Copyright Registration of AI-Generated Works Checklist

By: Kirk A. Sigmon, BANNER WITCOFF

THIS CHECKLIST OUTLINES KEY CONSIDERATIONS THAT ATTORNEYS should review when advising whether and how to copyright artificial intelligence (AI) and machine learning (ML)-generated works in the United States.

The checklist provides a framework for documentation of human involvement in the creative process of an AI-generated work and for the preparation of a copyright application. It focuses on collecting information useful for both the application and for responding to follow-up by the U.S. Copyright Office.

As a preliminary matter, applicants should exercise caution when trying to copyright works generated using AI or ML models. The U.S. Copyright Office (the Office) carefully scrutinizes such applications. Specifically, the Office has issued guidance stating that individuals using AI/ML technology to create a work may claim protection “for their own contributions to that work,” but if “a work’s traditional elements of authorship were produced by a machine, the work lacks human authorship and the Office will not register it.”¹ The Office thereby draws a line between work of an author’s “own original mental conception, to which [the author] gave visible form” and creative works of a machine (including simple mechanical reproductions).²

Documenting human involvement in the creation of an AI or ML-generated work is important because (1) the Office expects applicants to explicitly distinguish between human and AI contributions in copyright applications, and (2) the Office sometimes requests additional information from applicants when evaluating possible limitations on a copyright application involving AI-generated content.

Document the Nature of the AI

The training and capabilities of an AI model can have significant impact upon its ability to contribute—or not contribute—to a creative work. For example, if a model is rudimentary (e.g., designed to remove compression artifacts from existing images, designed to add makeup to a human face, or the like), then it might be fairly presumed to be less likely to provide creative output. As such, more human creativity might be implied in the resultant creative work. That said, if a model is highly sophisticated and trained based on previously published works, that model might be assumed to more readily provide what appears to be a creative work with relatively minimal human effort.

• Record model(s) used. If an existing model (e.g., a model downloaded from the internet) was used, collect information regarding the model such as:
  • When it was retrieved
  • Where it was retrieved from
  • A recorded version number
  • The date and/or time the model was used
  • Other similar information
• Document known model uses. Some generative models (such as Stable Diffusion) can generate wholly new images, whereas some other models (such as those used as plugins in photo editing suites) are trained to improve and otherwise modify existing images. It is generally easier to argue that the latter are similar to conventional photo editing tools.
• Document model training process. If available, document how the model was trained. This can include:
  • Documenting the training data that was used, including information such as:
    • Where the data originated
    • Who owned the data
    • The format of the data
  • Documenting the training process itself, such as:
    • Which algorithms were used
    • Which loss functions were used
  • Documenting, where applicable, whether the model is designed to continually learn, such as where it might receive further training as part of a feedback loop during use

Example: Some freely available Stable Diffusion models accessible through enthusiast websites are quite sophisticated and are trained to emulate specific authors’ work. It may be relatively difficult to copyright their output because those models require very little effort to produce output that appears quite creative and because the models are designed to, in effect, create permutations of another author’s previously published work. In those circumstances, applicants should endeavor to document as much human creative labor as possible (and should expect an uphill battle). With that said, other models, while equally sophisticated, are designed to simply clean up and/or otherwise enhance existing works. Copyrighting the output of these models seems significantly easier, in no small part because they are roughly analogous to an advanced photo filter.

To review the complete checklist, which includes tips for documenting the scope of the AI contribution and documenting the human creative labor, along with best practices and a draft application, subscribers may follow this link to read the complete article in Practical Guidance.

Not yet a Practical Guidance subscriber? Sign up for a free trial to read the complete article.

Kirk A. Sigmon is an attorney at Banner Witcoff’s Washington, D.C. office. His work in the United States and in Asia, tied with his experience with Fortune 500 companies and startups, provides him the know-how to counsel clients at all stages of invention, patent prosecution, intellectual property enforcement, and litigation.

To find this article in Practical Guidance, follow this research path:

RESEARCH PATH: Intellectual Property & Technology > Copyright > Checklist

Related Content

1. 88 Fed. Reg. 16190, 16192-193 (Mar. 16, 2023). 2. Burrow-Giles Lithographic Co. v. Sarony, 111 U.S. 53, 60 (1884).