Civil Law and Artificial Intelligence Copyright Infringement Claims in Europe

1. Introduction

Artificial intelligence has created a new category of copyright disputes in Europe. The main question is when the development, training, operation, or output of an AI system unlawfully uses copyright-protected material.

The disputes generally arise at two different stages:

Training stage – copyrighted books, photographs, music, articles, software, or other works are collected and copied into datasets or processed during AI training.

Output stage – an AI system produces text, images, music, code, or other material that may reproduce protected expression from an existing work.

EU law does not currently contain one single “AI copyright infringement test.” Ordinary copyright principles continue to apply, particularly the reproduction right, communication/making-available rights, originality requirements, and the text-and-data-mining (TDM) exceptions. The AI Act adds specific obligations for providers of general-purpose AI models, including copyright-compliance policies and summaries of training content. (AI Act Service Desk)

A particularly important recent development is GEMA v OpenAI, decided by the Regional Court of Munich I in November 2025, where the court granted substantial copyright claims concerning memorised and reproduced song lyrics. (justiz.bayern.de)

2. Meaning of AI Copyright Infringement

AI copyright infringement may occur where an AI-related activity involves an unauthorised act protected by copyright.

Typical examples include:

copying copyrighted works into an AI training dataset;

scraping copyrighted websites;

reproducing protected works during model training;

storing protected expression in model parameters in a form capable of reconstruction;

generating substantially identical passages from protected works;

reproducing photographs or illustrations;

generating protected song lyrics;

generating substantial portions of books or articles;

distributing AI-generated material that reproduces protected expression;

making infringing AI outputs available to the public.

However, mere similarity is not automatically copyright infringement. Copyright generally protects original expression rather than ideas, facts, concepts, themes, styles or general techniques.

3. Main European Legal Framework

A. EU Copyright Directive 2001/29/EC

The Information Society Directive provides the basic framework for:

reproduction;

communication to the public;

making available;

copyright exceptions;

temporary reproduction.

For AI disputes, Article 2's reproduction right is particularly important.

The fundamental question is:

Has the AI system reproduced protected expression belonging to the copyright owner?

The CJEU's jurisprudence makes clear that even a relatively small portion can constitute reproduction where it contains the author's own intellectual creation.

4. Digital Single Market Copyright Directive

Directive (EU) 2019/790, commonly called the DSM Directive, is particularly important for AI training.

Articles 3 and 4 establish text-and-data-mining exceptions.

Article 3

This concerns TDM undertaken by:

research organisations; and

cultural heritage institutions,

for scientific research purposes, subject to the statutory conditions.

Article 4

This provides a broader TDM exception, but rightsholders can reserve their rights against mining of publicly accessible works, subject to the Directive's requirements.

This is extremely important for commercial generative AI.

The EU AI Act expressly connects general-purpose AI development with these copyright rules. Its Recital 105 recognises that training large generative models requires large quantities of potentially copyrighted text, images, video and other material and explains the relevance of the TDM exceptions and rightsholder opt-outs. (AI Act Service Desk)

5. EU AI Act

The EU AI Act (Regulation (EU) 2024/1689) does not replace copyright law.

Instead, it introduces obligations relevant to copyright compliance for providers of general-purpose AI models.

Important obligations include:

adopting a policy to comply with EU copyright law;

respecting applicable TDM opt-outs;

preparing summaries of training content;

maintaining appropriate documentation;

complying with obligations concerning general-purpose AI models.

Therefore:

AI Act compliance ≠ automatic copyright clearance.

A provider may comply with AI Act obligations while still facing a copyright claim if a particular use infringes copyright.

6. Product Liability and Civil Liability

AI copyright disputes can also have a civil-law dimension.

Potential claims may involve:

injunctions;

damages;

restitution;

information/disclosure;

destruction or removal of infringing material;

licensing disputes;

contractual claims;

unjust enrichment, depending on national law;

intermediary/provider liability.

The precise remedy depends upon the applicable national copyright law and procedural rules.

