Editor’s Note: This article is part of a collaboration between the Harvard Art Law Organization and the Harvard International Law Journal.
Yelena Ambartsumian* and Maria T. Cannon**
The Terms of Service of many generative artificial intelligence (“generative AI”) tools, particularly those that produce illustrations and images, require that the user grant the AI tool an “irrevocable copyright license” to both the inputs and the outputs. While the content that the user provides to the generative AI tool (the input), and the data sets on which the generative AI tool was trained (training data) may consist of copyrighted material, the same is not true for the output—for example, the image that the generative AI tool generates in response to a user’s prompt. For such content to be copyrightable, most jurisdictions, including the United States, require some level of human creativity or originality in the selection and/or modification of the AI-generated content. That means AI-generated work, alone—in response to a human user’s prompt—is not afforded copyright protection.
In its 2023 Rule on Works Containing Material Generated by Artificial Intelligence, the U.S. Copyright Office stated copyright protects “only material that is the product of human creativity.” In late January 2025, the U.S. Copyright Office published its highly anticipated report on the copyrightability of works created using generative AI. Far from signaling a departure, however, the Copyright Office maintains that there is no need for changes to legislation and that existing law can resolve questions of copyrightability and AI. While determinations are made on a case–by–case basis, the Copyright Office clarified that most prompt-engineering will not suffice: this is because, for copyrightability, a human must determine the elements of creative expression, and, currently, “AI systems are unpredictable” given that the same prompt can create various outputs.
The U.S. approach is similar to that of many but not all other jurisdictions. We provide below a comparative analysis on the copyright laws of the United States, EU, UK, China, and Japan. In short, while all jurisdictions require some level of human involvement, one seeking to copyright AI-generated work would have the highest chances of success in China or Japan. When comparing different approaches to copyright protection, it is important to remember that the rights and ownership of copyright matter most when there is an alleged infringement. This practical concern—coupled with the United States’ policy interest in maintaining a monopoly on producing and exporting creative and entertainment goods—means it soon may be time to reevaluate U.S. copyright law’s human authorship requirement, particularly as AI technologies rapidly develop.
United States
While the Copyright Office has granted registration to hundreds of works that incorporated AI outputs, where the applicant properly disclaimed the AI-generated content, it is not possible to copyright a work generated solely by AI today. As the Copyright Office recently explained, “prompts alone do not provide sufficient human control to make users of an AI system the authors of the output.” Copyright does not protect ideas but rather the creative expression of those ideas. Accordingly, as the Copyright Office reasoned, “prompts alone do not provide sufficient human control to make users of an AI system the authors of the output.” For copyrightability, some level of originality is a prerequisite, though the “level of creativity is extremely low,” and not to be confused with “sweat of the brow” or industrious collection. See Feist Publications, Inc. v. Rural Telephone Service Co., 499 U.S. 340 (1991).
The hurdle to copyrightability stems from the “authorship” requirement in U.S. copyright law, found in the Constitution and the Copyright Act, and as interpreted by the courts. The Constitution gives Congress the power to promote the useful arts, by giving “authors” the exclusive right to their “writings” (Article I, Section 8, Clause 8). The Copyright Act, first enacted in 1790, thus protects “original works of authorship” (17 U.S.C. § 102(a)). The Copyright Office views this authorship requirement as “[m]ost fundamentally . . . exclud[ing] non-humans.”
But AI is not the first technology that has required us to re-think authorship and human involvement. While lower court cases over a century ago often sought to limit that which was protected, two early and important Supreme Court cases evidenced an expansive approach. See Burrow-Giles Lithographic Co. v. Sarony, 111 U.S. 53 (1884) (Miller, J.) (holding photograph of Oscar Wilde a “writing” and photographer an “author”); Bleistein v. Donaldson Lithographing Co., 188 U.S. 239 (1903) (Holmes, J.; Harlan, McKenna, J., dissenting) (holding posters advertising a circus are copyrightable).
In Burrow-Giles Lithographic Co., the Supreme Court reasoned that while some photographs result from purely mechanical actions and thus lack authorship, other photographs are the product of an author’s “intellectual conceptions” and design. The Copyright Office has since relied on Burrow-Giles and subsequent case law to imbue humanity into the authorship requirement for copyright registration. In its 1965 Annual Report, the Copyright Office explained “[t]he crucial question appears to be whether the ‘work’ is basically one of human authorship, with the computer merely being an assisting instrument, or whether the traditional elements of authorship in the work (literary, artistic, or musical expression or elements of selection, arrangement, etc.) were actually conceived and executed not by man but by a machine.”
