Item — HN Reader

One of these things is true:

1. Training AI on copyrighted works is fair use, so it's allowed no matter what the license says.

2. Training AI on copyrighted works is not fair use, so since pretty much every open source license requires attribution (even ones as lax as MIT do; it's only ones that are pretty much PD-equivalent like CC0, WTFPL, and Unlicense that don't) and AI doesn't give attribution, it's already disallowed by all of them.

So in either case, having a license mention AI explicitly wouldn't do any good, and would only make the license fail to comply with the OSD.

1 comments

Point 2 misses the distinction between AI models and their outputs.

Let's assume for a moment that training AI (or, in other words, creating an AI model) is not fair use. That means that all of the license restrictions must be adhered to.

For the MIT license, the requirement is to include the copyright notice and permission notice "in all copies or substantial portions of the Software". If we're going to argue that the model is a substantial portion of the software, then only the model would need to carry the notices. And we've already settled on accessing over a server doesn't trigger these clauses.

Something like the AGPL is more interesting. Again, if we accept that the model is a derivative work of the content it was trained on, then the AGPL's viral nature would require that the model be released under an appropriate license. However, it still says nothing about the output. In fact, the GPL family licenses don't require the output of software under one of those licenses to be open, so I suspect that would also be true for content.

So far, though, in the US, it seems courts are beginning to recognize AI model training as fair use. Honestly, I'm not surprised, given that it was seen as fair use to build a searchable database of copyright-protected text. The AI model is an even more transformative use, since (from my understanding) you can't reverse engineer the training data out of a model.

But there is still the ethical question of disclosing the training material. Plagiarism still exists, even for content in the public domain. So attributing the complete set of training material would probably fall into this form of ethical question, rather than the legal questions around intellectual property and licensing agreements. How you go about obtaining the training material is also a relevant discussion, since even fair use doesn't allow you to pirate material, and you must still legally obtain it - fair use only allows you to use it once you've obtained it.

There are still questions for output, but those are, in my opinion, less interesting. If you have a searchable copy of your training material, you can do a fuzzy search of that material to return potential cases where the model returned something close to the original content. GitHub already does something similar with GitHub Copilot and finding public code that matches AI responses, but there are still questions there, too. It's more around matches that may not be in the training data or how much duplicated code needs to be attributed. But once you find the original content, working with licensing becomes easier. There are also questions about guardrails and how much is necessary to prevent exact reproduction of copyright protected material that, even if licensed for training, isn't licensed for redistribution.

> The AI model is an even more transformative use, since (from my understanding) you can't reverse engineer the training data out of a model.

You absolutely can; the model is quite capable of reproducing works it was trained on, if not perfectly then at least close enough to infringe copyright. The only thing stopping it from doing so is filters put in place by services to attempt to dodge the question.

> In fact, the GPL family licenses don't require the output of software under one of those licenses to be open, so I suspect that would also be true for content.

It does if the software copies portions of itself into the output, which seems close enough to what LLMs do. The neuron weights are essentially derived from all the training data.

> There are also questions about guardrails and how much is necessary to prevent exact reproduction of copyright protected material that, even if licensed for training, isn't licensed for redistribution.

That's not something you can handle via guardrails. If you read a piece of code, and then produce something substantially similar in expression (not just in algorithm and comparable functional details), you've still created a derivative work. There is no well-defined threshold for "how similar", the fundamental question is whether you derived from the other code or not.

The only way to not violate the license on the training data is to treat all output as potentially derived from all training data.

> You absolutely can; the model is quite capable of reproducing works it was trained on, if not perfectly then at least close enough to infringe copyright. The only thing stopping it from doing so is filters put in place by services to attempt to dodge the question.

The model doesn't reproduce anything. It's a mathematical representation of the training data. Software that uses the model generates the output. The same model can be used across multiple software applications for different purposes. If I were to go to https://huggingface.co/deepseek-ai/DeepSeek-V3.2/tree/main (for example) and download those files, I wouldn't be able to reverse-engineer the training data without building more software.

Compare that to a search database, which needs the full text in an indexable format, directly associated with the document it came from. Although you can encrypt the database, at some point, it needs to have the text mapped to documents, which would make it much easier to reconstruct the complete original documents.

> That's not something you can handle via guardrails. If you read a piece of code, and then produce something substantially similar in expression (not just in algorithm and comparable functional details), you've still created a derivative work. There is no well-defined threshold for "how similar", the fundamental question is whether you derived from the other code or not.

The threshold of originality defines whether something can be protected by copyright. There are plenty of small snippets of code that can't be protected. But there are still questions about these small snippets that were consumed in the context of a larger, protected work, especially when there are only so many ways to express the same concept in a given language. It's definitely easier in written text than code to reason about.

> The model doesn't reproduce anything. It's a mathematical representation of the training data. Software that uses the model generates the output.

By that argument, a compressed copy of the Internet doesn't reproduce the Internet, the decompression software does. That's not a useful semantic distinction; the compressed file is the derived work, not the decompression software.