7. Training-Stage Copyright Infringement

Training generally involves several technical steps:

Web scraping → copying → dataset creation → preprocessing → training → parameter adjustment → model deployment

Copyright questions can potentially arise at several points.

For example:

A company downloads 10 million copyrighted photographs and uses them to train an image-generation model.

The legal questions include:

Were the photographs reproduced?

Was the reproduction authorised?

Does Article 3 or Article 4 DSM Directive apply?

Was there a valid opt-out?

Was the activity scientific research or commercial?

Was the material merely analysed or retained in a reconstructable form?

Can protected expression subsequently be reproduced through the model?

8. Output-Stage Copyright Infringement

A second category occurs when an AI produces material resembling or reproducing an existing work.

For example:

A user asks an AI system to reproduce a particular newspaper article, song lyrics or photograph, and the system generates substantial portions of the protected work.

Possible legal issues include:

reproduction;

communication to the public;

making available;

adaptation;

authorisation;

secondary liability;

user liability;

provider liability.

The distinction between inspiration and reproduction of protected expression is therefore fundamental.

9. AI Memorisation

One of the most important emerging concepts is memorisation.

A model may sometimes reproduce training material substantially verbatim when prompted appropriately.

This creates a stronger copyright problem than an output that merely resembles the general subject matter of training data.

The recent German GEMA litigation is particularly significant because the Munich court dealt with alleged memorisation and reproduction of protected song lyrics by an AI system. (justiz.bayern.de)

10. Case Law

Case 1 — GEMA v OpenAI

Regional Court of Munich I, 11 November 2025, Case No. 42 O 14139/24

This is one of the most directly relevant European AI-copyright decisions.

GEMA, representing authors and music rightsholders, brought claims concerning protected German song lyrics that could allegedly be reproduced by ChatGPT.

The court substantially upheld GEMA's claims for:

injunctive relief;

information;

damages.

The Munich court considered the reproduction and memorisation of protected lyrics by the AI system. It rejected additional claims concerning general personality rights arising from incorrect attribution of modified lyrics. (justiz.bayern.de)

Importance

The case demonstrates that:

AI-generated output can create a conventional copyright infringement claim where protected expression is reproduced.

It also makes model memorisation an important evidentiary issue.

The decision is subject to further proceedings/appeal, so it should not be treated as a final European-wide rule. (digital-client-solutions.hlc.com)

11. Case 2 — Kneschke v LAION

Hamburg Regional Court, 27 September 2024, Case No. 310 O 227/23

This is one of Europe's earliest important cases specifically concerning AI training data.

Photographer Robert Kneschke alleged that LAION had reproduced his photograph while creating a dataset used for AI training.

The Hamburg court accepted that reproduction of the photograph had occurred in the dataset-creation process but concluded that the activity was covered by the German TDM exception applicable to scientific research. (WIPO)

Importance

The case illustrates the crucial distinction between:

copyrighted work was copied

and

copying was legally permitted under a copyright exception.

Therefore, proving reproduction alone does not necessarily establish infringement.

The case is particularly important for:

AI datasets;

scientific research;

TDM;

non-commercial AI research;

copyright exceptions.

12. Case 3 — DPG Media / Mediahuis v Knowledge Exchange (HowardsHome)

District Court of Amsterdam, 3 October 2024

This dispute concerned AI-generated summaries and the use of protected news content.

The Amsterdam court considered the application of copyright and TDM principles to an AI-related news aggregation/summarisation service.

The decision illustrates that AI-generated summaries cannot simply be analysed by asking whether the final text is identical to the original. The underlying acts of obtaining, processing and reproducing protected material must also be examined.

The decision is important for:

AI-generated summaries;

news articles;

TDM;

opt-out mechanisms;

reproduction;

the three-step test.

The case has also become part of the emerging European discussion about the limits of TDM in AI-related activities. (DOI)

13. Case 4 — Infopaq International A/S v Danske Dagblades Forening

CJEU, Case C-5/08, 16 July 2009

Although this was not an AI case, it is one of the most important authorities for AI copyright disputes.

The case concerned the reproduction of newspaper extracts.

