Authored By: Duru Kurtulus
Istanbul Aydin University
Introduction
In 2023, the U.S. Copyright Office declined to grant full copyright protection to Kristina Kashtanova’s graphic novel Zarya of the Dawn, which included illustrations created with the help of the AI tool Midjourney. Cases such as this highlight an important unresolved issue in copyright law: when a work is created through the involvement of both a human and an AI system, how should the law distinguish between human authorship and expression generated by a machine? The U.S. Copyright Office currently considers whether a human has exercised “meaningful creative control” over the expressive elements of a work. As a result, parts of an AI-assisted work that are created by a human may qualify for copyright protection, whereas material generated autonomously by AI generally does not. However, this distinction between the use of AI as a tool and AI as an author, which was most recently reinforced by the Supreme Court’s refusal to hear Thaler v. Perlmutter, provides limited guidance for a much more typical situation: a novelist who uses AI during the drafting process, substantially revises AI-generated passages, and incorporates them into a completed manuscript. This article argues that when an author substantially and demonstrably revises AI-generated material, the resulting work should be presumed to be human-authored.
The Human-Authorship Requirement Under Existing Copyright Law
The constitutional and statutory foundation
Copyright protection in the United States is grounded in the Constitution’s grant of exclusive rights to “Authors,” a term that has traditionally been interpreted as referring to human creators. The Supreme Court addressed the role of human creativity in Burrow-Giles Lithographic Co. v. Sarony, when a lithography company argued that a photograph should not qualify for copyright protection because a camera simply captures what is placed in front of it. The Court rejected this argument and upheld Napoleon Sarony’s copyright, emphasizing the creative decisions he made in producing the photograph, including how he posed his subject, selected the costume and setting, and arranged the lighting. The camera itself only executed these choices. Thus, the photograph was protected not simply because a person operated the camera, but because the final expression resulted from Sarony’s own creative decisions.
The Copyright Office’s Human-Authorship Standard
The Copyright Office incorporated this requirement of human authorship into its formal registration standards through the Compendium of U.S. Copyright Office Practices, directing examiners to deny registration when a work is found not to have been created by a human. Courts have drawn the same distinction outside the context of AI. In Naruto v. Slater, the Ninth Circuit held that a monkey could not claim copyright over a selfie taken with an unattended camera, because copyright protection is limited to human authors, regardless of how the photograph itself was produced. The Copyright Office later applied this same principle, with little change, to works produced using generative AI.
Applying the Human-Authorship Standard to AI-Assisted Works
Zarya of the Dawn demonstrates how this standard applies when a work combines human and AI contributions, rather than involving a claim made entirely by a non-human creator. Kristina Kashtanova initially registered her graphic novel without disclosing that Midjourney had been used to generate its illustrations. After the Copyright Office became aware of this, it investigated the process behind the images and gave Kashtanova an opportunity to describe how she had created them. Her attorney argued that using Midjourney was comparable to using a camera or Photoshop, characterizing these technologies as tools through which human creative decisions are ultimately expressed in the finished image.
The Office did not accept this comparison. It distinguished photography from Midjourney by explaining that a photographer can control factors such as framing, lighting, exposure, and composition, all of which directly shape the final photograph. By contrast, a Midjourney user provides a text prompt but does not have the same ability to predict or control the particular image produced by the system. As a result, the Office maintained Kashtanova’s registration only for the parts of the work that she had created herself, including the written text and her selection and arrangement of the images within the completed book, while excluding the individual AI-generated illustrations from copyright protection. The Office also concluded that her later edits to those images could not be shown to involve sufficient creative input to receive copyright protection independently.
Judicial Confirmation of the Human Authorship Requirement
Thaler v. Perlmutter extended this issue to a situation in which no human had contributed to the creation of the work itself. Dr. Stephen Thaler attempted to register an image titled “A Recent Entrance to Paradise,” which he claimed had been produced entirely by an AI system known as the Creativity Machine. In his application, he identified the machine as the work’s sole author. The Copyright Office denied the application, and the district court upheld that decision, concluding that human authorship is a fundamental requirement under the Copyright Act. The D.C. Circuit affirmed the ruling in 2025, and the Supreme Court declined to hear the case in 2026. As a result, the requirement of human authorship remains established in copyright law, rather than being simply a policy adopted by the Copyright Office.
