German Court Sets Twin Test for AI Images: No Copyright Without Human Creativity, No Infringement Without Copied Specifics

A newly published decision from the Higher Regional Court of Düsseldorf (Oberlandesgericht, OLG) is shaping up to be the most concrete European answer yet to two questions every working photographer, illustrator, and AI product team has been asking. When does an AI image become a protected work? And when does feeding someone else's photo into an AI system cross the line into infringement?

In its decision of April 2, 2026 in case I-20 W 2/26, now widely reported in English-language coverage, the 20th Civil Senate of the OLG Düsseldorf gave a striking pair of answers. Pure machine output is not copyrightable absent meaningful human creative control. And the abstract motif of a photograph — the idea of what is depicted — sits in the public domain and can be reconstructed by generative systems, so long as the original's specific creative elements are not copied.

The Underwater Dog Case

The dispute came out of a niche corner of commercial photography. A photographer specializing in dogs diving underwater for toys had built a business around the carefully lit, post-processed images. A former cooperation partner uploaded one of her elaborately edited shots into an image-to-image AI system and posted the resulting comic-style output on its website. The new image kept the basic idea — a dog underwater reaching for a red toy — but rendered it in a clearly stylized, cartoon-like aesthetic.

The photographer sought interim relief, arguing the AI image infringed her copyright. The Regional Court of Düsseldorf rejected the urgent application as a permissible free adaptation. On appeal, the OLG reached the same outcome, but it substantially rewrote the legal reasoning along the way — and that is what makes the ruling significant for the wider industry.

The First Test: When Does an AI Image Become a Protected Work?

The OLG was direct. Copyright protection for AI-generated content is not categorically excluded, but the personality of the human user has to be visible in the final image. The court identified three pathways that can satisfy that requirement:

• Extremely detailed pre-settings that go beyond generic instructions.
• Continuous, highly specific corrections during prompting.
• A conscious, creative selection from a wider set of intermediate results.

What is not enough, according to the senate, is feeding the AI vague text instructions and accepting whatever the system spits out. That, the court said, is consumption of a result, not creation of a work. Anyone who later wants to claim copyright in court bears the full burden of presentation and proof of which creative decisions and prompts produced the image.

That procedural detail did the actual work in this case. The respondent could not — or would not — explain his prompts and choices when the court asked. As a result, the AI graphic was denied recognition as a work, which blocked the route to a so-called free adaptation defense.

The Second Test: Element-by-Element, Not Overall Impression

The reason the photographer still lost is the second, arguably more consequential half of the ruling. The OLG explicitly stepped away from the traditional German test that compared the visual overall impression of two images. In its place, the senate adopted an element-oriented analysis aligned with recent guidance from the European Court of Justice.

Under this approach, the question is not whether the AI image looks like the photograph in some general sense. The question is whether the creative elements that constitute the photographer's personal achievement — camera settings, specific lighting, depth of field, framing, sharpness, the way the post-processing was rendered — were adopted exactly. The motif itself, a dog underwater reaching for a toy, is treated as a mere idea and is not protectable.

Because the AI graphic deviated from the original in perspective, anatomy, and dynamics, the OLG could not identify a legal infringement. The court found the comic-style version picked up only the unprotected idea, not the photographer's protected execution.

What the Ruling Actually Changes

For photographers, the take is mixed. They retain strong protection against identical or near-identical reproduction of their technical and creative choices. But they have to accept that AI-assisted reconstruction of an idea — same subject, same setup, different rendering — is generally lawful in Germany under this framework.

For AI users and platforms, the practical message is twofold. Image-to-image imitation of an existing photograph is not automatically infringement, but it is not a blank check either. If the output retains specific protected elements, the analysis can flip. And if a user later wants to claim copyright over the result themselves, vague prompts and lucky outputs will not survive judicial scrutiny — the court will ask for receipts.

For the broader European doctrine, the ruling is a meaningful data point. It systematically builds on the recent CJEU emphasis on recognizable adoption of specific creative elements, and it dovetails with earlier German decisions denying copyright to AI-generated logos and similar machine-only output. It also sits comfortably alongside the position taken by the U.S. Copyright Office, which has consistently held that even intricate prompts are not enough to confer authorship over AI-generated content without sufficient human creative contribution.

How It Lines Up With the Wider AI Copyright Landscape

The decision does not directly address the most-litigated question in U.S. courts — whether using copyrighted works as training data is itself infringing. That fight is playing out in cases like Bartz v. Anthropic, Nazemian v. NVIDIA, the publisher class action against Meta over Llama, and a long list of suits against OpenAI. The OLG's holding focuses on the output side: whether a generated image infringes a specific input photograph and whether that generated image is itself a protected work.

Even on that narrower question, the ruling matters. It signals that European courts are converging on a relatively coherent test that distinguishes between protected expression and unprotected idea, and that places a real evidentiary burden on anyone claiming authorship of AI-generated content. Expect to see this decision cited the next time a German court has to deal with an image-to-image dispute, and expect platforms operating in the EU to lean on its reasoning when defending against takedown demands tied to motif-level similarity.

Open Questions

A few issues the OLG did not resolve are worth flagging. The ruling does not address training-data legality under the EU's text-and-data-mining exceptions. It does not address the platform's potential responsibility for outputs that do retain protectable elements. And it leaves open how a court will assess hybrid workflows — say, an AI-assisted base image that a human then meaningfully reworks. Those questions will move through the German and EU court system over the next year, and after this decision, the framework for analyzing them is noticeably sharper.

Case reference: OLG Düsseldorf, decision of April 2, 2026, file number 20 W 2/26 (sometimes cited as I-20 W 2/26).