Comment by carom | Hacker News Reader

Comment by carom | Hacker News Reader

Parent Full thread

There are three copyright issues here; datasets, model weights, and model outputs.

Dataset copyright is pretty well defined and things can often be used under fair use. Fair use decisions are done with a four prong test and really decided by the courts on a case-by-case basis.

Model weights cannot currently be copyrighted. They are the output of a mechanical process over the dataset. However, software faced a similar situation where the source code could be copyrighted but the compiled binary was not. US copyright law was updated to address this. We may see something similar for model weights.

Model outputs are less clear, but these are likely copyrightable by the user of the model. It is not possible for a non-human to hold copyright, so the model cannot. It is very unlikely that the company producing the model could assert copyright over the outputs. A good analogy here is someone using photo manipulation software.

Super interesting area. I think we will eventually see an update to the copyright code to make weights copyrightable. Also it will be interesting to see how court challenges (code generation, image generation) affect datasets in the future.

Who do you think should hold the copyright on the model weights, the copyright holders of the individual works comprising the dataset or the ones who assembled the dataset?

I don't think that model weights are copyrightable.

They're a mathematical transformation of the source material.

As a trivial example, applying gray = .299 r + .587 g + .114 b to a pixel is also a mathematical transformation but it doesn't create a new copyright.

And thus, the model is a mathematical transformation and a derivative work of its source material.

However I am also of the opinion that the resulting model is sufficiently transformative and fills a different purpose than the source material and so the model is not infringing.

From the Perfect 10 case against Google and Amazon:

> …We conclude that the significantly transformative nature of Google's search engine, particularly in light of its public benefit, outweighs Google's superseding and commercial uses of the thumbnails in this case. … We are also mindful of the Supreme Court's direction that "the more transformative the new work, the less will be the significance of other factors, like commercialism, that may weigh against a finding of fair use."

That was dealing with thumbnails being used in search and it was found that thumbnails being used for search was sufficiently transformative and applied to a different purpose that Google's use didn't infringe.

I believe that a model is even more transformative given that bar.

That doesn't mean that the output of the model isn't infringing, but that's a human with agency creating and publishing that output which is a different artifact to be considered than the model weights.

Indeed. Inputs are likely fair use but if you output Mickey Mouse, Disney will definitely he on you. I think that's the most sensible approach, as anyone can use a tool like a pencil to draw anything they want, doesn't mean that they'll get away with creating a drawing of Mickey and saying they own the copyright.

Depends who you think holds the copyright on the SHA256 hash of an image.

Model weights are even less specific then that number, since they don't represent any specific source input at all.

Biggest issue is the first one. Dataset copyrights are where most of the fight b/w regulators, AI companies & artists is happening.

Should you be able to train your model on an image or text someone uploaded on the internet without buying the copyright? If no, then most of current LLMs & stable diffusion models will have to go & I don't see big tech allowing it.

Realistically an AI model is basically just a very complicated piece of software. The model weights are akin to the software code, the model outputs are akin to the outputs a user of the software creates, and the datasets are akin to the intellectual property put into the software by the developer to create the code.

In the same way that a developer could not simply steal someone elses intellectual property in order to develop a feature of a piece of software, one cannot simply steal the intellectual property to adjust the model weights. The main difference is its generally quite easy to see in practice if a model has utilized some intellectual property (because for example you can ask ChatGPT to recite the first 100 words of Harry Potter) compared to another piece of software where you'd need access to the source code or developers thoughts (which could only be achieved through litigation, in most circumstances).

I think a great many people come up with convoluted answers to this question because they are uncomfortable with the reality that these very large organizations have essentially stolen hoards of intellectual property, and now that the horse has bolted people want to justify not closing the barn door. It seems to me very simple: to train an AI model on data, you must respect its copyright. The model weights should be copywriteable by the developers of the model (even if the law currently does not allow this), and the outputs of the model should be copywriteable by the person who interacted with the model (software) to produce the outputs.

The analogy with Photoshop is extremely simple: If some other software invented Gaussian blurring and copywrited it, then Adobe would have to license that technology from them to include it as a feature in Photoshop. The actual photoshop software/code would be copywrited by Adobe, and if someone created an blurry image with Photoshop they can copywrite it.

I think people only disagree with this due to some sense that the process of translating data to model weights is "automatic" or "computational" in nature. You could in principle get a person to, by hand, go through millions of data sets and compute the changes to the model weights. This is no different to someone writing a piece of code, checking someone elses approach, and adjusting their own code after the fact. It just happens that we have developed very effective tooling to automate the adjusting of the code.

Three points I want to make:

- Models are nothing more than a statistical distillation of facts that can be traversed. They are not like software at all. Calling them software is like calling pachinko machines software. Nonsense.

- Models are mechanically derived with no element of human authorship or creativity. You could argue that there is creativity in selecting the dataset or the process that derives the model, but neither is relevant to the final generated model. Even if we assumed for the sake of argument that a model is more that a just statistical distillation, it should still not be considered copyrightable due to this reason alone.

- Don't use the word Steal when you refer to the well-defined act of infringement. Stealing implies deprivation of property which does not and cannot occur in this case. Using the word Infringe is more honest and less manipulative.

It's not stealing, and the term "intellectual property" should be put to rest:

https://www.gnu.org/philosophy/not-ipr.en.html

Your opinions on what should be copyrightable are wrong, and fortunately the courts agree.

You've rather conspicuously failed to mention a fourth, with two parts, and the first listed in the article abstract: "the use of copyrighted works to train AI models, the appropriate levels of transparency and disclosure with respect to the use of copyrighted works".

Outputs are the third mentioned: "the legal status of AI-generated outputs."

Reply on news.ycombinator.com