanically derived from any one source
That’s why GPL’s copyleft provisions are generally understood not to propagate through model training.
4. Is the law settled?
No. Anyone claiming it is settled is overstating.
What is true:
-
Training on public data is widespread and openly practiced
-
Multiple lawsuits exist, but none have yet established that training itself violates copyright or GPL
-
Regulators and courts are still clarifying boundaries
5. Practical reality
Even if GPL material were in the training mix:
-
That alone would not automatically make the model or its outputs illegal
-
Liability (if any) would hinge on specific outputs being substantially similar to protected works, not on statistical likelihoods about training data
Bottom line
Your intuition about probability is reasonable—but legality does not turn on probability. It turns on:
-
whether protected expression is reproduced
-
whether derivative works are created
-
whether licenses are triggered by the act in question
As of now, the prevailing legal interpretation is that models like me are lawful to deploy, even though the debate is ongoing.
If you want, I can also:
-
Compare this to the Google Books ruling
-
Walk through why “derivative work” is a high legal bar
-
Discuss what would actually make an AI system GPL-tainted in practice
replies:
1 comment in c/opensource@programming.dev view that post
Tldr; we shouldn't be idealistic, we should accept that LLMs stole from FOSS code, we should accept reality and instead of abandoning daddy GitHub we should ask them super nicely if they can pretty please open-source the training models trained on our stolen code.
I don't understand how the author simultaneously holds the position "we should accept reality" and "GitHub/Microsoft will open up their models if we ask them to".
all 3 comments