By Jill Hubbard Bowman
A U.S. copyright owner has the exclusive rights to: [i]
The rights above allow a copyright holder to control copying, modification, distribution, and monetization of their work. These rights, however, are not unlimited. The scope of exclusive rights to a copyrighted work is only to the original expression in the work created by the human author. The monopoly does not extend to uncopyrightable subject matter in the work.
The right of reproduction protects the expressive elements of a work from being copied in some situations to allow a copyright owner to exploit markets for its expressive work.
The right to prepare derivative works protects the creative choices made in a work from being modified or copied and used to create a new work in new markets like the creation of a movie from a book. The right of distribution allows the copyright holder to control distribution of the work to others. The rights of public performance and display allow the copyright holder to control showing the work in public forums. All or some of these rights may be licensed or transferred to others.[ii]
In context of AI technology, the balance between public access to information and protection of markets for copyrighted expression in a work is in tension.
The term “derivative work” and the scope of the related exclusive right is being contorted in the AI technology copyright circus. The term is a complex, confusing, and commonly misused.
Everything downstream in an AI processing pipeline isn’t a “derivative work.”
Determining what AI technology may be a derivative work is being hotly debated and litigated. There is widespread confusion about the proper contextual use of the term “derivative work” and its implications. Some make broad conclusory assertions that generative AI models trained on copyrighted content and all subsequent output are derivative works of the training datasets whether or not the technology contains the expressive elements of a copyrightable work.[iii] A foundational assumption to these claims is that neural network AI models incorporate and compress the training data. This assumption is strongly disputed by AI engineers and generative AI makers.[iv]
When licensing AI technology, many simply assume that any modification of AI technology in the processing pipeline, from data to output, is a “derivative work,” whether or not the material contains original expression of a pre-existing copyrightable work or is independently copyrightable. The term “derivative work” is used widely, loosely, and often at odds with the U.S. Copyright Act and judicial opinions. This may have serious consequences if the business is relying on copyrights to control access, use and monetization of modified technology. It may unwittingly lose control over modified AI technology because it is not a copyrightable “derivative work.”
“Derivative work” is a legal term of art specifically defined in the U.S. Copyright Act as “a work based upon one or more preexisting works, such as a translation, musical arrangement, dramatization, fictionalization, motion picture version, sound recording, art reproduction, abridgment, condensation, or any other form in which a work may be recast, transformed, or adapted. A work consisting of editorial revisions, annotations, elaborations, or other modifications which, as a whole, represent an original work of authorship, is a "derivative work.”[v]
Law Professor Daniel Gervais, in his article, “AI Derivatives: The Application to the Derivative Work Right to Literary and Artistic Productions of AI Machines states: “[i]f there is one definition in the United States Copyright Act that raises eyebrows and uses language that cannot literally mean what it says, it is the definition of ‘derivative work.’”[vi] It can’t be anything “based upon” a pre-existing work. When we author literary or musical works, we don’t invent our own language or music notes. In a sense, everything humans create is based on what we have learned and experienced in many ways from many mediums. Indeed, multi-Grammy winner Ed Sheeran claimed he would quit songwriting if basic music chords were restricted by copyright law and “owned” by some allegedly original chord composer in the sixties.[vii]
Professor Gervais notes that copyright holders of works in training datasets are likely to broadly interpret and overextend the definition of “derivative work” to AI technology. He argues that the legal construction rule ejusdem generis limits what a derivative work is to the list of specific actions with the copyrighted work - recasting, transforming, or adapting - that follow the broader language “based upon” in the legal definition.
There is also a crucial debate as to whether there is a distinction between a copyrightable derivative work, as interpreted by the U.S. Copyright Office, and an infringing derivative work, as interpreted by federal courts in copyright infringement litigation. (The split in the Circuit Courts’ interpretations of “derivative work” and requirement of originality is discussed in the next Section on Copyright Infringement). [viii]
According to the U.S. Copyright Compendium, creating a derivative work requires “a process of recasting, transforming, or adapting ‘one or more preexisting works.’”[ix]
“[D]erivative works contain two distinct forms of authorship:
The U.S. legislative history describes “derivative works” as “every copyrightable work that employs preexisting material . . . of any kind . . .”[xi]
When an author is authorized by a copyright holder to make a derivative work from a copyrightable pre-existing work or the pre-existing work is in the public domain, a new copyright may arise to the modifications or new expressive content that independently qualifies for copyrightability. The status of the pre-existing work is unchanged. A new copyright arises only to the new copyrightable content.[xii]
Modifications to AI technology may not be a copyrightable derivative work.
AI technology is not like regular software. Unlike modifications to standard software, many modifications to AI technology are not copyrightable “derivative works.” In the U.S., to be a copyrightable derivative work, modified AI technology must meet the criteria for copyrightability. A new derivative work copyright will not arise if a) the author of the new material isn’t a human (like an autonomous AI system or machine) who is making creative choices and exercising ultimate control over the creation of the new material, and b) the new material does not have sufficient, independent, original expression (like purely functional and technical material).
In its recent registration guidance on AI-generated material, the U.S. Copyright Office stated:
“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. For example, when an AI technology receives solely a prompt from a human and produces complex written, visual, or musical works in response, the “traditional elements of authorship” are determined and executed by the technology—not the human user. Based on the Office’s understanding of the generative AI technologies currently available, users do not exercise ultimate creative control over how such systems interpret prompts and generate material. Instead, these prompts function more like instructions to a commissioned artist—they identify what the prompter wishes to have depicted, but the machine determines how those instructions are implemented in its output.” [xiii]
For modified AI technology consider if it lacks: a) incorporation of pre-existing work that was recast, transformed, or adapted, and b) new authorship where the human controlled the creation of the recast, transformed, or adapted original material by making creative choices.