The CJEU held that even an extract of 11 words could constitute reproduction in part if it contained elements expressing the author's own intellectual creation. (EUR-Lex)

Importance for AI

Suppose an AI produces:

a short paragraph;

a sentence;

a small portion of an article;

a short portion of lyrics.

The relevant question is not simply:

“How many words were copied?”

Instead, the court must examine whether the reproduced portion contains protected expression reflecting the author's intellectual creation.

Therefore:

Short output ≠ automatically non-infringing.

14. Case 5 — Pelham / Kraftwerk

CJEU, Case C-590/23, CG and YN v Pelham GmbH and Others, judgment of 14 April 2026

This case concerned music sampling rather than AI, but it is highly relevant to AI-generated music.

The CJEU addressed:

reproduction;

sampling;

copyright exceptions;

pastiche;

freedom of expression;

artistic freedom.

The Grand Chamber interpreted the EU copyright framework concerning the use of portions of phonograms for pastiche. (EUR-Lex)

Importance for AI

AI-generated music may incorporate characteristics or portions of existing recordings.

The case demonstrates that courts must distinguish:

copying protected expression;

artistic transformation;

permitted exceptions;

freedom of artistic expression.

However, the Pelham decision does not create a general AI-training exception.

15. Case 6 — Painer

CJEU, Case C-145/10, Eva-Maria Painer v Standard VerlagsGmbH and Others, 1 December 2011

This case concerned portrait photography.

The CJEU examined the concept of originality and the author's creative choices.

Importance for AI

AI image-generation disputes often involve questions such as:

Is the original photograph protected?

Has the AI output reproduced protected expressive elements?

Is the output merely based on general ideas?

Has the user's prompt caused the system to reproduce a protected composition?

Painer reinforces the importance of creative choices in determining copyright protection.

It is therefore relevant to AI-generated images even though the technology involved in the original case was not generative AI.

16. Case 7 — Football Dataco v Yahoo! UK

CJEU, Joined Cases C-604/10, judgment of 1 March 2012

The case concerned databases and the originality threshold.

The CJEU distinguished copyright protection based upon intellectual creation from protection based upon investment in database creation.

AI relevance

AI developers frequently use:

databases;

datasets;

catalogues;

structured information;

metadata.

The case helps explain that not every collection of data receives copyright protection simply because substantial resources were invested in creating it.

Therefore, an AI-training dataset must be analysed according to the actual intellectual-property right claimed.

17. Case 8 — Spiegel Online v Volker Beck

CJEU, Case C-516/17, 29 July 2019

This case concerned the use of copyrighted material online and the relationship between copyright and freedom of expression/information.

The CJEU examined exceptions and fundamental rights in the digital environment.

AI relevance

It is relevant where AI-generated or AI-assisted publication involves:

quotation;

news reporting;

public-interest material;

freedom of expression;

online dissemination.

The case illustrates that copyright protection must be interpreted together with applicable exceptions and fundamental rights.

18. Like Company v Google — Important Pending CJEU Reference

CJEU Case C-250/25

This is particularly important because it directly concerns generative AI.

The Hungarian proceedings concern allegations that Google's Gemini chatbot reproduced portions of protected press publications.

Questions referred to the CJEU include whether:

displaying protected content in an AI chatbot response constitutes communication to the public;

training an LLM on protected material constitutes reproduction;

Article 4 DSM Directive's TDM exception applies;

AI-generated responses reproducing protected content can be attributed to the chatbot provider.

The CJEU held an oral hearing in March 2026, but no judgment had been issued at the time of the latest available information. (IP Helpdesk)

This case could become particularly significant for European AI copyright law because it directly addresses the relationship between LLM training, reproduction and AI outputs.

19. Difference Between Training and Output Infringement

20. TDM Exception and AI Training

A central legal issue is whether AI training qualifies as text and data mining.

The answer is not simply:

“AI training is TDM, therefore everything is lawful.”

Instead, courts may need to examine:

purpose of mining;

nature of the organisation;

commercial/non-commercial character;

applicable national implementation;

rightsholder reservation;

technical method of reservation;

temporary or permanent copying;

whether the use remains within the exception;

subsequent use of the trained model.

