---
title: "AI and copyright: definition and example – Carl-Uwe Würfel"
description: "AI and copyright explained: when AI training is allowed, who owns AI output, what German courts have ruled, and what companies should check before publishing."
url: https://carl-wuerfel.de/en/glossary/ki-urheberrecht
author: Carl-Uwe Würfel
---

1. [Home](https://carl-wuerfel.de/en)
  2. [Knowledge](https://carl-wuerfel.de/en/knowledge)
  3. [Glossary](https://carl-wuerfel.de/en/glossary)
  4. AI and copyright

# AI and copyright

AI and copyright covers three questions: whether models may be trained on protected works, whether AI output is protected, and when an output infringes someone else's rights.

## On this page

  1. Training counts as text and data mining, with an opt-out
  2. The first rulings separate analysis from storage
  3. The AI Act requires model providers to have a copyright policy
  4. Only the human contribution is protected
  5. The publisher carries the risk of the output
  6. What a clean process looks like
  7. Where the term stops working
  8. Frequently asked questions
  9. Sources

Publishing is a use of its own, so the publisher needs the rights.

_Figure: A checklist using the example of a furniture retailer that drafts catalog copy and image backgrounds with generative AI. Before publishing, every draft passes three checks: whether the vendor 's terms allow commercial use, whether the draft resembles known works, characters, or trademarks, and whether the tool, prompt, selection, and edits are recorded. A purely generated draft can still go out, but copyright does little to stop others from copying it._

Debates about AI and copyright mix up three questions the law answers separately. First, training: German copyright law permits text and data mining, but rights holders can opt out. Second, the result: under Section 2(2) of the German Copyright Act (UrhG), only a person's own intellectual creation is protected, and under Section 7 the author is whoever created the work. Third, infringement: if an AI reproduces a protected work, that use needs the rights holder's permission.

German courts have ruled since 2024, and the Federal Court of Justice plans to issue its first judgment on an AI training dataset on December 17, 2026 (as of November 2026).

For companies, what counts is practice: rights in their AI content, checks before publishing, vendor contracts. This entry explains the basics and isn't legal advice; specific cases call for a copyright lawyer.

Example

A furniture retailer uses generative AI to draft catalog copy and image backgrounds. An editor checks every image against known works and trademarks and documents the edits. The vendor contract states that the retailer's inputs won't be used for training, and its website carries a machine-readable reservation against text and data mining.

Chapter 1 of 7

## Training counts as text and data mining, with an opt-out.

Section 44b UrhG defines text and data mining as automated analysis of digital works to extract information about patterns, trends, and correlations. Copying lawfully accessible works for it is allowed. The provision implements Article 4 of Directive (EU) 2019/790.

The permission has a limit. Under Section 44b(3), it doesn't apply if the rights holder has reserved the use. For works available online, that reservation only counts in machine-readable form. What exactly qualifies as machine-readable hasn't been settled. Signatories of the EU code of practice for AI models have committed to following robots.txt, the file a website uses to tell crawlers what not to fetch.

Section 60d adds a separate permission, with no opt-out, for research organizations that are non-commercial or reinvest their profits in research.

Chapter 2 of 7

## The first rulings separate analysis from storage.

In a photographer's case against the nonprofit LAION, the Hamburg Regional Court dismissed the claim (September 27, 2024, 310 O 227/23) and the Higher Regional Court the appeal (December 10, 2025, 5 U 104/24). LAION had downloaded a preview of the photo to check it against its caption for an AI training dataset. The appeals court found a reproduction covered by Sections 44b and 60d. It rejected the photo agency's plain-language reservation because nobody had shown it was machine-readable in 2021. The Federal Court of Justice now has the case (I ZR 281/25).

The Munich Regional Court went the other way on November 11, 2025, in GEMA v. OpenAI (42 O 14139/24). The models had memorized song lyrics and could reproduce them. The court said the text and data mining exception allows copies for analysis, not a copy of the work inside the model. The judgment is not yet final.

Chapter 3 of 7

## The AI Act requires model providers to have a copyright policy.

Since August 2, 2025, Article 53(1) of the [AI Act](https://carl-wuerfel.de/en/ai-act) (Regulation (EU) 2024/1689) has put two duties on providers of general-purpose AI models. Under point (c), they need a policy to comply with EU copyright law, in particular to identify and respect reservations under Article 4(3) of Directive (EU) 2019/790. Under point (d), they publish a sufficiently detailed summary of their training content using a template from the AI Office.

The copyright chapter of the voluntary code of practice (July 10, 2025) is more specific: signatories commit not to get around paywalls and to skip known piracy sites. They add safeguards against infringing output and a contact point for rights holders' complaints.

These duties fall on model providers. Companies that only use a model can read the summary when choosing one.

Chapter 4 of 7

## Only the human contribution is protected.

A prompt alone usually doesn't turn an AI image into a protected work. On February 13, 2026, the Munich Local Court (142 C 9786/25) denied protection for three logos a plaintiff had generated by prompting, partly with detailed, step-by-step instructions. Protection is possible, the court said, if human influence shapes the result in an objectively recognizable way. His claim against an acquaintance who copied the logos failed.

The U.S. Copyright Office's January 2025 report agrees: with current technology, prompts alone don't give enough control. Protection does extend to human works that remain perceptible in AI output, to the creative selection and arrangement of material, and to creative modifications of the output.

Copyright therefore does little to stop someone from copying a purely generated image. If you want protection, shape the work and document your edits.

