US Supreme Court: AI-only works have no author. The fight continues

Thesis: Lack of protection for fully AI-generated works is consistent, but it settles nothing about market value. The real dispute moves up to whether training on others' creative output is legal. The US and China rule in opposite directions.

On 2 March 2026, the US Supreme Court declined to hear the appeal of scientist Stephen Thaler regarding copyright for an image generated by his AI system. Thaler had applied to register "A Recent Entrance to Paradise" in 2018. He previously lost before the copyright office, which ruled in 2022 that a work must have a "human author". A federal judge in Washington called human authorship a "fundamental requirement" of protection in 2023. The DC Circuit Court of Appeals upheld that decision in 2025.

The ruling closes this thread but does not end the debate, for two reasons. First, the office consistently refused registration for Midjourney images when authors claimed "co-authorship with AI". The line is clear: no human, no right. Second, fully AI-generated output is a marginal part of the market. Value sits in the training data, not the result.

This is where the split appears, and nobody in Washington will settle it. A Chinese court in Wuhan held that the creator of AI-generated content can hold copyright if they controlled and anticipated the result by writing the prompt and choosing parameters. It was the first such case in that city. This doctrine is the opposite of the American one: what counts is not who physically drew the pixels, but who commanded the process. The two largest AI markets start from different definitions of authorship. They will eventually have to explain what that means for works crossing borders.

The European Union addresses the same problem elsewhere, in data protection law. The compromise from the Irish presidency on 3 September 2026, revealed by Politico, revives the idea of treating the use of personal data for AI training as a "legitimate interest" by default. Germany pushed for this solution, arguing it would help European companies. Max Schrems of noyb calls it a "digital expropriation of Europeans" and the solution a "complete opening of European data to global corporations".

Three models emerge. The American model requires a human, and the decision belongs to that human. Defendant companies conceded this, admitting they trained on protected material. The Chinese model counts control over the process. The European model, instead of ruling on authorship, tries to open input data wide. Three legal philosophies in one global value chain create a recipe for regulatory arbitrage. Companies will train where it is cheapest and sell where it is dearest, provided the courts keep up with the technology.