Copyright does not attach to a work an AI made on its own. Thaler v. Perlmutter (130 F.4th 1039, D.C. Cir. 2025) and the Supreme Court's cert denial of March 2, 2026 (No. 25-449) settled that. The live frontier question runs the other way: where is the authorship boundary for a work a human made using AI as a tool? Jason Allen entered and revised text prompts at least 624 times to produce Théâtre D'opéra Spatial — is he its author? Allen v. Perlmutter, No. 1:24-cv-02665-WJM (D. Colo., Martínez J.), is preparing the first judicial answer. Briefing closed on February 27, 2026, and a decision is expected.
Two Doctrinal Propositions
First, the Office's "re-rolling the dice" framing misreads the learning-responsive expressive control of prompt-engineering. Copyright and Artificial Intelligence, Part 2: Copyrightability (Jan. 29, 2025), at page 20, treats revising and resubmitting prompts as re-rolling dice that yields more outputs to choose among without changing the degree of control. That reduces a prompt to a draw from a fixed probability distribution. But a skilled user learns how the model reads language and tunes the next prompt to it — closer to responsive sculpting than to gambling. Edward Lee's amicus brief (ECF No. 47, Sept. 3, 2025) puts the critique at its sharpest: the metaphor "confuses more than it illuminates," implying randomness when the Office itself disclaimed outcome predictability as the relevant inquiry. The metaphor and the doctrine sit in contradiction.
Second, that framing manufactures a new duty of articulation imposed on AI works alone. Allen disclosed the categories of his prompts — overall subject, tone, composition, and the like (AR_007) — but kept the prompt text itself secret, arguing that disclosure would be like "explaining a magic trick" that loses its effect once known (AR_034). The government's brief (at 31) turns that nondisclosure into the ground for denial: without the prompts in the record, expressive choice cannot be assessed. Yet a photographer never files the instant of a viewfinder decision. The AI user is compelled to surrender the articulation of the process. Although Feist set the creativity bar "extremely low" (499 U.S. at 345), AI works are made to clear a higher threshold of evidentiary articulation. Whether that asymmetry is justified is the real question in Allen.
The Trans-Pacific Divide
The convergence on display in Thaler becomes divergence in Allen. China's Li v. Liu (Beijing Internet Court, Nov. 27, 2023) treats a human user as an author once intellectual investment is shown in the record — Li entered more than 150 prompts to Stable Diffusion and submitted his parameter adjustments. In September 2025 the Beijing Internet Court codified that burden of proof in its Eight Typical Cases Involving Artificial Intelligence. The EU's Infopaq–Painer standard of free creative choices has not met an Allen-equivalent head-on, but its direction resembles China's. The UK is moving from the deemed authorship of CDPA 1988 s. 9(3) toward the US and EU — a DSIT report of March 18, 2026, proposed removing s. 9(3). Only the US in Allen leans toward denying authorship even when the articulation duty is met, because a prompt is treated as an instruction at bottom.
How Industry Operationalizes the Divide
The doctrinal divide operates as a recording duty in industry workflows. The C2PA provenance standard (Adobe, Microsoft, OpenAI, and others) has become a de facto industry standard, and the US and EU voluntary standard converges with China's judicial guidance on a provenance-documentation duty. A Single Piece of American Cheese (USCO registration, Jan. 30, 2025) set the disclaimer model for hybrid works. Vendor indemnification — the Microsoft Copilot Copyright Commitment, the Anthropic Customer Copyright Commitment, and others — absorbs the risk that outputs carry no IP. The market splits bimodally: commodity content moves to AI-default with no IP, while premium creative work stays human-led and keeps its IP value.
The Test
Three tests set the provisionality of any ruling. Tenth Circuit appeal — Meshwerks v. Toyota (528 F.3d 1258, 10th Cir. 2008) (Gorsuch, J.) is the circuit's camera-copyright framework, the government's anchor. The next case — if a user who documented and articulated the prompts faithfully is still refused under the "re-rolling" logic, the Office concedes that articulation itself is meaningless. Office governance — the Perlmutter removal dispute (SCOTUS stay deferral, Nov. 26, 2025) and the Justice-Department-only signature on the brief (ECF No. 57) hint at daylight between the government and the Office.
Allen justified his secrecy with the magic-trick analogy, and in doing so decided his own case. Keep the magic as magic, and you forfeit the copyright. Read correctly, Allen is not a case about whether AI works can be authored — it is a case about zoning: about which procedural duties a user must discharge before a human-plus-AI work may enter the copyright system at all. If Thaler zoned AI-alone outputs as uncopyrightable, Allen zones the articulation duty for human-plus-AI hybrids. The second instance of law shaping an industry.