Accordingly, an artist that inserts a prompt into a generative AI model and receives a written, visual, or musical output in response is unlikely to have created a work capable of copyright protection. The Copyright Office would view this work as lacking “any creative contribution from a human actor.” (Although, in the United States, copyright is automatically secured upon creation of the work, registration is a prerequisite to filing suit.) But an artist that produces a work containing AI-generated material, which also required human involvement (by editing or modifying the output, combining the AI-generated elements with other elements, etc.) may have created an original work of authorship. Importantly, while the overall work may be protected (for example, a comic book, with its text and arrangement of elements), the individual AI-generated images within that work likely would not be, for now.
European Union
While no EU-wide unitary copyright exists, works receive protection according to the laws of the respective EU Member State. Currently, there is no prohibition on registering works made using AI as a tool (AI-assisted works). In fact, the recent EU AI Act does not directly address the question of registration of AI-assisted works.
On the subject of copyrightability of AI-generated works, there is little case law, apart from Infopaq International A/S v. Danske Dagblades Forening (Case C-5/08), in which the Court of Justice of the European Union (“CJEU”) held that copyright protection will only be available for works that are “the expression of the intellectual creation of their author.” What does this mean for the outputs of generative AI? In Infopaq, the CJEU suggests EU Member States should figure it out themselves (“[I]t is for the national court to make this determination”). Because the CJEU did not provide an exact formula, the States have some flexibility in interpreting and applying the law within their respective national frameworks. In a December 2024 policy questionnaire, the general view of the Member States was that AI-generated content could be eligible for copyright protection “only if the human input in their creative process was significant” (emphasis in original).
United Kingdom
The UK’s copyright laws have been shaped by EU harmonization, due to the UK’s nearly fifty-year membership in the EU, until 2020. Therefore, although early U.S. copyright law is rooted in the Statute of Anne (8 Anne c. 19, 1710), we see several departures between the American and UK systems today.
Per Section 1(1)(a) of the UK’s Copyright, Designs and Patents Act (1988) (CDPA), a literary, dramatic, musical, or artistic work must be an “original” authorial work. This originality requirement is interpreted in accordance with the relevant EU case law, including Infopaq and subsequent decisions, which hold originality is the “author’s own intellectual creation” and requires the author to make choices that “stamp the work created with their personal touch.” CDPA Sec. 9(1) also recognizes a separate category of works called “entrepreneurial works,” which include films, sound recordings and broadcasts; these works do not require originality to qualify for copyright protection but the term for their protection is shorter.
All of this said, the CDPA explicitly speaks to computer-generated works. Section 9(3) provides that for “literary, dramatic, musical or artistic work which is computer-generated, the author shall be taken to be the person by whom the arrangements necessary for the creation of the work are undertaken.” In 2021, after seeking public comment on whether computer-generated works should continue to be protected, the UK Intellectual Property Office elected to keep the law in place. Section 9(3) does not specify the originality required for computer-generated works. Future case law will hopefully resolve whether Section 9(3) simply designates the author—or owner—for such works that are entirely AI-generated and thus lack a traditional human author. (In that case, the person prompting the general-purpose AI tool simply would be the author.) Or, the courts may require the same originality as for other authorial works—“the author’s own intellectual creation”—which would be difficult to evaluate, particularly in the case of an entirely AI-generated work.
China
The Copyright Law of the People’s Republic of China approaches authorship through the lens of ownership, not originality. Historically, China’s Copyright Law lacked an “originality” requirement, to avoid confusion with the patent law requirement of “novelty” and “inventive step.” In 2002, the Regulations for the Implementation of the Copyright Law of the People’s Republic of China (amended 2013) introduced the concept of “originality” in Article 2 but with little guidance. Originality is interpreted through case law, with divergent interpretations by the courts but generally requiring that works be original, reflect intellectual achievement, and embody a concept of originality, among other factors—sometimes characterized as a “‘sweat of the brow’ plus” standard (effort and some creativity).
In Li v. Liu, (2023) Jing 0491 Min Chu No. 11279 (2023), the Beijing Internet Court unlocked the path for artists in China to obtain copyright protections for outputs of generative AI models. Critically, the Court relied on Article 3 of the Copyright Law to categorize an AI-generated image as a “wor[k] of fine art” and thus capable of copyright protection.
Li v. Liu involved a plaintiff who created a picture of a woman in springtime using an open source program called Stable Diffusion—a diffusion model which is trained on noising and denoising images, much like a human artist. The plaintiff exercised numerous choices in wording and phrasing when writing the prompt (including negative phrases, such as no “bad hands, text, error, missing fingers, extra digits”). He also adjusted the parameters to fine-tune the output. After the plaintiff posted the final image on social media, the defendant removed the watermark and published the same image in an article, on an alternate online platform, without obtaining permission or a license. The plaintiff sued for copyright infringement.