Under U.S. law, the following are likely not copyrightable “derivative works”:
In the first example, purely functional modifications lack sufficient originality for copyrightability. For example, AI processing software and model architecture code is often licensed under open-source terms and then modified and reorganized by subsequent authors based on technical considerations for new use cases. If a subsequent author made purely functional changes and rearrangements for technical reasons, like efficiency or performance, this type of modification likely wouldn’t qualify for copyright protection and be a copyrightable derivative work. But someone upstream may own a copyright to original, human-authored code. In the other examples above, if the “traditional elements of authorship” are AI-generated, copyright protection will not be available since the material lacks human authorship, which is required under U.S. copyright law.[xiv]
Human creative modifications to uncopyrightable AI technology may be copyrightable.
A work that has sufficient human originality and contains AI-generated material may be copyrightable. For example, a work that includes original, creative human modifications of code written by a generative AI system, which currently is not copyrightable, may be copyrightable. But the pre-existing AI-generated code must be disclaimed in a copyright registration for the human authored work. Its uncopyrightable status would be unchanged. Identifying and separating the human modifications from the AI generated output may be difficult in a mixed work.[xv]
Without other IP rights, a business may not have legal control over modified AI technology if the modification isn’t a copyrightable derivative work.
Calling all modifications of AI technology “derivative works,” whether or not the pre-existing material or new material was authored by a human and copyrightable is problematic. It can cause confusion and incorrect assumptions about the subsequent rights to control access, use, and monetization of the modified AI technology. Although the subsequent work may be “based upon” the pre-existing work, downstream legal rights may be missing, which may have serious consequences for a business simply relying on copyrights. Unwittingly, it may lose control over the unprotected, uncopyrightable modified AI technology. Without copyrights, a business should rethink how it approaches licensing and distribution of AI technology and carefully consider whether it should use a license agreement meant for traditional software.[xvi]
Navigation Tip: To assess whether AI technology is actually a copyrightable derivative work, it's important for AI technology licensing attorneys to understand where AI engineers obtained the original technology and specifically how that technology has been recast, transformed, or adapted. They should also understand what the human AI developers actually authored verses what was AI-generated without explicit human control of any expressive output. (See post on Copyright Analysis Questions.) In some situations, it is very helpful for AI developers to actually show the attorneys the code. They may be surprised.
[i] https://www.copyright.gov/circs/circ01.pdf.
[ii] 17 U.S.C. § 101. https://www.govinfo.gov/content/pkg/USCODE-2011-title17/pdf/USCODE-2011-title17-chap1-sec101.pdf.
[iii] Many copyright holders have sued makers of generative AI models asserting very broad copyright infringement claims in U.S. federal courts. Some courts have rejected some of the plaintiff’s overbroad arguments about the scope of possible “derivative works.” Paul Tremblay et v. OpenAI, Inc. et al., No. 3:23-cv-03223 (N.D. Cal. Feb. 12, 2024).
[iv]Microsoft Notice of Inquiry Comments to the US Copyright Office https://www.regulations.gov/comment/COLC-2023-0006-8750; OpenAI Copyright NOI Comments https://www.regulations.gov/comment/COLC-2023-0006-8906
[v] 17 U.S.C. § 101 (emphasis added. Definition of “derivative work”). https://www.govinfo.gov/content/pkg/USCODE-2011-title17/pdf/USCODE-2011-title17-chap1-sec101.pdf.
[vi] Daniel J. Gervais, “AI Derivatives: The Application to the Derivative Work Right to Literary and Artistic Productions of AI Machines” (hereinafter AI Derivatives), 53 Seton Hall Law Review. (2022) https://scholarship.law.vanderbilt.edu/faculty-publications/1263.
[vii] https://www.youtube.com/watch?v=NcCKlsTgjeM (See Ed Sheeran playing his guitar and demonstrating similar chords in popular songs.)
[viii] AI Derivatives, p. 7.
[ix] Citing H.R. REP. NO. 94-1476 at 57, reprinted in 1976 U.S.C.C.A.N. 5659, 5670; S. REP. NO. 94-473 at 55.
[x] Copyright Compendium 507.1 (where a new author modifies a preexisting work, the “new authorship . . . may be registered, provided that it contains a sufficient amount of original authorship”) https://www.copyright.gov/comp3/chap500/ch500-identifying-works.pdf.
[xi] H.R. REP. NO. 94-1476 at 57, reprinted in 1976 U.S.C.C.A.N. at 5670; S. REP. NO. 94-473 at 55.
[xii] Copyright Compendium 507.2
[xiii] See generally Copyright Registration Guidance: Works Containing Material Generated
by Artificial Intelligence https://copyright.gov/ai/ai_policy_guidance.pdf.
[xiv] U.S. Copyright Office Review Board, Decision Affirming Refusal of Registration of a Recent Entrance to Paradise at 2 (Feb. 14, 2022), https://www.copyright.gov/rulings-filings/review- board/docs/a-recent-entrance-to-paradise.pdf.
[xv] https://www.federalregister.gov/documents/2023/03/16/2023-05321/copyright-registration-guidance-works-containing-material-generated-by-artificial-intelligence.
[xvi] See the subsequent licensing section in this U.S. Copyright Law AI Map for more information.
