Back to Blog

Copyright in AI-Generated Content: Current Legal Status

April 26, 2026 Aruna 12 min read
copyright in AI-generated content — human authorship and AI copyright registration

Artificial intelligence has fundamentally challenged one of copyright law’s most basic assumptions: that creative works are authored by human beings. As AI systems generate images, music, text, and code with increasing sophistication, questions about copyright in AI-generated content have become urgent legal and business issues. The answer affects creators who use AI tools, companies that develop AI systems, and anyone who commissions or purchases AI-generated works. PerspireIP tracks rapidly evolving developments in copyright law as it applies to AI-generated content and helps clients position themselves for the emerging legal landscape.

The Human Authorship Requirement

US copyright law has long required that protectable works be the product of human authorship. The Copyright Office’s Compendium of US Copyright Office Practices explicitly states that the Office will not register works produced by a machine or mechanical process that operates randomly or automatically without any creative input or intervention from a human author. This principle has been tested and upheld in court. In Thaler v. Perlmutter, the D.C. District Court affirmed the Copyright Office’s refusal to register an image autonomously generated by an AI system called DABUS, holding that human authorship is an indispensable element of a valid copyright claim.

The US Copyright Office has issued guidance specifically addressing copyright in AI-generated content. In its February 2023 policy statement on AI-generated works and its March 2023 guidance on works containing AI-generated material, the Office clarified its approach. Works that are entirely generated by AI without human creative control cannot be registered. However, works that contain AI-generated elements combined with human authorship may be registrable, but only with respect to the human-authored portions. The Copyright Office also announced a major study of copyright and artificial intelligence to inform future policy, with interim reports covering training data, copyrightability, and licensing issues.

The Human Creativity Spectrum

The copyright analysis for AI-generated content exists on a spectrum based on how much human creative control went into the work. At one end, a user types a simple text prompt and accepts whatever the AI generates. At the other end, a human artist uses AI tools as a sophisticated brush, making hundreds of creative selections, iterations, modifications, and additions. The Copyright Office and courts are working through where on this spectrum sufficient human authorship exists to support a copyright claim. Factors that may weigh toward protectability include the specificity and creativity of prompts, human selection and arrangement of AI outputs, significant human modification of AI-generated material, and integration of AI elements into a larger human-authored work.

The Zarya of the Dawn Case

One of the most instructive early cases involves the graphic novel Zarya of the Dawn, which was partially created using Midjourney AI image generation. The Copyright Office initially registered the work, then reconsidered, ultimately granting copyright only in the text and the selection and arrangement of the work, while refusing protection for the individual AI-generated images themselves. This case illustrates the Copyright Office’s current approach of parsing AI-generated works and granting copyright only to the human-authored portions, treating AI-generated elements as unprotectable.

The same parsing approach governs commercial design output.

For a practical walk-through of who owns an AI-generated logo, product rendering, or design deliverable — and how to document the human contribution that makes it registrable — see our guide to AI copyright law and design ownership.

Business Implications of the AI Copyright Gap

The current state of copyright in AI-generated content creates significant business uncertainty. Companies investing in AI-generated marketing content, product designs, music, or software face the risk that their investments in AI-generated outputs cannot be protected by copyright, leaving them available for competitors to copy freely. This affects business models, valuation of AI-generated assets, and investment decisions. Some companies are responding by ensuring that human creative involvement is documented throughout AI-assisted workflows, making the human contribution explicit and substantial enough to support copyright claims.

Separate from the output question, copyright in AI-generated content is also implicated on the input side. Training large AI models on copyrighted works without license is the subject of numerous ongoing lawsuits. Authors, artists, musicians, news organizations, and software developers have filed class action claims arguing that using their copyrighted works to train AI models constitutes copyright infringement. These cases are working through the courts and will have profound implications for how AI systems can lawfully be trained. Depending on their outcome, AI developers may be required to obtain licenses for training data, which could significantly affect the development and cost of AI systems.

