Someone ran your photo through AI: what a German appeal court said, and which records help

This is general information, not legal advice. Copyright rules differ by country, and the decision discussed here is a court decision in interim-injunction proceedings, not a statute. Check the sources at the end, and ask a copyright lawyer in your jurisdiction about a real dispute.

What happened in the Düsseldorf case

A photographer who sells underwater dog photographs applied for an interim injunction against someone who had previously cooperated with her (the court notes that the nature, extent and reason for ending that cooperation were disputed). The respondent had uploaded one of her photos, a dog diving toward a red toy, to AI software and published the resulting comic-style image on a website. The court of first instance (the Landgericht) rejected her application for an interim injunction (einstweilige Verfügung) on 22 December 2025. She appealed (a sofortige Beschwerde), and on 2 April 2026 the Oberlandesgericht (OLG) Düsseldorf dismissed that appeal. The decision is catalogued as 20 W 2/26 (ECLI:DE:OLGD:2026:0402.20W2.26.00), and the full German text is published on the North Rhine-Westphalia court database. It was decided in interim-injunction proceedings, so it rules on that application rather than after a full trial on the merits, and the decision records that it cannot be appealed further on points of law.

The ruling was reported by specialist outlets in April, including The Decoder and the law firm D Young & Co. It reached a much wider audience in July 2026, when PetaPixel covered it and a one-line summary on X (this post, showing about 8.0 million views at the time of our check) said that transforming someone else’s image with AI “does not constitute as copyright infringement”. That summary is broader than the decision. PetaPixel’s own text says the court ruled that doing so does not automatically infringe.

What the court actually decided

Reading the court’s text alongside the two law-firm and trade summaries, the reasoning has three parts.

1. A photograph’s protection is about expression, not the scene. PetaPixel reports the court as saying the scope of protection for a photograph is generally limited to the choice of image composition, perspective, lighting, and the sharpness or blur achieved through aperture and exposure time, while “the subject matter and motif are generally not eligible for protection” (that wording is PetaPixel’s translation of the German). Here the AI image kept the motif, a dog reaching for a red toy underwater, but used different framing, a different angle and a flat cartoon rendering. D Young notes the court found the AI image lacked the dynamic quality the photographer achieved through exposure and aperture choices.

2. The AI image was not itself a “work”. The court said that for AI output to count as a protected work, the human user must make creative decisions that shape the specific result. Merely selecting one of several AI suggestions is not enough. Because the comic image was not a work, the German “free adaptation” rule in section 23(1) sentence 2 of the Copyright Act did not apply. The court also noted that whoever relies on the image being a “work” has to set out the creative decisions involved, and the user in this case did not do so, even after being told at the hearing on 3 March 2026 that the image looked AI-made. The photographer lost on a different route: no protected element had been taken. The court also found no breach of the separate neighbouring right that protects photographs, which is section 72 of the same Act.

3. European case law shapes the test. The decision cites a December 2025 judgment of the Court of Justice of the EU (Mio and konektra, C-580/23 and C-795/23) on what counts as a work. On that authority the court said an infringement finding needs the creative elements of the protected work to have been recognisably taken, and that comparing the overall impression is not the deciding test. The Decoder summarises the ruling the same way.

The Decoder also says the decision lines up with earlier German rulings and with the position of the US Copyright Office. That office’s report on copyrightability of AI outputs (Part 2, 29 January 2025) is listed on its AI initiative page.

What the ruling does not say

• It does not say AI copying is always allowed. A closer AI copy that reproduced the framing, perspective or lighting might be analysed differently. The court looked at one image on its facts.
• It does not bind courts elsewhere. It is an appellate decision in one German region. US, UK, Irish or French courts apply their own infringement tests, and nothing in this article predicts how they would treat a similar image.
• It does not decide training questions. This was about an image made from one photo. Whether scraping many photographs to train a model is lawful is a separate fight. PetaPixel mentions an earlier German ruling against a stock photographer who sued the LAION dataset organisation, where the court treated the free dataset as scientific research use. Read the underlying decisions before relying on any summary.
• It does not say dates or records were irrelevant. The court’s question was what was taken. Records only matter if you later need to describe your own choices and show which version came first.

