Isn’t AI and already rendering those licenses moot? AI data sets don’t care what type of license the code it scraped caries. and probably most AI code being generated for private companies is based on non-permissive licenses. yet nothing can be done about it.
to look at it another way: since anything an llm spits out is public domain, whoever commanded it do so is automatically in breach of the gpl unless they can show evidence that there is no copyleft material in the training data at all.
I hate seeing this kind of doomerism around AI discussions. A future where we have no rights is not inevitable.
We know for a fact that OpenAI, Anthropic, Google etc trained on GPL code. They haven’t claimed otherwise, and this isn’t something they’re trying to hide.
We also know what the various open source models were trained on, since they’re open source.
We can pass laws that require LLM providers to disclose their data set in some way, eg through independent audits. Trump is the one who didn’t want AI regulation, and he basically just lost all ability to make that a reality. The final six weeks of the pedophile congress isn’t likely to accomplish anything on that front, and even if they do, the incoming congress can reverse it.
This is the time to energize people, not complain on the internet.
I recall articles saying that significant input/prompt may influence that. Whether this is true or not I can’t say.
That’s not true. Until these issues have run through the courts, we won’t know what is or isn’t legal. The biggest court case I’m familiar with atm is the New York Times’ case that has yet to be resolved, and even then the ruling there will probably be too narrow to decide anything for code licensing. Hell, OpenAI/Microsoft might even settle to avoid a bad ruling, leaving all those questions unanswered.
So the Tl;Dr is that anyone using LLMs to code (in any capacity) is taking a risk, since the legality of it is still uncertain. It’s not outside the realm of possibility that some court will rule that LLMs trained on stolen data are illegal, and the outputs of those models are infringing.
If you can compare the code and it matches you probably have a good case if the AI user can’t demonstrate it’s not coming from you.
There’s different kinds of datasets. Some respect, select and disclose licenses, others do not.
Some model trainers use any data they can, others are selective on what they train on.
It can also be a decision on principle rather than current or future enforcability. If you say “don’t copy this” and someone blatantly does, the morality is clearer than if you didn’t specifically say that.
There’s also a question of how court rulings and law may change the situation of enforcability and usage/ignorance risk.
MIT requires attribution too. GPL is, at the very least, a stronger statement of intention and demand. It can also be seen as protest or indication of a broader move of corporate AI exploitation leading to damage to the whole FOSS ecosystem, with additional restrictions being applied. Corps that would use AI now can’t use the MIT product anymore like before.
Not being able to compare some proprietary is besides my point. Not all code is or remains hidden. I specifically wrote “if you can match” to remain neutral on that aspect.
If you can compare the code and it matches you probably have a good case if the AI user can’t demonstrate it’s not coming from you.
It’s not necessarily that simple.
Copyright protects creative expression, not underlying ideas or concepts or functionality.
The famous examples are that a phone book or map get limited copyright protection, because the text or image on the page corresponds to a ground truth that can’t be copyrighted. Recipe books also have to deal with that legal reality.
In software, functionality, including documented APIs, are not copyrightable in themselves. There was a phenomenon where AT&T asserted copyright over their Unix executable /bin/true, despite that script doing nothing except exit successfully, but there’s really nothing stopping anyone else from doing the same thing with a file with identical non-comment contents. So copyright doesn’t protect bare functionality.
And the easiest way to prove that you didn’t copy any nonfunctional expression is to have someone take the documentation and reimplement that functionality without ever looking at the original code, aka a “clean room implementation.” Plenty of examples of that happening in the past, surviving challenges in court: Compaq reverse engineering IBM’s BIOS and kicking off the IBM clone PC revolution, Google reimplementing the Java APIs from documentation to free themselves from Oracle’s oppressive contracts in Android, etc.
And because copyLEFT licenses rely on the force of copyright law to work, these types of areas where code doesn’t get copyright protection are also the areas where FOSS code wouldn’t get copyleft protections, either.
Why would a AI user copy the code though? They can just launder it through intelligence (brain and artificial both work) and the code will look different, but do the same.
For actual copying they wouldn’t need AI. They would just need “git clone”.
Isn’t AI and already rendering those licenses moot? AI data sets don’t care what type of license the code it scraped caries. and probably most AI code being generated for private companies is based on non-permissive licenses. yet nothing can be done about it.
I think the answer is “yes”.
This feels like a token response to a systemic unsolvable problem.
