GenAI and Copyright
Risk
Both training and using AI services—especially GenAI—can put you at risk for copyright, right to privacy, or right of publicity lawsuits. Rest assured that there are many ways that you can minimize the risk of assuming liability. These mitigations include an understanding of substantial similarity," fair use, and what kinds of things have or haven’t been settled legally regarding the use of GenAI.
Substantial Similarity
Substantial Similarity refers to the degree of similarity between two works and whether it is significant enough to conclude that one work may have been copied from the other. Whether or not this similarity is intentional, a high level can lead to a copyright lawsuit, especially if the similarity was intentional. Because a GenAI service could produce something very similar to an already existing work, this is an important issue to consider. This standard of similarity is somewhat subjective as there is no set rule for determining how much similarity constitutes what is “substantial”. Courts assess this on a case-by-case basis.
Fair Use
Uses of GenAI that fall under the fair use provision of U.S. copyright law are not copyright infringement. It's important to understand, however, that the principles of fair use aren't a checklist of things to be done. Rather, they are factors that are weighed holistically by a judge to determine whether a specific use is fair and justified. Careful analysis of the fair use factors can help you make an educated guess about whether a use is fair and can help you assess copyright risks associated with that use. The four fair use factors are:
To better understand the principles of fair use, the BYU Copyright Licensing Office has extensive resources available to help you understand each of these principles and the legal subtleties of each.
Right of Publicity/Right to Privacy
Another area of law to consider in relation to GenAI is Right of Publicity. This idea states that people have a right to their own name, image, voice, and likeness, and it is up to them to determine how those aspects are used. Even though there isn’t yet a federal law protecting this right, there are several states that have passed legislation in this category. For example, if you used AI to create an image that looked very similar to someone (especially a famous person), and used that image in an advertisement, that image could result in a lawsuit.
The United States Copyright Office has published a Report on Copyright and Artificial Intelligence, Part 1 of which addresses digital replicas of people and how they might be addressed going forward. The report does not have the force of law, but it reflects the expertise of the government agency tasked with administering the copyright regime of the United States. It is expected that this report will influence courts’ reasoning in AI-related copyright lawsuits going forward.
Training GenAI
On the surface, it is copyright infringement to use other people's content as training inputs for your own AI models without permission. It's possible that the fair use defense could apply, but it's difficult to know for sure. Courts so far have only provided a few decisions on how fair use applies to AI training (For example, see Kadrey v. Meta and Bartz v. Anthropic. For now, be careful about relying on fair use; it may be best practice to only use data that you own or have licensed when training AI models.
Part 3 of the U.S. Copyright Office’s Report on Copyright and Artificial Intelligence (which exists now in a prepublication version) addresses the copyright considerations of training AI models, including the possible applicability of the fair use defense.
Using Outputs
If someone uses a trained AI model or service, and an output yields results that someone thinks violates their copyright (like a work that was used in the training without permission), it's still unclear whether that liability would fall on the creator of the model or the user generating the content. Please be aware that this is a risk. The forthcoming Part 4 of the U.S Copyright Office’s report will attempt to answer this question.
As the user of a service, it's safest to use services that use their own data for training. For example, Adobe Firefly, which owns or licenses all of its training data, may be safer than Midjourney.
The table below outlines scenarios to consider when deciding the level of caution warranted by GenAI-related activities.
| Risk Factor | Effect | Example |
|---|---|---|
| Non-profit use | Lower Risk | Creating an image to put on the wall of your room |
| Educational use | Lower Risk | Generating text to use as inspiration for an assignment |
| Minimal distribution | Lower Risk | Sharing an image with classmates or friends |
| Moderate distribution | Higher Risk | Posting generated content online |
| Not knowing the source of training data | Higher Risk | Creating a work of art using an AI service that uses copyrighted pieces of art in the training data |
| Imitating a specific piece of work | Much Higher Risk | Trying to imitate Harry Potter |
| Imitating someone or their style | Much Higher Risk | Creating an AI imitation of Taylor Swift's voice and sharing it |
| Not following terms and conditions | Extreme Risk | Publishing content in a way that is expressly forbidden |
Copyrighting GenAI Outputs
The output from a GenAI service is not currently copyrightable. The United States Copyright Office does not deem the level of human authorship to be sufficient for copyright eligibility. That being said, if you create a work using GenAI, and make significant additions or edits, those additions or edits could be copyrightable if they are sufficiently creative. The base GenAI work, however, would still not be copyrightable.
In addition, if a prompt for a GenAI service was long, complex, and creative enough to be considered a literary work, it could be copyrighted. However, the content generated by that prompt would remain unprotected.
Part 2 of the U.S. Copyright Office’s Report on Copyright and Artificial Intelligence provides guidance on the copyrightability of work related to AI.