Hacker News Reader

Comment by pcthrowaway | Hacker News Reader

She wasn't just prompting the AI. That comic required significant work on the user's part to edit the images and re-synthesize portions of each frame repeatedly while doing things like inpainting, outpainting, re-prompting, etc. And likely required use of some other tools like Photoshop or Indesign (midjourney absolutely did not generate the bubbles or comic book text).

View on HN

I’m guessing that the copyright equivalent is whether the editor of someone else’s photograph or the director of a photo shoot should have copyright privileges. The answer is pretty clearly not.

It’s not that she didn’t do (a lot) of work, but that the creation of the underlying work happened by AI. If it were human then it would hold the copyright, but it’s not so it doesn’t. The logical misstep is thinking, “because it can’t have copyright, someone else can.”

But then by this logic photos shouldn’t have copyright either because a camera can’t hold copyright

I think prompting the AI is like taking a photo in that you set the parameters and subject and press the button to get an image

Exactly, generative AI, like the camera, is just a fancy way of transforming ideas from the real-world into a digital representation, but with a lot more parameters.

> That comic required significant work on the user's part to edit the images and re-synthesize portions of each frame repeatedly while doing things like inpainting, outpainting, re-prompting, etc. And likely required use of some other tools like Photoshop or Indesign (midjourney absolutely did not generate the bubbles or comic book text).

The issue is that the AI's output is deterministic, and the prompts it uses are commonly too short to be eligible for copyright, not least because someone else could plausibly use the same prompt independently. With the same seed, the AI would then produce the same image. Or produce a similar image in response to a similar prompt.

If you could copyright that, you then have an enforcement problem, because the usual way to show copying is similarity, but you can then have two images that are similar yet independent. And some light Photoshop work is not going to eliminate the similarity.

If someone builds a device that sprays paint onto a canvas based on some parameters, and then a person use those parameters, writes a small controller that sprays the paint a specific way, manually adjusts the position of the device over multiple sessions, and does their own touchups to create a piece of art, would they not hold the copyright on the resulting piece?

Yes, there is a deterministic component (although, to my knowledge, the AI tools aren't actually deterministic over their parameters). But the resulting work involves a lot of human creativity and intervention.

Suppose you have a printer with a programmable head which can be used to print out arbitrary images. Can you copyright the pictures it can print, even though the printing is done by machine? Obviously. But then it's not the machine choosing what you want on the page.

Perhaps my analogy wasn't clear. I meant, the painting device uses your parameters, but also supplies some modifications. Like a stenograph.

At the end of the day, generative AI, Stenograph, and this painting device are all tools.

If you shake a kaleidoscope and then take a picture through the lens, I'm pretty sure you have copyright on that photograph, even though someone else made the kaleidoscope

It's not the presence of a machine which makes it ineligible for copyright. It's the absence of some independent creative expression. If you add significant creative expression to what the machine does, that's not the same thing as the output of a machine produced from a minimal input.

What we're talking about here is definitely art where a human is providing independent creative expression, in combination with generative tools

"The issue is that the AI's output is deterministic"

Point a digital camera at a still scene under controlled lighting conditions and it's pretty deterministic also.

Anyway, it's easy enough to add randomness to the AI image generation process (and Midjourney does do that by default, as a matter of face), so the process need not be deterministic at all.

It's not about the physics of light, it's about what the artist chooses to have in the scene.

And it's not randomness that fixes it. Do you even want to get into the information theory problems if you could copyright the output of a random number generator?

What you need is for someone to be making creative choices with enough entropy to be considered original and not be reproducible by feeding a short sentence to an existing algorithm.

I really like how you put it. Good thought! But I don't quite agree that entropy is a criteria of originality. Because originality is not ranodmness, it's... distinctiveness.

If I commission an artist and give them exacting directions, they are still the artist and copyright owner of the work. They can then assign it to me, but I am not the artist simply because I commissioned the work and gave input.

Typically commissioned works are "work-for-hire" where the person paying for the work gets the copyright, no?

Same as how when you freelance, the customer (usually) owns the code you write for them.

Reply on news.ycombinator.com