AI Art Copyright: Thaler v. Perlmutter and Human Authorship Requirement

Thaler v. Perlmutter is the leading US case on whether AI-generated works qualify for copyright protection. Stephen Thaler, a computer scientist, used his Creativity Machine neural network to generate an image titled "A Recent Entrance to Paradise" and tried to register it with the US Copyright Office (USCO), listing the AI as sole author and himself as owner under work made for hire. The USCO refused. Thaler sued; Judge Beryl Howell of the DC District Court affirmed the refusal in August 2023, and the DC Circuit unanimously affirmed on 18 March 2025. The US Supreme Court denied certiorari in 2026, ending the case. The courts read the Copyright Act of 1976 as requiring a human author: although the statute does not define "author," provisions on ownership, transfer, duration, termination, and nationality only make sense if the author is a human being. The DC Circuit stressed it was deciding the narrow question Thaler presented — a work that he himself stipulated was autonomously generated by a machine with no traditional human authorship. The court explicitly left open how copyright should treat works where a human uses AI as a tool. The USCO's March 2023 Statement of Policy on works containing AI-generated material formalized the same human-authorship line. Applicants must disclose which portions were generated by AI and disclaim them; USCO will register only the human-authored elements — selection, arrangement, modification, or original text written by a person. The agency's Zarya of the Dawn decision (February 2023) was the template: Kris Kashtanova's Midjourney-illustrated comic kept copyright for the human-written story and the human selection/arrangement of panels, but the individual AI images were excluded. The USCO doubled down in its January 2025 USCO Copyright and AI Report Part 2 Copyrightability, concluding that prompts alone — even long, detailed, iterative ones — do not give a user enough creative control over the output to make the user its author, because generative models fill in too much expressive content the prompter did not specify. Human selection and arrangement, meaningful modification, or human-authored elements inserted into AI output can still be protected. Grey-zone questions get resolved case by case. What counts as substantial human input? Examples the USCO treats as protectable: a human artist heavily reworking an AI draft, using AI for one layer of a larger composition, or arranging many AI outputs into an original whole. Examples treated as unprotectable: clicking re-roll until a prompt produces something pleasing, or curating a single best generation. Tools that give users tighter control — inpainting, ControlNet, sketch-to-image, img2img from a human drawing — push toward protectable authorship; pure text-to-image with default sampling pushes the other way. The ruling has commercial bite. Companies producing AI-generated marketing assets, stock images, music, code, or video without meaningful human shaping cannot sue infringers under US copyright law — the work falls into the public domain by default. Studios and publishers responded by documenting human-in-the-loop workflows and prompt histories, and by treating raw model output as draft material that humans must edit before publication. AI-generated code, music, and synthetic voices face the same question; GitHub Copilot suggestions, for instance, are treated by most legal commentators as uncopyrightable on their own, with protection attaching to the surrounding human-authored code. Thaler's earlier DABUS cases (named for his Device for the Autonomous Bootstrapping of Unified Sentience) tested the parallel question in patent law and lost in every major jurisdiction. The USPTO, EPO, UKIPO, and IP Australia all rejected DABUS as a named inventor. The US Supreme Court denied cert in April 2023, and the UK Supreme Court ruled unanimously in Thaler v Comptroller-General ([2023] UKSC 49) on 20 December 2023 that only a natural person can be a patent inventor. Courts in both regimes were careful to say AI-assisted inventions are still patentable — only AI-as-sole-inventor is barred. Internationally, the picture is splitting. China's Beijing Internet Court took the opposite path in Li v. Liu (November 2023): a user who shaped a Stable Diffusion image through ~150 prompts, parameter tuning, and iterative selection was deemed the author under Chinese Copyright Law, satisfying the "intellectual achievement" and originality tests. The UK has long had Section 9(3) of the CDPA 1988, which assigns authorship of computer-generated works with no human author to "the person by whom the arrangements necessary for the creation of the work are undertaken" — a 1988 provision designed for an earlier era of computing whose application to modern generative AI is contested. The EU has no harmonized rule on AI authorship yet; the The EU AI Act: The World's First Comprehensive AI Regulation sits alongside copyright rather than replacing it, and its Article 50 transparency obligations (effective August 2026) require machine-readable marking of synthetic content and clear deepfake disclosure but do not address copyrightability. The practical takeaway: in the US, AI output is uncopyrightable unless a human contributed expressive choices that go beyond prompting. Document the human contribution, register only the protectable elements, and assume raw generative AI output is freely copyable by anyone.