Chapter 5 of 7

## The publisher carries the risk of the output.

AI output can reproduce someone else's work, such as song lyrics, a character, or a photo. In the Munich GEMA case, the original elements of the lyrics were always recognizable in the output. The court held the model operators responsible, since simple prompts were enough.

That doesn't shield a company that publishes such output. Under Section 19a UrhG, the right to make a work available to the public belongs to the author. Publishing is a use of its own, so the publisher needs the rights unless a statutory exception such as quotation applies.

Whether a vendor covers third-party claims is a matter of contract. Labeling content as AI-generated under the AI Act doesn't replace a rights check, because it answers a different question.

Chapter 6 of 7

## What a clean process looks like.

Contracts with AI vendors spell out what rights the company gets in the output, whether its inputs are used for training, and whether the vendor covers third-party claims. Terms change, so they need a fresh look after every update.

Before publishing, a person checks images and text for similarity to known works, characters, and trademarks, then revises them. The tool, prompt, selection, and edits are recorded. That record helps prove the human contribution later.

A company that doesn't want its own content used for AI training states the reservation on its website in machine-readable form, for example in robots.txt, and also in its terms of use. A short policy lists approved tools and which outside content must never be uploaded.

Chapter 7 of 7

## Where the term stops working

This entry describes the law in Germany and the EU as of November 2026 and isn't legal advice. AI labeling under Article 50 of the AI Act, data protection, trademark law, and personality rights in deepfakes follow their own rules. Search performance of AI-written text depends on quality and provenance ([E-E-A-T](https://carl-wuerfel.de/en/glossary/eeat)).

## Frequently asked questions

Are AI-generated images and texts protected by copyright?

Usually not if they were created by prompting alone. German law requires a human to shape the result in an objectively recognizable way.

Can an AI be an author?

No. Under Section 7 of the German Copyright Act, the author is whoever created the work, and that can only be a person.

Is AI training on copyrighted works allowed?

Under conditions, yes: as text and data mining under Section 44b UrhG, as long as the rights holder hasn't declared a machine-readable reservation. The Munich Regional Court held that storing entire works in the model isn't covered.

How do you opt out of AI training for your own content?

With a machine-readable reservation, such as an entry in robots.txt, because only that counts for online content under Section 44b(3) UrhG. A plain-language statement wasn't enough for the Hamburg appeals court in the LAION case.

Who is liable when an AI reproduces protected content?

Model operators can be liable for the output itself, as the Munich Regional Court ruled in GEMA v. OpenAI. A company that publishes the output also uses the work and needs the rights.

What did the Munich court decide in GEMA v. OpenAI?

On November 11, 2025, it largely ruled for GEMA: song lyrics had been memorized in the models and reproduced in output, and no exception applied. The judgment is not yet final.

What does the AI Act say about copyright?

It requires providers of general-purpose AI models to have a copyright policy and to publish a summary of their training content (Article 53). Copyright law still governs protection and infringement.

Can you use AI-generated content commercially?

That depends on two checks. The vendor's terms have to allow commercial use, and the output must not reproduce someone else's works or trademarks.

## Further reading

  * [**AI glossary**](https://carl-wuerfel.de/en/glossary)
  * [**AI Act: classifying how you use AI**](https://carl-wuerfel.de/en/ai-act)
  * [**General-purpose AI (GPAI)**](https://carl-wuerfel.de/en/glossary/gpai)
  * [**AI labeling and disclosure**](https://carl-wuerfel.de/en/glossary/ki-kennzeichnung)
  * [**Generative AI**](https://carl-wuerfel.de/en/glossary/generative-ki)
  * [ Term**E-E-A-T**](https://carl-wuerfel.de/en/glossary/eeat)
  * [ Term**Deepfake**](https://carl-wuerfel.de/en/glossary/deepfake)

## Sources

  1. [German Federal Ministry of Justice: Copyright Act (Urheberrechtsgesetz, UrhG), Sections 2, 7, 16, 19a, 44b, and 60d, as of November 2026 (German)](https://www.gesetze-im-internet.de/urhg/)
  2. [European Union: Directive (EU) 2019/790 on copyright in the Digital Single Market, Articles 3 and 4, 2019](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32019L0790)
  3. [European Union: Regulation (EU) 2024/1689 (Artificial Intelligence Act), Articles 50, 53, and 113, 2024](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689)
  4. [European Commission: Code of Practice for General-Purpose AI Models, Copyright Chapter, July 10, 2025](https://ec.europa.eu/newsroom/dae/redirection/document/118115)
  5. [Federal Court of Justice (Bundesgerichtshof): hearing notice I ZR 281/25 (creating a dataset for AI training), with the lower-court rulings LG Hamburg, 310 O 227/23, and OLG Hamburg, 5 U 104/24, as of November 2026 (German)](https://www.bundesgerichtshof.de/SharedDocs/Termine/DE/Termine/IZR281-25.html)
  6. [Munich Regional Court I: press release 11/2025, GEMA v. OpenAI, 42 O 14139/24, November 11, 2025 (German)](https://www.justiz.bayern.de/gerichte-und-behoerden/landgericht/muenchen-1/presse/2025/11.php)
  7. [Munich Local Court: judgment of February 13, 2026, 142 C 9786/25 (AI-generated logos) (German)](https://www.gesetze-bayern.de/Content/Document/Y-300-Z-BECKRS-B-2026-N-1513)
  8. [U.S. Copyright Office: Copyright and Artificial Intelligence, Part 2: Copyrightability, January 29, 2025](https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-2-Copyrightability-Report.pdf)
