Comment by II2II | Hacker News Reader

Comment by II2II | Hacker News Reader

Parent Full thread

> Use of the output of systems like Copilot or Stable Diffusion becomes a violation of copyright.

That really should depend upon the output.

Many, if not most, people learn an art by imitating the style of established artists. Some will carry on with that style. Others will develop their own, though it will probably always carry elements from those they imitated. Should injecting a machine into the process automatically make it illegal?

There are going to be clear cut cases where it should be, cases where so much is imitated that it goes beyond style and into substance. Yet that means we should have a human looking at the output to determine if it is too close to a copy, rather than banning AI generated art altogether. To do so would put the creative process in peril. This is not because machine learning reflects our definition of creativity. Rather it is because it is difficult to define what human creativity itself is.

(That said, I do believe that using the artist's name as a way of promoting their own work is stepping over a line.)

> That really should depend upon the output.

Except, in copyright law it depends on the _input_.

These models would not exist if they were not first fed the source material.

Until we have systems that are not trained on a pre-existing corpus this will remain true. No matter how clever the algorithm, without the source material you have no output. Zilch. Nada.

Now, when the source material is someone else's property this means that without - someone else's property - you would have had no output.

So, when you want to use someone else's property, which you do not own, the general rule is that you a) first ask them if you may and b) pay them for the right to use their property.

In this sense it is no different than using a photocopier.

It's the copyright ownership of the material you put into the machine that will interest the judge not the quality of the copy.

I'm really looking forward to the first court cases and predict that much hilarity will ensue!

Trained models don't have the actual images inside, they have summed up gradients. So what they are doing is far from a copy&paste job, it's more like decomposing and recomposing from basic concepts, something made clear by the "variations" mode.

Among the things the model learned are some un-copyrightable facts, such as the shapes of various objects and animals, how they relate, their colours and textures - general knowledge for us, humans. Learning this is OK because you can copyright the expression, but not the idea.

Trained models take little from each example they learn. The original model shrunk 4B images to a mere 4GB file, so 1 byte/image worth of information learned from each example, a measly pixel. The DreamBooth finetuning process only uses 20-30 images from the artist, it's more like pinning the desired style than learning to copy. Without Dreambooth its harder but not impossible to find and use a specific style.

And the new images are different, combining elements from the prompt, named artists and general world knowledge inside. Can we restrict new things - not copies - from being created, except in patents? Isn't such an open ended restriction a power grab? To make an analogy: can a writer copyright a style of writing, and anything that has a similar style be banned?

A digital photo of the Eiffel tower at night doesn't have the real Eiffel tower inside, only weights and pixels - still you don't have the rights to publish your photo of the Eiffel tower in France.

Drawings and art including the Eiffel tower are still ok, right?

> Trained models don't have the actual images inside, they have summed up gradients. So what they are doing is far from a copy&paste job, it's more like decomposing and recomposing from basic concepts, something made clear by the "variations" mode.

Doesn't matter. JPEG of the work is just a bit of numbers to feed equation, doesn't change the fact it's copyright infringement

>Now, when the source material is someone else's property this means that without - someone else's property - you would have had no output.

Exactly the same happens with artists. The only artists who can claim not to have been influenced by seeing the work of other artists lived tens of thousands of years ago. So what makes it okay to process artwork via some processes but not others, when the ultimate output may in some way copy the input anyway?

>Except, in copyright law it depends on the _input_

I am not a lawyer, but my understanding of copyright law is that this is explicitly the opposite.

https://www.wipo.int/edocs/pubdocs/en/wipo_pub_909_2016.pdf

>• reproduction of the work in various forms, such as printed publications or sound recordings; >• distribution of copies of the work; >• public performance of the work; >• broadcasting or other communication of the work to the public; >• translation of the work into other languages; and >• adaptation of the work, such as turning a novel into a screenplay

None of these rights, to me, indicate that copyright protects the input. The AI model is not reproducing any specific works, distributing copies of it, performing it in public, broadcasting it, translating it to another language, or adapting the work from one format to another.

Reply on news.ycombinator.com