The Kneschke decision demonstrates the importance of these distinctions. (WIPO)

21. Copyright Opt-Out

Under the DSM framework, rightsholders can, in relevant circumstances, reserve rights against certain forms of TDM.

For AI companies, this creates practical compliance questions:

Was the work publicly available?

Was a reservation expressed?

Was the reservation made in an appropriate manner?

Could the AI crawler detect it?

Was the content collected despite the reservation?

Was the model trained using that content?

The AI Act reinforces the importance of respecting applicable copyright reservations for GPAI providers. (AI Act Service Desk)

22. Machine-Readable Reservations

Modern AI copyright disputes may involve technical mechanisms such as:

robots.txt;

metadata;

HTML instructions;

API restrictions;

platform terms;

machine-readable copyright reservations.

However, the legal validity of a particular mechanism depends upon the applicable EU and national law.

Therefore, an AI company should not assume that:

“The website was publicly accessible, so we could freely train on everything.”

Public accessibility and copyright permission are different legal concepts.

23. AI Output and Substantial Similarity

An AI output can create risk where it reproduces protected expression.

Example 1 — Low risk

A user asks:

“Write a poem about rain.”

The AI creates an independently generated poem.

There is no obvious reproduction of a particular protected work.

Example 2 — Higher risk

The user asks:

“Give me the exact lyrics of a copyrighted song.”

The AI reproduces the lyrics.

This creates a substantially different copyright question because protected expression is being reproduced.

Example 3 — Intermediate situation

A user asks:

“Write a novel in the same general genre as a famous author.”

Similarity in:

genre;

themes;

atmosphere;

general literary technique

does not automatically amount to copyright infringement.

Copyright generally focuses on protected expression, not abstract style.

24. AI-Generated Images

Image-generating AI creates several possible claims.

Training claim

A photographer alleges:

“My photographs were copied into the training dataset.”

The court must consider reproduction and applicable TDM exceptions.

Output claim

A photographer alleges:

“The AI generated an image that reproduces protected elements of my photograph.”

The court must examine:

originality;

similarity;

protected expression;

reproduction;

transformation;

exceptions.

Style imitation

A third claim might be:

“The AI generated an image in my artistic style.”

Style imitation alone does not necessarily establish copyright infringement because copyright protection generally concerns protected expression rather than an abstract artistic style.

25. AI-Generated Music

AI music creates additional questions involving:

musical compositions;

lyrics;

sound recordings;

neighbouring rights;

sampling;

performance rights;

phonograms;

voice imitation.

The Pelham jurisprudence is relevant to the treatment of sampled sound recordings and copyright exceptions. (IP Helpdesk)

The GEMA v OpenAI litigation is particularly important because it concerns reproduction of protected song lyrics through an LLM. (justiz.bayern.de)

26. AI-Generated Software Code

Copyright claims can also concern AI-generated source code.

Potential issues include:

training on copyrighted source code;

reproduction of open-source code;

licence conditions;

attribution requirements;

copyleft obligations;

reproduction of substantial portions;

confidential source code;

database rights.

A developer cannot assume that an AI-generated answer is copyright-free merely because it was produced automatically.

The analysis depends upon what protected material, if any, has actually been reproduced.

27. Who Can Be Liable?

Potential defendants can include:

1. AI model developer

The company that developed and trained the model.

2. AI provider

The company offering the AI system to users.

3. AI deployer

A business integrating the AI system into its own product.

4. End user

The individual or company generating and distributing infringing material.

5. Dataset provider

An organisation that collected or distributed copyrighted training material.

Liability depends upon the particular act, applicable national law and the person's involvement.

28. Provider Liability vs User Liability

A useful distinction is:

User asks → AI generates → user publishes

The user may potentially be responsible for the subsequent use.

But if:

AI provider's system itself reproduces memorised protected material

the provider may also face direct copyright claims.

GEMA v OpenAI illustrates the importance of this distinction because the Munich proceedings addressed both the model's handling of protected lyrics and their subsequent reproduction through the chatbot. (justiz.bayern.de)

29. Causation and Evidence

Copyright litigation involving AI creates unusual evidentiary problems.

