EU courts say AI output isn't copyrightable. That doesn't mean it's free to copy.

Three German court rulings this year rejected copyright claims on purely AI-generated works, and the European Parliament adopted a 10 March 2026 resolution on AI and the creative sector that points the same direction. The rule: Article 2 of the EU Copyright Directive (the bloc's main copyright harmonization law) ties protection to "the author's own intellectual creation," and a prompt alone does not meet that bar. The decision is narrow. The discourse has flattened it into a half-truth.

The half-truth is that "no copyright" on AI output means the output is unrestricted. The Hacker News discussion of an @maxpool post on the EU position kept returning to a version of that claim, and a common counter-argument ran: if no one owns the output, the output must be in the clear. That argument collapses two distinct copyright questions. The first is whether the AI's output gets its own protection. In the EU the answer is no, absent a qualifying human author. The second is whether the output infringes someone else's rights. That depends on the input, not the output.

The German line is the cleanest example. The Two Birds analysis of three 2026 decisions, on AI-generated photographs, text, and music, draws the same line each time: prompts, parameter tweaks, and post-hoc curation are not creative contribution in the legal sense, and the output lands in the public domain only because no one authored it. The Morgan Lewis writeup reads the rulings as a unified doctrinal turn. The PhotoWorkout coverage of the Oberlandesgericht Düsseldorf ruling (case I-20 W 2/26, April 2 2026) makes the point concrete: a photographer who used AI to assist a real shoot kept her copyright; a user who prompted an image into existence and stopped there did not.

The same logic lands outside Germany. The European Audiovisual Observatory's IRIS database treats output-side non-protection as a settled feature of EU doctrine, paralleling (not identical to) the US position. The Jones Day review of the European Commission's 2025 report on generative AI and copyright reaches the same conclusion: the human-authorship requirement is doctrinally firm, and the wiggle room sits in what counts as a human creative act, not in whether AI involvement disqualifies protection outright. The Wisemen analysis of Dutch case law reaches the same point: the Court of Justice of the EU's originality threshold applies, and Dutch courts have refused protection to purely model-generated images.

Training a model on copyrighted photographs without a license is a separate infringement, governed by the European Parliament's March resolution and pending national implementation. Translating a copyrighted novel with a model does not launder the underlying rights; the model is a tool, not a co-author. A platform that hosts AI-generated images does not escape the upstream rights of the works those images were trained on, reworked from, or visually resemble. Until a harmonized EU instrument lands, the answer to "can I copy this AI image?" is the same as the answer to "can I copy any image?": check the upstream rights, not the output.