Copyright and generative AI

Copyright ownership

In the US, copyright ownership for generative AI outputs is already fairly settled. Human creativity is required for a copyright to exist, so most outputs of generative AI tools are not eligible for copyright on their own. That is, most generative AI outputs do not have any copyright. However, when you add human creative expression, that creative expression can have a copyright - and many generative AI outputs are then manipulated by people in ways that can create a copyright.

For example, a series of images produced by an image generator are not eligible for a copyright on their own. (That is, they do not have a copyright, there is no copyright owner.) But if human authors use that same series of images to make a comic book, the process of selecting and arranging the images is human creative expression and the whole book is eligible for a copyright. (That is, the book as a whole has a copyright, and a copyright owner. The pictures individually still don't.) 

Contracts about ownership and use

Although there may be no copyright owner for many AI outputs (which usually means anyone can use a creative work), there may be contracts that claim to establish ownership, or that limit how the outputs of AI systems can be used. For example, most general AI tools have terms of use that say who can use the output, and for what purposes. In many cases, you can't use the output commercially without a paid subscription. 

Copyright issues in training

Training most generative AI tools involves use of enormous sets of text, images, audio, video, and other types of copyrightable creative works. While a few providers of AI tools have trained their models using only materials which are in the public domain, or for which the training company has a relevant license, most training is done on in-copyright materials. 

Most companies and other organizations training on in-copyright materials believe that this training is legal under the United States copyright provision for fair use. Their legal argument is derived from principles that were developed in cases about search engines (which copied web content to make it findable), and about book scanning projects (which copied books to make it possible to search inside them.) Broadly speaking, this kind of copying often does not involve humans having access to the entirety of the original creative works, and does involve using the works for an unusual (or "transformative") purpose, which both tend to strengthen fair use claims. (It is also slightly more likely to be fair use for non-profit researchers to do this, than for commercial companies.) 

Currently, numerous copyright lawsuits are pending in the United States and elsewhere over the legality of copying existing materials without permission for training AI systems. There is a lot of marketing rhetoric from general AI companies that suggests there are no copyright issues with using existing materials as training data, and that's not really correct. The fact that the suits are underway does indicate that the arguments both for -and- against fair use are strong - if it was immediately obviously fair use, the cases would have wrapped up by now. However, some of the early verdicts out of these cases are affirming fair use even in commercial applications, especially of legally-acquired (rather than pirated) original works. 

Non-profit research projects using existing materials to train AI systems are even more likely to be able to rely on fair use than commercial actors. Similarly, due to their more strongly transformative purpose, researchers developing specialized AI tools may be more able to rely on fair use. However, researchers should pay careful attention to developing case law, as some unexpected rulings may have implications even there. 

Contract issues in training

While the copyright law on using existing materials as training data is evolving, many existing content owners are taking action to prevent their materials from being used this way via contracts, such as terms of use or subscription contracts. For example, most subscribers to streaming TV and movie services have agreed to terms of use that say you cannot use those materials for anything but personal, noncommercial viewing. And the research journals and databases that the University Libraries subscribe to on behalf of campus also often only allow for noncommercial research use. In these cases, a contract provides a legal limit on using the materials as training data, even though the copyright law question is still open. 

If you violate a contract, the most likely result is that your business relationship with the other parties to the contract will end - although lawsuits are still possible, they're rare. So if you violate your streaming media terms of service, you will likely find yourself cut off from that service. If you violate the subscription terms for Libraries susbcription resources, all campus users may find themselves cut off from acccess. Some of our subscription resources -do- have options for using them with AI, but please read more about that, or get in touch, if you are interested in doing so. 

Contractual limitations on using existing materials to train AI tools exist regardless of the developers' commercial or non-profit intentions, and often even regardless of research or scholarly intentions - they depend in particular on the specific terms of a specific contract, so researchers should be careful in this area. 

Occasionally, a contract or license may expand the possibilities of using existing materials to train AI tools - an open license such as a Creative Commons license pre-approves some kinds of uses. Note that Creative Commons licenses often prohibit commercial use, and it's unclear how far down the chain of a training process those prohibitions might persist. Note also that most variations on Creative Commons licenses require attribution as a condition of use - it's currently unsettled how this interacts with training uses (since most LLMs and related systems are not very capable of true attribution.)

Creative Commons licenses also explicitly preserve fair use and other copyright exceptions, so if at some point it becomes settled law that fair use covers training uses, a lot of these considerations specific to Creative Commons (or other open licenses) may become irrelevant. But at the moment, CC licenses and other open licenses can be one way to create certainty in training sets.