Practical Strategies for AI Content Creators

  • Document your creative process when using AI tools, including prompts, iterations, selections, and modifications
  • Add substantial human creative elements to AI-generated base material to strengthen copyright claims
  • Be transparent with the Copyright Office about AI involvement when registering works
  • Consider whether trade secret protection is available for valuable AI-generated works that cannot be copyrighted
  • Consult PerspireIP before investing heavily in AI-generated content assets to understand the current protection landscape
  • Monitor ongoing Copyright Office rulemaking and court decisions for developments that may affect your strategy

International Perspectives on AI Copyright

Other countries are taking different approaches to copyright in AI-generated content. The United Kingdom has a specific provision in the Copyright, Designs and Patents Act that grants copyright in computer-generated works to the person who made the necessary arrangements for the work’s creation, potentially covering AI-generated outputs. China has courts that have found copyright in AI-generated content under certain circumstances. The European Union is grappling with these questions in the context of its AI Act and ongoing copyright reform discussions. As the global legal landscape evolves, international copyright strategy for AI-generated works will become increasingly complex.

Thaler v. Perlmutter and the Statutory Basis for Human Authorship

The question of copyright in AI-generated content now has an appellate answer in the United States. In Thaler v. Perlmutter, No. 23-5233 (D.C. Cir. Mar. 18, 2025), the Court of Appeals for the D.C. Circuit affirmed the Copyright Office’s refusal to register “A Recent Entrance to Paradise,” an image generated by Stephen Thaler’s Creativity Machine and submitted with the machine named as sole author. The Supreme Court subsequently declined to review the decision.

What makes the ruling durable is its reasoning. The court did not rest on policy or on the Office’s expertise; it held that human authorship is required as a matter of statutory law, reading the Copyright Act’s machinery — the term of protection measured by an author’s life, the transfer of ownership by inheritance, the signature requirements — as provisions that presuppose a human being. That lineage runs back to Burrow-Giles Lithographic Co. v. Sarony, 111 U.S. 53 (1884), where the Court upheld copyright in a photograph precisely because a human had conceived and arranged it.

The holding is narrower than the headlines suggested. Thaler decides that a machine cannot be an author. It does not decide how much human contribution is enough when a person uses a generative tool, which is the question almost every real applicant actually faces.

The Copyright Office has been working through copyright in AI-generated content in a three-part study. Part 1, on digital replicas, was published in July 2024. Part 2, on copyrightability, followed in January 2025 and is the one that governs registration practice. Part 3, addressing the use of copyrighted works in generative AI training, was released in pre-publication form in May 2025.

Part 2’s conclusion is more permissive than it is often reported to be. The Office confirmed that existing law is adequate and that no legislative change is needed, and it drew the line in three places. Prompts alone, however elaborate or iterated, do not confer authorship, because the user does not control how the model converts the prompt into expressive output. Human selection, coordination and arrangement of AI-generated material can be protected, as a compilation, even where the underlying elements are not. And human-authored material that is perceptible in the final work remains protected regardless of what the model contributed around it.

Two registration decisions illustrate the line. In Zarya of the Dawn (February 2023), the Office cancelled and reissued the registration for a graphic novel so that it covered the author’s text and the selection and arrangement of the images, but not the individual Midjourney-generated images themselves. In Théâtre D’opéra Spatial (September 2023), the Review Board refused registration of Jason Allen’s prize-winning image outright, notwithstanding that he reported using more than six hundred prompt iterations. Volume of prompting is not authorship.

Disclosing AI Material in a Registration Application

Since the Office’s guidance on works containing AI-generated material, published at 88 Fed. Reg. 16190 (March 16, 2023), applicants have had an affirmative duty to disclose. More than a trivial amount of AI-generated content must be identified and excluded from the claim, using the Limitation of Claim space to disclaim it, and the human contribution must be described briefly.

  • Identify which elements were generated by a model and which were authored, modified or arranged by a person.
  • Disclaim the AI-generated elements rather than staying silent about them.
  • Describe the human contribution concretely — selection, arrangement, substantial editing — not as “creative direction”.
  • Keep contemporaneous records of the human editing process, because the burden of showing authorship falls on the applicant.

The stakes for getting this wrong are not merely a refused application. A registration obtained without disclosing AI-generated material can be cancelled, and a knowing misstatement in an application can be raised to defeat the registration’s evidentiary value in later litigation. For anyone building a commercial catalogue, the disclosure discipline matters more than the marginal scope it costs.

