# Copyright and AI output: who owns what the model writes
> A generated text, image or function: who owns it, whether it can infringe someone else's work, and the status of the data the model learned from. The court cases that matter as of November 2026, and practical rules for a team.
Source: https://rfriedmann.de/blog/who-owns-ai-output/
Published: 2026-11-08 · Track: learn · Level: Ops
The [lesson on GDPR and the AI Act](/blog/ai-and-the-law-gdpr-ai-act/) covered what goes
*into* a model. This one is about what comes *out*. Marketing wants to know whether the
campaign image from the generator is "ours". A developer asks whether the function the
assistant wrote can go into the product. Legal asks whether we can be sued over any of
it. Three different questions, and copyright answers each one differently.
Same caveat as before: this is not legal advice. It is the map an IT lead needs to ask
the right questions, with the court cases as they stand in November 2026.
## Question one: is the output anyone's?
In the EU a work is protected when it is the author's *own intellectual creation*. The
Court of Justice said so in *Infopaq* (C-5/08, 2009) and spelled it out in *Painer*
(C-145/10, 2011): the author has to make free and creative choices that stamp the work
with a personal touch. German law says the same in its own words: § 2 (2) UrhG protects
only a *persönliche geistige Schöpfung*, a personal intellectual creation. A machine is
not a person, so pure machine output has no author.
The US got there by a different road and reached the same place. The Copyright Office's
report *Copyright and Artificial Intelligence, Part 2: Copyrightability* (29 January
2025) concluded that prompts alone do not give the user enough control to be the author:
a prompt is closer to an instruction or an idea, and the same prompt gives different
results. On 18 March 2025 the D.C. Circuit held in *Thaler v. Perlmutter* that the
Copyright Act requires a human author, and on 2 March 2026 the Supreme Court declined to
hear the case. That ruling stands.
What courts and the Copyright Office leave open is the middle: how much human work makes
an AI-assisted result protectable. The direction is consistent, though.
The more of you is in the result, the more of it is protectable
[diagram omitted — see the page for the chart]
A rough scale, not a legal test. Choosing among outputs is a choice, but the US Copyright Office has so far treated it as too thin. Where you rewrite, combine and arrange, your additions can be protected; the parts the model produced on its own stay unowned.
The practical consequence is often unexpected: a raw generated logo,
product photo or text block may be free for anyone to copy, including your competitor.
If exclusivity matters, a human has to do real creative work on it, and you want a
record that they did.
## Question two: can the output copy someone else's work?
Yes, and that risk is yours. Models can [memorise](/blog/why-models-make-things-up/)
pieces of their training data and repeat them almost word for word: song lyrics, well
known passages, code snippets complete with their licence.
The leading European case is *GEMA v OpenAI*. On 11 November 2025 the Regional Court
Munich I (42 O 14139/24) held that lyrics memorised in the model are a reproduction, and
that ChatGPT reproducing them on request infringed the rights GEMA manages. OpenAI has
appealed. On 3 November 2026 Advocate General Szpunar gave his opinion in the first
generative AI case at the Court of Justice, *Like Company v Google* (C-250/25): a
chatbot may freely report the facts from a news article, but an answer that reuses its
expression so closely that the reader no longer needs the original can infringe the
press publisher's right (Art. 15 of the Copyright in the Digital Single Market
Directive). That
opinion does not bind the Court; the judgment comes later.
For a team that means: whoever publishes the output is the one publishing a copy. The
vendor may be liable too, as *GEMA v OpenAI* shows, but that does not take the
responsibility off you. Code is the everyday case. A suggestion that
reproduces a GPL-licensed function brings the GPL's conditions along, whether or not
the licence header survived. Some code assistants can filter suggestions that match
public code; switch that on.
## Question three: was the training allowed?
This is mostly the model maker's problem, and it is still being decided. EU law has two
text and data mining exceptions in the 2019 DSM Directive: Art. 3 for research, Art. 4
for everyone else, unless the rightsholder has reserved their rights. Germany implements
them as § 60d and § 44b UrhG. For works online, § 44b only accepts a reservation in
*machine-readable* form.
That detail decided the first German case. Photographer Robert Kneschke sued LAION, the
non-profit behind a large image-text dataset. The Regional Court Hamburg dismissed the
claim on 27 September 2024 (310 O 227/23). The Higher Regional Court Hamburg dismissed
his appeal on 10 December 2025 (5 U 104/24), held that § 44b covered building the
dataset, and found that a plain-language reservation in a website's terms was not
machine-readable. The Federal Court of Justice heard the case on 3 September 2026
(I ZR 281/25) and has set 17 December 2026 to announce its decision; a referral to the
Court of Justice is possible.
The AI Act adds a duty for the makers of [general-purpose models](/blog/ai-and-the-law-gdpr-ai-act/):
a copyright policy that respects these reservations (Art. 53). It becomes your problem
when you [fine-tune](/blog/fine-tuning-with-lora/) or build a dataset yourself. And if
you run the company website and want it left out of training: say so in a
machine-readable way, robots.txt rules for the AI crawlers at least, not only a
sentence in the imprint.
## What the vendors promise
The big providers offer to defend business customers against copyright claims over
output: Microsoft's Customer Copyright Commitment (it started as the Copilot Copyright
Commitment in 2023 and now covers Azure OpenAI too), OpenAI's Copyright Shield for
ChatGPT Enterprise and the API, and Google's generative AI indemnities for Workspace and
Cloud. Read the conditions before you count on them: paid business plans only, the
vendor's content filters and required mitigations switched on, and no deliberate
attempt to produce infringing material. A free account an employee signed up for
carries none of this.
## The rules I'd give a team
Do
Have a human edit anything you want to own exclusively
Keep prompts, drafts and edits as a record of the human part
Use business plans with the indemnity, filters on
Turn on the public-code filter in code assistants
Label AI content where you publish it
Don't
Ask for lyrics, book passages or "exactly like" a known work
Paste third-party protected text in just to "rewrite" it
Treat a raw generated logo as a protectable brand asset
Merge generated code that carries someone's licence text
Rely on a free personal account for anything published
None of this needs a lawyer for everyday use. It needs the same habit as with
[model licences](/blog/open-weights-and-licenses/): know where a thing came from before
you ship it. For anything with significant financial value, a brand, a product, a book, the
lawyer gets the last word.
## Primary sources
- **EU**: Directive (EU) 2019/790 (DSM Directive), Art. 3 and 4, on EUR-Lex; CJEU
judgments C-5/08 and C-145/10 on curia.europa.eu, where C-250/25 will also appear.
- **Germany**: §§ 2, 44b and 60d UrhG on gesetze-im-internet.de; the BGH press releases
for I ZR 281/25.
- **US**: the Copyright Office's AI report, Part 2, on copyright.gov.
- **AI Act**: Art. 53 of Regulation (EU) 2024/1689 and the General-Purpose AI Code of
Practice, copyright chapter.
Two decisions are due soon after this lesson: the BGH in *Kneschke v LAION* and, later,
the Court of Justice in *Like Company*. Either can move the lines drawn here.