AI & Copyright

Can someone copyright AI-generated art? Can AI own a copyright of its own art?

I’ve been pondering these questions over the past year or so with the advent of LLMs and other generative AIs. My general conclusion is that while an AI couldn’t be the owner of a copyright, whoever uses the AI - at least, prompt-driven generative AI in its current form - would own the copyright.

AI, at least in my opinion, is a tool. Much like Microsoft Paint, or Adobe Photoshop, or cameras. When someone uses one of these to create something, it’s not the camera, not Photoshop, and not Paint that owns the copyright, it’s the person who used the tool who owns it. And it’s definitely not Microsoft, Adobe, or the camera manufacturer who owns the copyright.

I’m not the first one to think that AI-generated art is subject to copyright in much the same way photography is. Timothy B. Lee made a similar argument is his article on Ars Technica Opinion: The Copyright Office is making a mistake on AI-generated art. He even mentioned the 1884 Supreme Court case Burrow-Giles Lithographic Company v. Napoleon Sarony where the lithographic company argued that a photo Napoleon Sarony took wasn’t subject to copyright. The reasoning? Because photography was merely a mechanical process.

The Burrow-Giles Lithographic Company did not win - after all, photographs were explicitly added to copyright law in a 1865 amendment to US Copyright law:

Thirty-Eighth Congress. SESS. II. CH. 125, 126. 1865

March 3, 1865

CHAP. CXXVI. - An Act supplemental to an Act entitled “An Act to amend the several Acts respecting Copyright,” approved February third, eighteen hundred and thirty-one, and to the Acts in Addition thereto and Amendment thereof.

Be it enacted by the Senate and House of Representatives of the United States of America in Congress assembled, That the provisions of said act shall extend to and include photographs and the negatives thereof which shall hereafter be made, and shall enure to the benefit of the authors of the same in the same manner, and to the same extent, and upon the same conditions as to the authors of prints and engravings.

I’m not sure how controversial during the 19th century the idea was that photographs could be copyrighted. Was there as much debate and differing opinions as we see online today regarding AI-generated pictures and text?

What is Copyright? What is the purpose of Copyright?

Before I continue, I want to try to answer these two questions. I say try because even these two questions seem to have differing opinions.

Etymologically, copyright is the right to make copies. So, whoever owns a copyright has the right to make copies of whatever it is they own a copyright on. Typically this is whoever created it.

In addition to simply copying, Wikipedia also includes “distribute, adapt, display, and perform” in its list of rights held by copyright owners.

Some people (don’t you love the wishy-washiness of those two words) believe that copyright holders also hold the right to what is done with their work - copy or otherwise. For instance, someone might not like someone else reselling their book. This belief is especially relevant in the AI-art debate since people often claim that AI-art is unethical due to training on copyrighted work - without the copyright owner’s permission.

Finally - in the question of “what is copyright” - copyright is a legal framework. Something can be or can’t be copyrighted - or something is copyrighted or in the public domain - based on what the law says.

Why copyright? This question isn’t as clear (despite the “what is copyright” being clear as mud).

A common argument I’ve seen (again with the wishy-washiness) is that copyright exists so that writers, artists, etc, can make money from their copyright.

The U.S. constitution says that copyright is meant “To promote the Progress of Science and useful Arts” (Article 1, Section 8, Clause 8).

Arguments Against Generative AI Copyright

Perusing the comments in the above-mentioned Ars Technica article I’ve seen a few common arguments against AI-copyright:

• allowing AI-generated works to be copyrighted would put people out of jobs
• “troll farms” would generate every possible art, thus owning every conceivable copyright
• generative AI completely random; it’s impossible to give the same inputs and get the same results
• generative AI is too quick and too easy
• generative AI trained on copyrighted data without the copyright holder’s permission
• choosing among a selection of AI-generated work does not constitute creativity; neither does refining AI-generated output

A few of these seem moralistic - whether we, as a society, should allow such copyrights depends on whether it would improve or impair society as a whole.

A few are based on false premises. Troll farms wouldn’t be able to generate “every possible image” - and, indeed, you can generate “every possible image” itself without AI. It would just take a long time.

Generative AI is not completely random. It’s just that the end-user typically only enters a prompt. There are weights in an AI system (typically a neural network) that are randomized, but if you set set the weights the the same thing again, the same output will be generated.

And just because someone can create something quickly doesn’t mean they shouldn’t be entitled to copyright - AI-generated or otherwise.

A couple of the arguments do hold merit: can something that was trained on copyrighted work itself be copyrighted? From what I understand, derivative works themselves are subject to copyright, regardless of permission. But is generative AI derivative? Or is it something else completely?

Regarding choosing among a selection of AI-generated work does not constitute creativity - the following comment by Onigato on the Ars Technica article sums it up best:

I appreciate the authors opinion, and they do raise several valid arguments, to which I only have a single counter.

Naruto the monkey.

When a person is using a tool like Dall-E or Midjourney they are not creating the image. They may be setting several parameters for the image-space to be selected from but they aren’t themselves creating the image. Similarly the photographer who had his camera stolen by a monkey didn’t create any of the images that were subsequently made, they only set up the parameters for the image-space to have all those photos taken (if accidentally, happy little accidents!).

That photographer then curated the final images, IIRC Naruto took nearly a hundred pictures and the photographer only actively released a dozen of so of them, and only three or four were considered “important” to the photographer.

Same thing with these generated images, the final selection is curated from dozens, hundreds, possibly thousands of “results”. The curator may narrow or broaden the curation field with more or less prompt, but the actual creation is performed by a non-thinking application, even further removed from sentience and sapience than Naruto’s entire species combined.

TL:TR; The photographer who had his camera stolen by a monkey curated pictures and didn’t get copyright on the final product, a person using a generative application curates the pictures and shouldn’t get a copyright on the final products.

Conclusion?

So far, the US Copyright office has refused to accept registrations for generative AI. But copyright no longer requires registration. So the question remains - can AI outputs be subject to copyright?

Ultimately, the question of “can AI outputs be subject to copyright” will fall to politicians, lawyers, and diplomats - whether in the courts, the legislature, or international convention.