Training Generative AI Models on Copyrighted Works Is Fair Use - Change My Mind

I think we should have a rule that says if a LLM company invokes fair use on the training inputs then the outputs are public domain.

Google scanned millions of books and made them available online. Courts ruled that was fair use because the purpose and interface didn't lend itself to actually reading the books in Google books, but just searching them for information. If that is fair use, then I don't see how training an LLM (which doesn't retain the exact copy of the training data at least in the vast majority of cases) isn't fair use. You aren't going to get an argument from me.

I think most people who will disagree are reflexively anti AI, and that's fine. But I just haven't heard a good argument that AI training isn't fair use.

What constitutes fair use?

17 U.S.C. § 107

Notwithstanding the provisions of sections 17 U.S.C. § 106 and 17 U.S.C. § 106A, the fair use of a copyrighted work, including such use by reproduction in copies or phonorecords or by any other means specified by that section, for purposes such as criticism, comment, news reporting, teaching (including multiple copies for classroom use), scholarship, or research, is not an infringement of copyright.

GenAI training, at least regarding art, is neither criticism, comment, news reporting scholarship, nor research.

AI training is not done by scientists but engineers of a corporative entity with a long term profit goal.

So, by elimination, we can conclude that none of the purposes covered by the fair use doctrine apply to Generative AI training.

Q.E.D.

Selling an AI model (or usage of that model) that allows for producing works that are clearly based upon those copyrighted works and would be considered copyright infringement if a person did the same thing is not fair use.

If a person creating the same thing as generative AI would be infringing, then it isn't magically not infringing because it is on the internet or done by a program. Basically, AI needs to follow the same rules and restrictions as a person would. That does mean that the AI also needs to be trained to not create copyright infringing works if the use of the AI is being sold.

As a downloadable model that anyone can use at no cost? Sure, whatever is fine. Then it is on the person who uses it and tries to infringe. But if someone pays a company to use their AI to create infringing work, that is on the company and they are just as at fault as if they sold T shirts that infringed on copyright.

It isn't fair use, See most of faq @ fairuse faq.

"Fair Use" is often the subject of discussion when talking about online copyright with regards to online video content or music sampling, but it's notably a flawed defense as it generally has no legal definition for how much of certain content can be used or referenced. The very first line of that faq has the following note:

How do I get permission to use somebody else's work?
You can ask for it. If you know who the copyright owner is, you may contact the owner directly. If you are not certain about the ownership or have other related questions, you may wish to request that the Copyright Office conduct a search of its records or you may search yourself. See the next question for more details.

All artists / writers and others are asking LLM model producers to do is a) Ask for permission or B) Attribute the artists work in some kind of ledger, respecting the copyright of their work. Every work you make (write/play/draw/whatever) has a copyright that should be respected by companies and are not waived by EULA or TOS (ever) and must be respected in order for author attribution as a concept to work at all. There is plenty of free, permissive copyrighted content on the internet that can be used instead to train an LLM, but simply asking for permission or giving attribution would at least be a step in the right direction for these companies and for the industry as a whole.

Defenders of AI will note that the "use" of art in LLM is limited and thus protected by fair use, but that is debatable based on the content of the above listed FAQ.

How much of someone else's work can I use without getting permission?
Under the fair use doctrine of the U.S. copyright statute, it is permissible to use limited portions of a work including quotes, for purposes such as commentary, criticism, news reporting, and scholarly reports. There are no legal rules permitting the use of a specific number of words, a certain number of musical notes, or percentage of a work. Whether a particular use qualifies as fair use depends on all the circumstances. See, Fair Use Index, and Circular 21, Reproductions of Copyrighted Works by Educators and Librarians.

You can see that the use cases above (commentary, criticism, news reporting and scholarly reports) does not qualify LLM companies to use or train their models with copyrighted data for privatized industry. Additionally, you'll note that "market disruptive" uses cannot be protected by fair use in it's definition, meaning that displacing artists with AI automatically makes LLM use of copyrighted material an infraction of copyright that is not protected by the fair use clause.

Regardless, this will need to be proved in court and even if it passes certain criteria, it will not apply to all infractions. Fair use is a defense, not a protection, and thus LLM producers will have to spend time in court in order to defend individual infractions. There's no way for them to catch all copyright infringement with one ruling, it needs to be proved on a case-by-case basis.

IANAL but this is my 2 cents on the matter.

Fair use doesn't exist in Japan, so I wonder how it applies to content made there.

Here's another good one: https://www.eff.org/deeplinks/2023/04/how-we-think-about-copyright-and-ai-art-0

Agreed. I would also argue that trained model weights are not copyrightable.

I agree with your statement as stated. (EDIT: title has changed since posting)

The part I have an issue with is the output that tends to fall into 4 categories:

1. output that is creative and unique is acceptable

2. output tricked into revealing source is infringement

3. output not creative, as in biometric identification, utilizes the entirety of copyright work in a non-transformative way and needs to be licensed.

4. output of creative work, while not copyrightable, can still infringe trademark.

Do you mean "fair use" in the legal sense or the ethical sense?

Sure, that can be fair use, but only if using them can also be fair use