Why the Existing Framework Works in Clearly Defined Cases
The tool-versus-author approach works relatively well in Thaler and Zarya of the Dawn because both cases allowed the Copyright Office to identify a fairly clear boundary between human and machine contributions. In Thaler, there was no human creative contribution to consider alongside the AI-generated work; the analysis essentially ended with the fact that the claimed author was a machine. In Zarya of the Dawn, the contributions were also relatively easy to distinguish. Kashtanova wrote the text, chose and arranged the images, while Midjourney generated the individual illustrations. Each contribution could therefore be identified as belonging to a separate part of the final work. The binary approach is easier to apply when human and AI contributions are already clearly separated.
The Problem of AI-Assisted Drafting and Revision
A novelist who uses AI during the drafting process does not create a work that can be divided into clearly separate layers. She might ask an AI system to generate a scene, rewrite some of its sentences, remove entire passages, reorganize the remaining material, and repeat the process throughout the manuscript. By the end, the final prose may be neither entirely human nor entirely AI-generated. Instead, it reflects an ongoing process of human editing and decision-making applied to AI-generated material. Unlike the illustrations in Zarya of the Dawn, there is no distinct “AI layer” that can simply be removed from the completed book. The AI-generated material may have already been incorporated into sentences that the author has rewritten, rearranged, and modified.
Neither Zarya of the Dawn nor Thaler provides a clear answer for this type of situation. Zarya of the Dawn recognizes protection for human-created text and the selection and arrangement of images, but it does not address prose that originates from AI output and is later substantially rewritten by a human author. Thaler, meanwhile, deals with the opposite situation, where there was no human contribution to the work at all. Applying the existing framework to a blended manuscript would therefore require the Copyright Office to conduct a type of analysis for which its current standard provides little guidance. Examiners would potentially need to identify which parts of the manuscript originated with AI and then assess, sentence by sentence, whether the author’s revisions were sufficient to make each passage human-authored. The concept of “meaningful creative control” does not provide a clear method for doing this because it describes the level of human involvement in the final work without explaining how that involvement should be traced throughout the writing process.
The Practical Consequences of the Existing Gap
Leaving this gap unresolved is more than a theoretical problem. It creates genuine uncertainty for the authors who are expected to work within the existing framework. An author who has extensively revised a manuscript with the help of AI cannot know with confidence how the Copyright Office will ultimately treat the finished work. It may be considered fully protected, protected only in part, or partially unregistrable, since the current standard does not explain how human authorship should be identified when AI-generated text has been substantially revised. Publishers are left with a similar problem. A work accepted by one examiner as showing “meaningful creative control” could potentially face a different assessment from another examiner, or from a court, even where the underlying facts remain the same.
This uncertainty may also encourage authors to avoid disclosing their use of AI. An author who openly acknowledges using AI risks attracting the kind of scrutiny that led to Kashtanova losing protection for her illustrations, while an author who does not disclose that use may avoid such scrutiny altogether. This creates a tension with the Copyright Office’s own registration system, which depends on authors accurately identifying AI-generated material. If the existing framework cannot adequately address works involving both human and AI contributions, it creates more than a legal gap; it may encourage authors to be less transparent about the way their works were created.
III. A Presumption-Based Standard
The Proposed Presumption
The problem identified in Part III suggests the need for a standard that focuses on what an author does with AI-generated material, rather than simply asking whether AI was involved in creating the work. This Article proposes that substantial revision of AI-generated material should be sufficient to establish a presumption of human authorship. Where an author can demonstrate that she substantially revised AI output by rewriting its language, restructuring its content, and incorporating it into a broader manuscript, the finished work should be presumed to be human-authored. This would eliminate the need for the Copyright Office or a court to trace the origin of individual sentences throughout the manuscript.
This approach better reflects the way AI-assisted manuscripts are actually created. Instead of focusing on whether a particular passage originally came from an AI system, a question that may become difficult, if not impossible, to answer once the material has been extensively revised, the inquiry would focus on the author’s subsequent creative and editorial decisions. The central question would be whether those decisions were substantial enough for the resulting passage to be treated as the author’s own expression, regardless of how the initial material was produced. In this way, the proposed standard replaces an inquiry that is difficult to apply in practice with one that can be assessed through the author’s actual revision process.
Why Substantial Revision Should Establish Human Authorship
Basing the standard on the author’s revision of AI-generated material would not require a departure from established copyright doctrine. In fact, it is consistent with the Supreme Court’s reasoning in Feist Publications, Inc. v. Rural Telephone Service Co. There, the Court explained that originality requires only a minimal level of creativity; it does not require an author to produce something completely new or create the work entirely from scratch. What matters is that the author exercises some independent creative judgment in selecting, arranging, or shaping the material being used.