After determining that AI-generated images are fine art, the Beijing Internet Court focused on examining the work’s originality. It analyzed factors such as the specificity of the prompts, the actual descriptions of elements of the finished image that could be generated, and the unique formula and method the plaintiff applied to obtain the final result. Here, the plaintiff demonstrated a trial–and–error creative process well known to all artists. The Court found that the cumulative impact of the plaintiff’s choices caused the contributions to meet the threshold of “original,” as applied to creative works.
In the end, the Beijing Internet Court ruled in favor of the plaintiff and his AI-generated work, but with the following caveat: AI-generated works will only meet the qualifications for copyright registration in “appropriate” (read: not all) cases, and whether or not a work meets this threshold will be determined case–by–case. Despite the guardrails, this decision gives artists greater freedom to use generative AI to create copyrightable works in China than in the United States (and perhaps even in the EU and UK).
Japan
Japan amended its Copyright Act in 2018 in response to the development of new technologies. Notably, Article 30-4 of the Act gives broad rights to use copyrighted material for information analysis, including to train AI models for commercial use—so long as the use of the copyrighted works does not unreasonably prejudice the interests of the copyright owner. In May 2024, the Copyright Subdivision of the Cultural Council published guidelines on AI and copyright. The Copyright Subdivision noted that copyrightability of AI-generated material rests on whether the human author has provided “creative contributions that surpass mere effort.” Examples include the amount and specificity of the instructions or inputs, the number of generation attempts, and the selection from multiple output materials.
Conclusion
Regardless of the jurisdiction, most courts seem to grapple with the human’s role in the AI-generated output, rejecting that a simple prompt is enough to constitute authorship or originality. We may need to rethink human authorship, as the AI Revolution changes the paradigm of how humans can express their creativity. By encoding neural network patterns and decision-making processes in such a way that closely mimics human thought, AI-powered technology can create new categories of creative goods—which, as we saw in Li v. Liu, also require protection against theft, copying, and unfair competition (just as the photograph in Burrow-Giles and the circus poster in Bleistein).
Apart from numerous examples of technology assisting artists in creating new possibilities of creative expression and copyrightable works (e.g., the camera, Photoshop, or Adobe Illustrator), there is another precedent for expanding our notion of authorship: conceptual art. As in the dispute between Maurizio Cattelan and Daniel Druet—the wax sculptor and artist who painstakingly executed some of Cattelan’s most famous works—we accept that Cattelan is the author, based on his instructions (the prompt) to Druet. Similarly, putting aside the U.S. doctrine of “works made for hire,” we also accept that Jeff Koons is the artist, and not his dozens of studio assistants—many of whom are sophisticated designers and engineers (and some of which are robots). Today’s copyright law is able to ignore that the studio assistant inevitably leaves her touch, even when following the artist’s instructions (or prompt). Why is AI treated differently?
Ultimately, the copyrightability of AI-generated outputs will rest not on the law catching up to current events but on the technologies’ further development necessitating change. Advances in prompt engineering (i.e., adaptive prompting, human-in-the-loop) will inevitably allow humans to have more control over the creative expression produced by the AI tool. Soon enough, we will be prompted to rethink the human’s participation in creative expression and whether (and how) we want to incentivize and protect such creation.
*Yelena Ambartsumian is the Founder of AMBART LAW, a New York City law firm offering outside general counsel services to startups, with a focus on data privacy, AI counseling, and intellectual property. Prior to founding AMBART LAW, Yelena founded the art-tech startup Origen, a collection management and analytics platform for emerging contemporary art. Yelena is a certified Information Privacy Professional (CIPP/US) by the International Association of Privacy Professionals (IAAP), and a co-chair of IAPP’s New York KnowledgeNet chapter. She has also worked as General Counsel at an engineering consulting firm, a Senior Associate at a premier global law firm handling complex commercial litigation and regulatory investigations, and Of Counsel at an art and cultural heritage law boutique. Yelena is a frequent contributor to the leading arts magazine Hyperallergic, on topics including copyright and cultural heritage destruction.
**Maria T. Cannon is an Associate at AMBART LAW and is a certified Artificial Intelligence Governance Professional (AIGP) by the International Association of Privacy Professionals. Maria frequently writes on the intersection of art, law, and technology and has been published by the American Bar Association, and the New York State Bar Entertainment, Arts and Sports Law Section (EASL), among others. She is an attorney admitted in New York State, and prior to her work as an associate in AI counseling and data privacy, she worked as a student extern in the Legislative Drafting Division of the North Carolina General Assembly.