Hacker News Reader

Comment by Someone1234 | Hacker News Reader

I don't understand this...

If I open Photoshop then via my inputs it indirectly generates pixels using various tools/brushes/automations, it is copyrightable.

If I open an LLM, enter a bespoke prompt, then the tool generates pixels using that prompt, it isn't copyrightable.

Things can be copyrightable even if there isn't a 1:1 between human input and copyright output. Where is the cut-off exactly? What about Photoshop's Content-Aware Fill, is that copyright or not?

What if I take LLM output then color correct it, is that now my copyright? Since it effectively has no authorship before I color shifted it.

View on HN

> Where is the cut-off exactly?

This is the difference between legal thinking and CS. There is a spectrum between 100% human-generated (and copyright-eligible) and 100% machine-generated (and ineligible), and in between there are gray areas.

To us, that is a problem, because it implies ambiguity. Where is the cut-off?

In law, that is normal. No actions are the same; everything is on a spectrum from allowed to illegal. Judging the boundary is made on a case-by-case basis through legal argument, evaluated by a judge or jury.

So to the Copyright Office, it is sufficient to say that this particular way of AI generating art doesn't involve enough human creativity. But you won't get them to define enough in the abstract, for any general case, because that is essentially impossible, and can only be done by evaluating concrete cases.

Sorry, not everything is evaluated on a "case-by-case" basis per say, there is a difference between a "Bright-Line" (clearly specified rule in court) and a "Balancing-Test" (what you describe).

I think it's actually pretty damn easy to define "enough" precisely. If the AI + human generated image cannot be easily recreated by AI, than it's enough.

So, for example, random image from civit.ai which is nothing more than a prompt, and negative prompts. This fails the test because I can run the image through CLIP and get (almost) the same exact prompts used to generate it

Now let's say I generate an image using a bunch of custom tools and extensions, such as regional color-loss (i.e. top right of the image must be blue and I will force the model to prefer blue outputs), and leveraging a lora model at 0.3 weight trained on my kids sketch-book drawing style. There's no way for any AI model that hasn't been configured exactly by my human hands to deduce anything approaching the prompt used to generate it (CLIP/Other LLMs doesn't know about extensions and loras).

Courts need a bright-line here, balancing tests lead to miscarriages of justice, which is what looks like is going to happen here a lot.

There are fairly few bright-line tests, because they're hard to write.

For instance, consider your proposed test: The copyrightability of works would change retroactively depending on the capabilities of AI. Something that can't be reproduced by AI today could be next year. Is that a desirable outcome?

>I think it's actually pretty damn easy to define "enough" precisely. If the AI + human generated image cannot be easily recreated by AI, than it's enough.

That's a terrible test, because the capabilities of LLMs are constantly evolving. And the words "easily" and "recreate" just creates more ambiguity.

I struggle to come up with a better definition of “enough” than his. It could be clarified to “the original AI.”

> I think it's actually pretty damn easy to define "enough" precisely. If the AI + human generated image cannot be easily recreated by AI, than it's enough.

Would you be on-board with the inverse of this?

You cannot copyright what cannot be generated by a human.

Very eloquently put.

I often find it interesting how when these types of issues come up, technologists often interpret it using "transitive rules" that are natural in math and engineering, but "the real world" doesn't follow this logical progression. That is, if you have "A implies B, and B implies C, and C implies D, then A implies D". The problem is that in "the real world" there are enough slight differences at each step that most people feel that A does not imply D.

Point being, there are lots of comments here along the lines of "If creating art with a tool can be copyrighted, how is art created with Photoshop copyrightable but art created with AI is not". The difference is that most people (and, most importantly, the US Copyright Office) believe that AI crosses the line between being "just a tool" to actually being the author of significant parts of the work. The fact that that line is gray is not a fundamental problem, and as you point out it occurs all the time in the legal realm.

> That is, if you have "A implies B, and B implies C, and C implies D, then A implies D". The problem is that in "the real world" there are enough slight differences at each step that most people feel that A does not imply D.

So you just got it wrong and A does not imply D. Obviously logic applies to the world outside math and computer science.

This is by no means settled legally either.

> In law, that is normal.

Normal, sure, but still a problem in the sense that they need to clarify it and try not to write ambiguous laws. The existence of the law, lawyers, and the whole judicial and lawmaking systems is proof that its difficult and complicated.

I just dont see the contrast with computer science here.

And that's the problem with the law.

This seems like a misunderstanding of the argument. It has nothing to do with inputs, it has to do with the creative process of making of content. It doesn't have to with the method but the creative input and making of content.

Example used above: I tell an artist to make me a picture of a dog. They created the image, and they own the copyright. They would have to give the copyright to me, I do not own it merely because I made the request.

If you use photoshop, we can identify the source of every pixel. It was you.

If you use a generative tool, we can no longer identify the source of every pixel, or if it derives from someone else's copyrighted work.

You can claim rights over the first, you cannot over the second. This line is pretty bright and identifiable.

No. You don't know who wrote the software that implemented that algorithm. In effect, me as the engineer who created photoshop was the creator of that. You needed me in this example to do that. You don't know if ai created that algorithm - it has been happening for decades that some software was created by other software - and I'm not just talking about optimizing compilers.

There was a creative algorithm, and 99.999% of the time there was not a specification that determines every single bit.

The work output of the programmers, the thousands of people involved in that pipeline to create the software - "that's different", say the artists.

You also don't know who created the paint or what their secret formula is in traditional art.

Drawing your art in Photoshop is clearly authorship. You are manipulating each pixel from scratch to be how you want it. There may be algorithms between your mouse clicks and the final product, but you as the author still have complete control of the whole art. You can tweak each pixel until its exactly how you want.

Asking an LLM for a picture of a duck is clearly not authorship. You did not create any pixels in that image, the LLM did. You just asked for it. Its no different than googling for a picture of a duck and claiming what you find because you engineered the google search query.

There is a lot of murkiness in between, where you take assets from an LLM and use them in your own creation. I think having to disclaim the assets which were not created by you is a very reasonable solution.

so if i tell an artist to do some revisions i own that work now? what exactly did the human do in this case? wrote some words in a certain order? so the prompt is copy writable then? if that's the case can i just copywrite certain prompts that are commonly used?

It seems highly likely that the prompt itself may be copyrighted, yes

If a prompt is commonly used, then it wasn't created by you, and isn't copyrightable. It has to have enough original content so it could be copyrighted. I am sure a short poem is copyrightable.

But then again, fair use also allows citing most (if not all) of the short poem.

The difference is the underlying source material. With photoshop it comes from your brain, with AI tools like Midjourney it is derived from other peoples brains.

Reply on news.ycombinator.com