The article says itself talks about having models launder code so it can’t be considered a copy of the source.
to look at it another way: since anything an llm spits out is public domain, whoever commanded it do so is automatically in breach of the gpl unless they can show evidence that there is no copyleft material in the training data at all.
Afaik that’s not exactly so. I recall articles saying that significant input/prompt may influence that. Whether this is true or not I can’t say.
can’t be. if there is gpled code in the training data, all code must be gpl since the model is one work. and since it can’t, that’s a breach.
Yeah, in theory. In practice, good luck proving it.
I hate seeing this kind of doomerism around AI discussions. A future where we have no rights is not inevitable.
We know for a fact that OpenAI, Anthropic, Google etc trained on GPL code. They haven’t claimed otherwise, and this isn’t something they’re trying to hide.
We also know what the various open source models were trained on, since they’re open source.
We can pass laws that require LLM providers to disclose their data set in some way, eg through independent audits. Trump is the one who didn’t want AI regulation, and he basically just lost all ability to make that a reality. The final six weeks of the pedophile congress isn’t likely to accomplish anything on that front, and even if they do, the incoming congress can reverse it.
This is the time to energize people, not complain on the internet.
the question is if that theory is enough for a general ruling.
That’s not true. Until these issues have run through the courts, we won’t know what is or isn’t legal. The biggest court case I’m familiar with atm is the New York Times’ case that has yet to be resolved, and even then the ruling there will probably be too narrow to decide anything for code licensing. Hell, OpenAI/Microsoft might even settle to avoid a bad ruling, leaving all those questions unanswered.
So the Tl;Dr is that anyone using LLMs to code (in any capacity) is taking a risk, since the legality of it is still uncertain. It’s not outside the realm of possibility that some court will rule that LLMs trained on stolen data are illegal, and the outputs of those models are infringing.
If you can compare the code and it matches you probably have a good case if the AI user can’t demonstrate it’s not coming from you.
There’s different kinds of datasets. Some respect, select and disclose licenses, others do not.
Some model trainers use any data they can, others are selective on what they train on.
It can also be a decision on principle rather than current or future enforcability. If you say “don’t copy this” and someone blatantly does, the morality is clearer than if you didn’t specifically say that.
There’s also a question of how court rulings and law may change the situation of enforcability and usage/ignorance risk.
MIT requires attribution too. GPL is, at the very least, a stronger statement of intention and demand. It can also be seen as protest or indication of a broader move of corporate AI exploitation leading to damage to the whole FOSS ecosystem, with additional restrictions being applied. Corps that would use AI now can’t use the MIT product anymore like before.
First, you can’t even compare the code if you don’t have access to it, you don’t have access to google’s or facebook’s code.
Second who can even legally challenge entities with pockets as deep as the ocean.
Third courts will probably classify AI generated code based of non-permissive licenses as fair use, so It’s a lost cause anyway.
Nothing is set. Until there’s a challenge in court, all you stated as certainties are just conjectures.
Not being able to compare some proprietary is besides my point. Not all code is or remains hidden. I specifically wrote “if you can match” to remain neutral on that aspect.
It’s not necessarily that simple.
Copyright protects creative expression, not underlying ideas or concepts or functionality.
The famous examples are that a phone book or map get limited copyright protection, because the text or image on the page corresponds to a ground truth that can’t be copyrighted. Recipe books also have to deal with that legal reality.
In software, functionality, including documented APIs, are not copyrightable in themselves. There was a phenomenon where AT&T asserted copyright over their Unix executable /bin/true, despite that script doing nothing except exit successfully, but there’s really nothing stopping anyone else from doing the same thing with a file with identical non-comment contents. So copyright doesn’t protect bare functionality.
And the easiest way to prove that you didn’t copy any nonfunctional expression is to have someone take the documentation and reimplement that functionality without ever looking at the original code, aka a “clean room implementation.” Plenty of examples of that happening in the past, surviving challenges in court: Compaq reverse engineering IBM’s BIOS and kicking off the IBM clone PC revolution, Google reimplementing the Java APIs from documentation to free themselves from Oracle’s oppressive contracts in Android, etc.
And because copyLEFT licenses rely on the force of copyright law to work, these types of areas where code doesn’t get copyright protection are also the areas where FOSS code wouldn’t get copyleft protections, either.
Why would it match? There’s thousands of ways to program any feature, the chance that it would be the same is tiny.
Copied code often matches, often look the same.
Why would a AI user copy the code though? They can just launder it through intelligence (brain and artificial both work) and the code will look different, but do the same.
For actual copying they wouldn’t need AI. They would just need “git clone”.