A claimant may need evidence concerning:

training datasets;

source URLs;

copies of protected works;

model versions;

training procedures;

model parameters;

prompts;

output logs;

system architecture;

memorisation tests;

similarity analysis;

copyright ownership;

licensing agreements.

Important evidence

Training data + model documentation + prompts + output records + technical expert evidence

can become central to the dispute.

30. Discovery and Transparency

AI copyright litigation may require information that is normally held by the defendant.

For example:

“Was this copyrighted book included in your training dataset?”

or:

“Why does the model reproduce this exact paragraph?”

or:

“What version of the model generated this output?”

This creates tension between:

copyright enforcement;

trade secrets;

confidential algorithms;

privacy;

cybersecurity;

intellectual-property protection.

European courts may therefore need to balance evidence rights against legitimate confidentiality interests.

31. Damages

Where infringement is established, national copyright law may provide remedies including:

damages;

compensation;

injunctions;

information orders;

removal of infringing material;

destruction of infringing copies;

publication of judgments;

accounting of profits or other monetary remedies, depending on national law.

The precise calculation differs among Member States.

Possible factors include:

commercial scale;

duration;

number of works;

economic benefit;

licensing value;

seriousness of infringement;

repeated infringement.

32. Defences Available to AI Developers

An AI developer may potentially rely upon:

A. TDM exception

Where statutory requirements are satisfied.

B. Authorisation/licence

Where appropriate rights were obtained.

C. Lack of protected expression

The material may not qualify for copyright protection.

D. Lack of reproduction

Similarity alone may not establish reproduction.

E. Copyright exception

For example, quotation, parody, pastiche or other applicable limitations.

F. Lack of causation

The claimant may be unable to establish that the allegedly protected material was actually used or reproduced.

G. Independent creation

Where technically and factually supportable.

33. Three-Step Test

European copyright exceptions are also influenced by the three-step test.

Generally, exceptions must be confined to:

certain special cases;

uses that do not conflict with normal exploitation;

situations that do not unreasonably prejudice the legitimate interests of the rightsholder.

This can become particularly controversial where AI systems process enormous quantities of commercially valuable works.

The exact application depends upon the specific exception and statutory framework.

34. Copyright and AI Style Imitation

A particularly difficult issue is:

Can an AI imitate an artist's style?

The answer requires distinguishing style from protected expression.

For example:

“in the style of impressionism” → generally an artistic concept/style;

“paint a picture containing the same original composition, characters and expressive elements as X's protected painting” → potentially much stronger copyright concerns.

Thus, a claimant generally needs more than proof that an AI output has a similar artistic “feel.”

35. Personality Rights and AI

AI copyright disputes can overlap with:

moral rights;

attribution;

integrity rights;

false attribution;

personality rights;

performers' rights.

GEMA v OpenAI is notable because the Munich court separately considered personality-right claims arising from incorrect attribution of altered lyrics and rejected those additional claims. (justiz.bayern.de)

This demonstrates that:

copyright infringement and personality-right infringement are separate legal questions.

36. Collective Management Organisations

Collecting societies may play an important role in AI copyright disputes.

Examples include organisations representing:

musicians;

composers;

authors;

publishers;

photographers.

GEMA's litigation illustrates how collective-management organisations may attempt to establish licensing or enforcement mechanisms against AI providers. (justiz.bayern.de)

37. Contractual AI Copyright Claims

Copyright is not the only source of liability.

A website may contain contractual terms stating:

“Content may not be used for automated training.”

A company scraping that content may face:

contractual claims;

copyright claims;

database-right claims;

unfair-competition claims;

potentially other national-law claims.

Therefore, an AI company's compliance programme should consider both statutory copyright rules and contractual restrictions.

38. Database Rights

AI training frequently involves databases.

European law also protects certain databases through the EU Database Directive.

Potential claims may arise where an AI developer:

extracts substantial portions of a protected database;

repeatedly extracts insubstantial portions;

systematically reproduces database contents;

commercially exploits the database.