If it happened to you: records worth keeping now

The practical lesson is that the more of your concrete choices you can point to, the clearer your description of what was or was not taken. A short routine:

1. Keep the native originals. For a photographer, that is the RAW file plus any sidecar or catalogue edits; for an illustrator, the layered file and the exported versions you published. Do not overwrite them.
2. Date the originals before you need to. Make a TimeBind of the exact files. Start with the free SHA-256 checker if you want to see a file’s fingerprint first, and see how to prove a RAW photo existed before a date.
3. Capture the other image as you found it. Save the page, the image file if you can, the web address and the date you saw it. Fingerprint what you saved so you can show later that your copy has not changed. The file comparison checker shows whether two files are identical.
4. Write down the elements you chose. A plain note of framing, lens, perspective, lighting and post-processing decisions is easy to produce while you remember them. A lawyer may use it; a TimeBind of the note can be made too.
5. Do not rely on metadata alone. Camera dates can be edited. See how to check when a photo was taken and whether a file’s date created can be changed.
6. Get advice before you send a demand. In this decision, what mattered was which protected elements were taken, not simply that AI was used. A lawyer can tell you which, if any, were. For general routes, see how to prove someone copied your work and EU copyright without registration.

If you use AI on a photo or your own art: the other half of the ruling

The same decision cuts the other way for people who generate images. D Young’s takeaway is that a user who wants copyright in AI output has to be able to demonstrate and document the creative choices exercised in prompting: the prompts, the settings and the selection among outputs. The court said simply picking an output from several suggestions is not enough.

That is a record-keeping point, not a promise of protection. If you work with AI tools, keep the prompts, settings, intermediate files and your edits, and date the final files you publish. In the United States, the Copyright Office’s Part 2 report is the place to start on what is registrable, and our guide on disclosing AI on a US copyright application explains why records of your own contribution matter. None of this tells a court that you made something; it only helps you describe what you did.

How a TimeBind fits, and what it does not show

A TimeBind is a dated, checkable record that an exact file existed no later than a recorded time and still matches. The fingerprint is made in your browser, and the file stays on your device. It helps with questions like “did this exact RAW or layered file exist before that AI image appeared?” and “is the copy I saved of the other image the same one I saved last month?”

It does not show that anyone copied you, that you made or own the work, that a use is infringing, or that an image is AI-generated. TimeBinder does not detect fakes. Whether something infringes is a legal question for a court and depends on the country. For the creator workflow, see creator timestamping and timestamping for illustrators.

Sources checked

• OLG Düsseldorf, decision of 2 April 2026, 20 W 2/26 (German text): nrwe.justiz.nrw.de
• D Young & Co, “AI output and copyright infringement: protection regarding an AI-generated image of real photograph (Underwater Dog)”, 23 April 2026: dyoung.com
• The Decoder, “German court rules AI comic adaptation of copyrighted photo doesn’t violate the original”, 19 April 2026: the-decoder.com
• PetaPixel, “Dog Photographer Loses Copyright Case Over AI-Generated Comic Version of Her Image”, 22 July 2026: petapixel.com
• U.S. Copyright Office, Copyright and Artificial Intelligence (Part 2, copyrightability, 29 January 2025): copyright.gov/ai
• X post summarising the ruling, showing about 8.0 million views when checked on 11 November 2026: x.com/TheMG3D/status/2079946004582007104

Photo and video originals are grouped on the photos and video timestamping hub. A free 10-question quiz on recent viral AI fakes walks through real cases, with sources; it is not a detector and does not check a file you upload.

Start a dated record of your originals

Your first Pooled TimeBind is free, with no credit card required. After that, TimeBinds start from US$5, and an ASAP TimeBind is US$15. Click here to start, or see plans and pricing.

Related: A night-sky photo accused of being AI: what photographers should keep.