Meditation on Copyright

NOTE: This is an OLD post I wrote more then half a year ago. I never finished because I never found a good conclusion. But I decided to publish it as the AI situation has become dire for not just art and computer scient, but for the entire humanity. So here's my little contribution to the discussion. The following is a raw log of my thought at the time.

Time to think more about art. AI has impacted the art industry the most. Artists see their income dwindle. Customers with low budget got sucked away by AI with much lower cost of production. And high budget customers are figuring out they sometimes can get away with lower quality results. Not good for the artists. One thing comes up often during the discussion of artists wanting to ban AI generated art - it is not art and the AI stole their content in violation of copyright. Last time I hope I expressed my distaste for why I think that's a bad take. The entire view is anthropocentric and what works... works. This time I want to make my attempt to improve on the subject on copyright. The end goal is to arrive at a (set of) definitions with philosophical rigor. So we know someone violated copyright when we see it, with high confidence. Or not, and we prove it to be impossible. At least to create a design boundary on what's bad and should not exist in any good final decisions.

IMPORTANT: This post assumes you are familiar with how image/audio generation AI works internally. Latent space, path over manifold, embedding, computation, etc..

My interest in copyright starts with 3 core needs:

• Open source and Copyleft depends on Copyright working
• GPL code doesn't become proprietary via coding assistants
• Artists have put labor into the work. Their labor ought to be protected. Unless the market deems the result has 0 value, which is currently not the case.

That's despite my recent work being released under ISC or BSD-0, basically public domain releases.

Yes, it means the idea within this post is motivated reasoning. But we can't arrive at any conclusion if we do not assert any preference about a world state we desire - that's the entirety of the is-ought problem.

First, let's examine what the Copyright law says about reproducing (which is what the artists are complaining about) and see how broken the definition we are using is. IANAL, but it doesn't matter here - the law should be a model of the reality. Not the other way around. And a baseline understanding should be good enough. Since I live in Taiwan, let's see what Taiwan's copyright law says. In section 3.5.

"Reproduce" means to reproduce directly, indirectly, permanently,
or temporarily a work by means of printing, reprography, sound
recording, video recording, photography, handwritten notes, or
otherwise. This definition also applies to the sound recording or
video recording of scripts, musical works, or works of similar nature
during their performance or broadcast, and also includes the
construction of an architectural structure based on architectural
plans or models.

Congratulations! This definition is useless. It is lacking critical definitions that we need right now - given two works, what's the procedure to determine if one is a reproduction of the other? It amounts to "I know what reproduction is when I see it". Great. Let's start from first principles then.

So what's the point of Copyright?

The core of the problem copyright is trying to solve being - mental work is work. And we ought to protect the rights of the fruit of intellectual activity. Alice wrote a book about growing corn. It's the result of her life's experience and expertise all about growing the most amount of corn per unit of resource poured into the land. She should be able to say "You are free to practice and share what I teach. But only I can print the book. You should not be able to make your own copy".

That is a reasonable ask and baseline. However it faces serious challenges in the computation world we live in. What exactly is a copy? Sure making an exact 1:1 copy is a copy. What if it's the same content but a different cover? What if someone rephrases what she taught into a local language? What if an academic looks at it, finds several errors despite it's generally correct so made a corrected copy? What if Bob decided to change letters one by one into something random? 1 letter? 1%? 2%? 5%? 20%? 99%? When does the resulting book become garbage that Bob came up with painstaking dice rolls instead of Alice's book? At 100% it definitely is Bob's garbage. What if a monkey with a typewriter accidentally types out Alice's book out of sheer luck?

We are back to the original observation - there is no stable and coherent definition of a copy/reproduction. "I know when I see it" does not work when we spend 2 paragraphs probing the boundary of the problem.

The same problem exists for paintings too. Alice could spend 10 years of her life drawing the best art humanity could ever create. And Bob could change it pixel by pixel. At what point does it become Bob's garbled painting? Even worse, what if Charlie, a hacker, inspired by his trip to the local science museum, decided to paint in the frequency domain. After who knows how long, Charlie comes up with a painting that when DCT is applied becomes Alice's painting because that just what peak of art being? Does Charlie infringe on Alice's copyright even if Charlie never landed his eyes on Alice's painting, and on a different domain? Does the Library of Babel infringe on every possible book because no one can find the already existent book in the library? Or is it the other way around?

But wait it gets worse. We not only do not know how much transformation turns one work into another. We also don't have a coherent definition on what kind of transformations constitute creating a new work. It can't be "if it's possible to turn work A into work Alice's painting, then A infringes on Alice". The problem being that's asking if f exists such that f(B) = Alice's painting then B infringes on Alice. f could simply be unconditionally make everything Alice's painting or an arbitrary function something made up that happens to include B in the image (in the mathematical sense) or some exotic one-to-one-and-onto function that guarantees something maps to B. And the pre-image to the function f could be arbitrary.

Likewise it also can't be "if something looks exactly like Alice's painting". Then I could throw the copyrighted work into a low quality JPEG and claim it's my new work. However we as the human society seems to agree that "repaint in the style of Alice" is definitely infringement. The only coherent and legible difference between them I could find is "can a human write down the procedure to turn B into Alice's painting". And things have changed. Humans still can't. But we learned how to make computers write down the procedure, and it's 4GB in size. The law might argue with the "The Reasonable Person" standard but that Reasonable Person™ now has an army of state of the art Large Language Models and nearly infinite computing power.

You might also argue "but that's my style. It exists". Unfortunately Kolmogorov Complexity is uncomputable so it's useless for this debate.

My gut feeling is, copyright was always incoherent but only held up by the high cost of work creation. It is no longer true in the world of AI and computation. That cost is as low as it has ever been. And that exploded the wound wide open.

Hope we can all agree on how broken the definition being used right now is, how anthropomorphism being at the center of the entire problem, where the incoherence lies and it is so vague that it's impossible to draw a line between what is and what is not a copy. And to that point some quick answers to questions I expect to be asked by readers of this blog.

But some line (even if incoherent) is better than no line

No. An incoherent line that breaks unpredictably will cause more harm than good. See what DMCA did, the pain it caused and the barely helpful legislation that people actually benefit from it. And ask the person who figured out "fair use" means "what the judge's mood feels that day".

A philosophy war right now will not fix artist's pain

Yes. But I have been raising this problem since very early days of mine. Pre this blog existing. If my memory serves me correctly I argued about this ambiguity in the Intellectual Property and Patent course in my college in 2020 and my thought experiment along the lines of above in 2017. No one wanted to engage with the problem I proposed. And the philosophy is the problem that leads to the pain.

You are not being helpful. Actually help solve the problem.

But I am. Let's agree that the copyright model is broken. And let's rule out all the dumb ideas to constrain our design space.

We can just ban generating images with AI

You can't. And that's not me being hard. It's governments tried the Crypto War back in the 80s and 90s. Utter failure. And you can't uninvent latent space and walking vectors through a manifold. Also there's many ways to generate images besides diffusion. GANs, VAEs, etc..

Let's setup some design constraints. We can't define copyright based on transformations as shown above, it's a losing battle. We can't define it based on the similarity of the final result either. That runs into the fun problem of multiple parties can independently converge on the same product.

This is where I left my thought more 6 months ago. I have made no progress since then. I feel that I have exhausted all enumeration and permutation of the problem space, each solution is far from working. Maybe.. the idea of copyright is an incoherent idea from the beginning that only worked because humans are slow at producing new martial. Once materials are computationally enumerable, the only tie, that we didn't know exists or refused to recognize the existence of, broke and all hell broke loose.