Thus:

AI copyright compliance ≠ only copyright compliance.

It may require examination of:

copyright;

database rights;

trade secrets;

contracts;

personal data;

confidentiality.

39. Territorial Problems

AI systems are international.

For example:

Copyrighted work in France → scraped by US company → model trained in the US → model deployed in Europe → output delivered to Germany.

Courts may have to consider:

applicable national law;

territorial jurisdiction;

location of reproduction;

location of training;

location of output;

place of harm;

cross-border enforcement.

The geographical location of AI infrastructure therefore does not automatically determine whether European copyright law is relevant.

40. Civil Liability Framework

A simplified European AI copyright claim can be represented as:

Protected work

↓

Copyright ownership

↓

AI-related reproduction/use

↓

Restricted copyright act

↓

No valid licence

↓

No applicable exception

↓

Infringement

↓

Injunction / information / damages / other remedy

This framework is useful for examination and legal analysis.

41. Important Case-Law Principles at a Glance

The first three are especially useful as direct AI-related European cases, while the others provide foundational CJEU copyright principles that courts can apply to AI disputes. (DOI)

42. Key Legal Issues for Future European AI Copyright Litigation

Future litigation is likely to focus on:

What exactly is copied during AI training?

Does model training constitute reproduction?

When does TDM protect AI training?

How must copyright opt-outs be expressed?

Does machine-readable reservation become mandatory in practice?

Does memorisation constitute reproduction?

Who is liable for an infringing AI output?

Is the user or provider the relevant infringer?

Can AI-generated style imitation infringe copyright?

How should damages be calculated?

How can claimants prove training-data use?

How should trade secrets be balanced against disclosure?

How do database rights interact with AI datasets?

How do copyright exceptions apply to AI outputs?

What happens when training occurs outside Europe but outputs are supplied in Europe?

43. Special Importance of GEMA and Like Company

Two developments are particularly important for understanding the direction of European AI copyright litigation.

GEMA v OpenAI

The German case has already produced a first-instance judgment directly addressing memorisation and reproduction of copyrighted lyrics by an LLM. (justiz.bayern.de)

Like Company v Google

The pending CJEU reference directly asks questions about:

LLM training;

reproduction;

AI-generated responses;

communication to the public;

TDM exceptions.

Consequently, the eventual CJEU judgment could have implications far beyond the parties involved. (IP Helpdesk)

44. Conclusion

European civil-law copyright disputes involving AI are developing through a combination of existing copyright doctrine, TDM exceptions, the AI Act, national copyright legislation and emerging national judgments.

The central distinction is between:

AI training

Whether copyrighted works can lawfully be copied and processed to train a model.

AI memorisation

Whether protected expression remains reproducible from the trained model.

AI output

Whether the generated response reproduces protected expression.

AI distribution

Whether an infringing output is subsequently communicated or commercially exploited.

The Kneschke v LAION decision demonstrates the importance of TDM exceptions for AI training, while GEMA v OpenAI demonstrates the emerging importance of model memorisation and verbatim reproduction. Infopaq supplies the foundational rule that even a relatively short extract can be protected if it contains the author's intellectual creation. (WIPO)

At the EU level, Like Company v Google (C-250/25) is particularly significant because it places several core generative-AI copyright questions directly before the CJEU; as of the latest available information, it remains pending. (IP Helpdesk)

Exam-Ready Keywords

AI copyright infringement, generative AI, LLM, AI training, training dataset, copyright-protected works, reproduction right, communication to the public, making available, memorisation, AI output, text and data mining, DSM Directive, Articles 3 and 4, AI Act, GPAI, copyright opt-out, machine-readable reservation, originality, intellectual creation, substantial reproduction, protected expression, style imitation, AI-generated images, AI-generated music, database rights, moral rights, licensing, damages, injunction, disclosure, evidence, model parameters, training data, provider liability, user liability, TDM exception, three-step test, GEMA v OpenAI, Kneschke v LAION, Infopaq, Pelham, Painer, Football Dataco, Like Company v Google.