Training Data: The Second Front

The output question sits alongside an input question that is being litigated in parallel: whether training a model on copyrighted works is itself infringement, and whether fair use excuses it. The two are analytically distinct — a work can be unprotectable as AI-generated output while the model that produced it was trained in a way that infringes.

The first substantive merits ruling came in Thomson Reuters v. Ross Intelligence (D. Del., February 2025), where the court rejected a fair use defence for the use of Westlaw headnotes to train a legal research tool, weighing heavily that the tool was non-generative and competed directly with the plaintiff in the same market. Subsequent decisions involving generative models have treated training as more transformative while distinguishing sharply between lawfully acquired copies and pirated corpora — a distinction that has driven settlement value more than the fair use holding itself.

For businesses, the practical exposure is indemnity rather than doctrine. Read the provider’s terms: whether the vendor indemnifies output-based claims, what it warrants about training data provenance, and whether the indemnity survives your fine-tuning of the model.

How Other Jurisdictions Diverge

The American human-authorship rule is not universal, and multinational publishers cannot assume a single answer on copyright in AI-generated content.

The United Kingdom is the clearest counterexample. Section 9(3) of the Copyright, Designs and Patents Act 1988 provides that for a literary, dramatic, musical or artistic work that is computer-generated, the author is taken to be the person by whom the arrangements necessary for the creation of the work are undertaken; section 12(7) gives such works a term of fifty years from creation rather than life plus seventy. A work refused registration in Washington may therefore be protected in London, on a shorter clock and with authorship attributed to the arranger.

China has moved in a third direction. In Li v. Liu (Beijing Internet Court, November 2023), the court found an AI-generated image protectable where the user’s selection of prompts and parameters and iterative adjustment reflected sufficient intellectual investment — effectively the opposite conclusion to Théâtre D’opéra Spatial on comparable facts. For anyone distributing generated material internationally, the safe planning assumption is that protection varies by territory and that the human-contribution record is the asset worth preserving in all of them.

What Copyright in AI-Generated Content Means for Your Contracts

The doctrine matters far less to most businesses than the paperwork built on top of it. If a deliverable turns out to be substantially unprotectable, the loss lands wherever the contract put it — and most template agreements were drafted on the assumption that everything created under them would be copyrightable.

Work-made-for-hire clauses are the first casualty. A work-for-hire provision, and the assignment that normally backs it up, can only transfer rights that exist. Where a vendor delivers substantially machine-generated material, the clause conveys a much thinner bundle than the client believes it bought, and no amount of assignment language creates authorship that the statute does not recognise. The practical fix is not better assignment drafting but a representation about how the work was produced.

  • Disclosure obligation. Require the contractor to disclose which deliverables used generative tools and to what extent, rather than warranting generally that the work is original.
  • Human-contribution records. Require retention of drafts, edits and version history. Where protection depends on selection, arrangement and modification, the evidence of that contribution is the asset.
  • Registration responsibility. Say who applies, and who bears the duty to disclaim AI-generated material correctly. A defective registration is worse than none.
  • Fallback protection. Where copyright may be thin, lean on contractual exclusivity, confidentiality and trade-secret protection, which do not depend on authorship at all.
  • Indemnity scope. Check whether the vendor’s AI provider indemnifies output claims, and whether that indemnity survives your own fine-tuning or downstream modification.

Internally, the same logic applies to employee policy. Staff using generative tools should know that unreviewed output may be unprotectable and that pasting confidential material into a third-party model can forfeit trade-secret status independently of any copyright question. A short written policy naming approved tools, prohibited inputs and the record-keeping expectation is cheap, and it is the control that auditors and acquirers now ask to see.

None of this requires a settled answer on copyright in AI-generated content. It requires accepting that the answer is currently uncertain and allocating that uncertainty deliberately, rather than leaving it to be discovered during diligence.

Conclusion

Copyright in AI-generated content is one of the most rapidly evolving areas of IP law. The current US framework requires human authorship, leaving purely AI-generated works outside copyright protection, but significant questions remain about the degree of human involvement required and how to evaluate the creative contributions of complex human-AI collaborative processes. PerspireIP helps clients navigate this uncertain landscape, structure AI workflows to maximize copyright protection, and develop comprehensive IP strategies for AI-generated assets.