The same reasoning can apply to the revision of AI-generated text. When an author rewrites an AI-generated sentence, selects one version of a passage instead of another, or reorganizes AI-produced material to fit the rest of a manuscript, she is making the kind of creative choices that Feist recognizes as sufficient for originality. The only difference is the source of the initial material. Rather than working from facts, data, or material provided by another person, the author is working from text initially produced by an AI system. If copyright law does not require an author to create every element of a work from nothing, there is no clear reason to impose such a requirement simply because the initial material came from AI rather than from an editor, collaborator, or co-author.
Defining “substantial” and its evidentiary basis
A presumption is only useful if there is a clear basis for applying it, and “substantial revision” cannot simply depend on what an author says she did. This Article therefore proposes that substantial revision should be demonstrated through concrete and verifiable evidence rather than the author’s own description of her creative process. A manuscript’s draft history, for example, can demonstrate how much the final version has changed from the original AI output. Track-changes files and version-control records can provide even more detailed evidence by showing the specific words, sentences, and passages that the author changed. If these records are unavailable, the original AI output can be compared directly with the completed manuscript. The examiner or court would not need to determine the author’s reasons for making each change; the relevant question would be whether the final text has been substantially transformed rather than merely altered in wording or appearance.
This approach is consistent with the way courts already evaluate evidence of authorship in joint-authorship disputes, where objective evidence of creative contribution is generally more useful than a party’s own statement about what she intended to contribute. Requiring verifiable evidence would also prevent the term “substantial” from becoming as vague as “meaningful creative control,” which is precisely the uncertainty this Article seeks to address. If an author cannot provide any evidence showing how the final manuscript differs from the AI’s original output, there would be insufficient basis for applying the presumption of substantial revision.
Objections and Responses
Administrability
One possible objection is that courts and Copyright Office examiners may not be well equipped to determine how extensively an author revised AI-generated material. However, this type of assessment is not unfamiliar to copyright law. In fair use cases, courts already evaluate degrees of transformation rather than relying on a simple yes-or-no determination. Requiring authors to provide concrete evidence of their revisions, instead of relying solely on their own descriptions, would make the proposed standard similarly workable in practice.
Line-Drawing
Another concern is that the term “substantial” simply replaces the uncertainty of “meaningful creative control” with another vague standard. The difference, however, lies in what the standard is based on. Rather than leaving the examiner to make an open-ended judgment about the author’s level of control, the proposed rule connects substantial revision to evidence that can actually be examined, such as draft histories, tracked changes, and side-by-side comparisons of AI output and the final text. These records give the examiner a concrete basis for determining whether substantial revision occurred.
Doctrinal Fit
A final objection may be whether this presumption is consistent with Feist and Thaler. It is. Neither case addresses the situation in which AI contributes to the initial expression but a human subsequently makes substantial creative changes to that material. Thaler concerns a work produced without human authorship, while Feist establishes the relatively low level of creativity required for originality. Neither decision therefore resolves the intermediate situation addressed by this Article. The proposed standard fills that gap without altering the holdings of either case.
Conclusion
The current copyright framework can distinguish between a work created entirely by a human and one produced entirely by an AI system, but it provides much less clarity when the two contributions are intertwined. This is particularly important for authors who use AI during drafting and then substantially rewrite, reorganize, and develop the resulting material. This Article proposes that substantial and documented revision of AI-generated material should give rise to a presumption of human authorship, with the burden shifting to the party challenging that presumption. Such a rule would allow courts and the Copyright Office to focus on the author’s actual creative contribution to the finished manuscript rather than attempting to trace every sentence back to its original source. It would provide a more workable way to recognize human authorship in AI-assisted writing while remaining consistent with the basic principles of existing copyright law.
Reference(S):
U.S. Const. art. I, § 8, cl. 8.
Burrow-Giles Lithographic Co. v. Sarony, 111 U.S. 53 (1884).
Naruto v. Slater, 888 F.3d 418 (9th Cir. 2018).
Thaler v. Perlmutter, 687 F. Supp. 3d 140 (D.D.C. 2023).
Thaler v. Perlmutter, 143 F.4th 187 (D.C. Cir. 2025).
Thaler v. Perlmutter, No. 25-449, 2026 WL [pending] (U.S. Mar. 2, 2026) (cert. denied).
Feist Publ’ns, Inc. v. Rural Tel. Serv. Co., 499 U.S. 340 (1991).
U.S. Copyright Office, Compendium of U.S. Copyright Office Practices § 313.2 (3d ed. 2021).
Letter from Robert J. Kasunic, Assoc. Register of Copyrights & Dir. of Registration Policy & Pracs., U.S. Copyright Office, to Van Lindberg, Re: Zarya of the Dawn (Registration No. VAu001480196) (Feb. 21, 2023).
