A German court ruled that an AI-generated comic version of a dog photo did not infringe the photographer's copyright. That's not the interesting part. The interesting part is why: the court could tell exactly which elements of the original the AI version kept, and which it changed — because the case gave it enough to work with. Most disputes won't be so lucky, and neither will most creators.
What the Ruling Actually Decided
The Higher Regional Court of Düsseldorf (Oberlandesgericht Düsseldorf, case I-20 W 2/26) dismissed a photographer's application for a temporary injunction over an AI-generated comic-style version of her photo of a dog underwater with a red toy. A former collaborator had uploaded her photo to an AI tool, prompted it to generate a new image from it, and published the result.
The court's reasoning followed a standard tool of copyright analysis called the idea-expression dichotomy: copyright protects the specific expression of an idea, not the idea itself. The AI output had transformed the protected elements — framing, angle, lighting, sharpness — while the similarities were limited to the subject matter itself: a dog underwater with a toy. Being able to recognize the same dog and setting wasn't enough on its own to show copying; the dog, the toy, and the pose are the unprotectable idea layer, freely available to anyone, AI-assisted or not. The AI-generated comic changed the expression enough — while keeping the idea — to count as a new work rather than an infringing copy.
There's a second finding that's easy to miss and often gets dropped from summaries of this case: the AI-generated version wasn't found to be protected either. The court wasn't handing out a copyright to the AI output — it was declining to find infringement of the original. Those are two different questions, and conflating them is the single most common way this kind of ruling gets oversimplified on social media.
The Question Courts Are Actually Asking
Strip away the “AI” framing and this is a burden-of-proof case, not a technology case. The photographer had to show what was hers in the original — specific, demonstrable creative choices — for the court to weigh against the new work. Where a claimant (or a defendant) can point to concrete evidence of a specific creative decision, courts have something to reason about. Where they can't, courts default to the unprotectable idea layer, and the claim collapses.
This is the same test showing up in two other recent German rulings. The Local Court of Munich (142 C 9786/25, February 2026) rejected copyright protection for three AI-generated logos: the prompts behind them were “overly generic,” so the court attributed the creative output to the AI rather than the human prompter. It left the door open for a different outcome — human contribution can come through prompting or through subsequent editing — but only where that contribution is clear, objectively recognizable, and decisively shapes the result.
The Regional Court of Frankfurt (2-06 O 401/25, December 2025) put the burden-of-proof question even more directly. In a dispute over AI-assisted song lyrics, the court held that once a defendant presents specific, concrete indications that AI was used, the claimant has to respond with a detailed account of their own creative process — or lose the presumption that the work is human-authored. The lyrics in that case survived, because the songwriter could actually account for how she wrote them. The ruling is a clean statement of the exact mechanism this piece is about: courts shift the burden onto whoever needs it once AI enters the picture, and only a demonstrable record satisfies it.
Courts aren’t asking whether AI was used. They’re asking whether you can reconstruct which stage you contributed at — prompt, selection, post-processing, hand-edit — and show your work.
That's not a copyright-law observation. It's a record-keeping problem. Most creative teams can't answer that question about their own output six months later, AI-assisted or not, because nothing in a typical workflow captures which decisions were whose. The file gets exported, the metadata gets stripped, and the chain between “here's what I did” and “here's the output” is gone.
Is Germany an Outlier?
The instinct — including in some of the commentary on this ruling — is to ask whether Germany is isolated or whether other courts are converging on the same approach. The honest answer doesn't sort cleanly into either bucket, and claiming it does is exactly the kind of overreach this piece is trying to avoid.
What does look consistent: the idea-expression dichotomy itself — the analytical tool the German court used — shows up as the starting point in copyright analysis across most jurisdictions with functioning copyright systems. That's not a coincidence; it's a shared doctrinal ancestor, not a coordinated new consensus about AI specifically.
What doesn't look settled: whether AI-generated output itself clears the bar for protection, and how much human involvement that requires. This threshold is still moving — including within single jurisdictions. China is the clearest example: the Beijing Internet Court found, in Li v. Liu (November 2023), that a user's iterative prompting and parameter refinement — starting from an initial prompt, reviewing the output, and repeatedly adjusting it — met the originality threshold and granted the prompter authorship. Roughly sixteen months later, the Zhangjiagang People's Court (March 2025) reached the opposite result on a single, un-iterated prompt: no original records of a creative process, no demonstrated individualized input, protection denied. Same country, same basic legal test, opposite outcomes — because the evidence of process was different. That's not two jurisdictions disagreeing. That's one jurisdiction still working out where its own line sits.
Treat any claim of a clean global consensus — in either direction — with suspicion, including claims made in this space by other commentators. The convergence is in the method courts use to ask the question. The answer is still fragmented, including inside individual legal systems.
What This Means If You Can't Produce a Court Case as Evidence
Nobody wants to find out whether they can meet this bar by being sued. The practical version of the question is: if someone challenged a piece of your creative output tomorrow — AI-assisted or not — could you show which decisions were yours?
For most teams, no. That's not a legal failure; it's an infrastructure gap. Prompts get lost in chat history. Reference images get overwritten. The specific model version, seed, and post-processing steps that turned a rough generation into a finished asset live nowhere durable. The lineage — the sequence of human and machine decisions that produced the final file — is exactly the thing that would let you answer a court's question, and it's exactly the thing most workflows never capture.
Where the Record Should Already Be Sitting
This is the gap Numonic is built to close. For ComfyUI and Midjourney workflows specifically, Numonic retains that lineage inside the platform automatically: the workflow's node chain, including the prompt and seed values that produced it, parent-child relationships between assets, and the evolution chain from one generation to the next. That record lives in Numonic, queryable later, instead of scattered across chat history and overwritten files.
Worth being precise here: that's the internal record. Numonic also lets you choose to strip that same workflow and prompt data from anything you export to share publicly — the record staying internal by default is a deliberate privacy choice, not a gap. The two aren't in tension: what you can prove and what you choose to publish are different questions, and Numonic is built to let you answer the first without being forced into the second.
None of this settles any court's decision. No tool can promise a legal outcome, and this piece isn't legal advice. What it does is put you in the position the German court's winning party was in: able to point to a specific, demonstrable record of which choices were whose, instead of reconstructing a workflow from memory under pressure, potentially years after the fact.
Key Takeaways
- The German ruling didn't create a special AI carve-out — it applied the ordinary idea-expression dichotomy, and found the AI output changed enough expression to avoid infringing, while also declining to grant the AI output its own copyright.
- Courts are testing creative contribution stage-by-stage: prompt, selection, post-processing, hand-edit. The more of that sequence you can evidence, the more a court (or anyone challenging your work) has to reason with.
- The “is this global now” question doesn't have a clean answer. The analytical method (idea-expression dichotomy) travels well across jurisdictions; the protectability threshold for AI output specifically does not — it's still moving even within single jurisdictions like China.
- The underlying problem isn't legal, it's a record-keeping one: can you reconstruct which stage you contributed at, months after the fact?