Copyright and AI in the EU: TDM Exceptions, Authorship, and Risks

#

Research Date: February 25, 2026 Jurisdiction: European Union / Portugal Scope: Copyright implications of AI-generated content

Executive Summary

#

Copyright law in the EU distinguishes between AI training (covered by TDM exceptions) and AI output (subject to standard copyright rules). Key findings:

• Text and Data Mining (TDM) exceptions (Articles 3-4, DSM Directive) apply to AI training, not to publishing output
• AI-generated content without substantial human contribution is not protected by copyright in EU
• Human authorship requirement — only “author’s own intellectual creation” qualifies for protection (Infopaq standard)
• Recent case law: GEMA v. OpenAI (Germany) — “memorization” is copyright infringement; Kneschke v. LAION (Germany) — machine-readable opt-out required
• CJEU case pending: Like Company v. Google (C-250/25) — first EU AI copyright case, decision expected 2026-2027

1. DSM Directive: TDM Exceptions

#

1.1 General Context

#

Directive (EU) 2019/790 of April 17, 2019 on copyright and related rights in the Digital Single Market (DSM Directive) introduced two key exceptions for text and data mining (TDM).

TDM Definition (Article 2(2)): any automated analytical technique aimed at analyzing text and data in digital form to generate information, including but not limited to patterns, trends and correlations.

1.2 Article 3 — TDM for Scientific Research

#

Subjects: Research organizations and cultural heritage institutions (libraries, museums, archives, film archives).

Conditions:

• Applies only to organizations acting in public interest
• Materials must be lawfully accessible
• Exception is mandatory — rights holders cannot opt out
• Only for scientific purposes

Relevance for content creators: Article 3 not directly applicable, as individual publishers are not research organizations or cultural heritage institutions. However, it applies to AI providers who may use research partnerships.

1.3 Article 4 — TDM for Any Purpose

#

Subjects: Any person, including commercial organizations and individuals.

Conditions:

• Materials must be lawfully accessible
• Rights holders may opt out by reserving their rights “in an appropriate manner” (machine-readable)
• Article 4(3): reservation must be expressed explicitly, including in machine-readable form

Opt-out Mechanism:

• Robots Exclusion Protocol (robots.txt) — de facto standard, but interim solution
• No single harmonized standard at EU level (as of 2026)
• EUIPO planning creation of centralized machine-readable opt-out registry

1.4 Application to AI Training vs. AI Use

#

Critical Distinction:

Articles 3 and 4 regulate training of AI models, not use of already trained models for content generation. This confirmed by:

• EU AI Act (2024/1689): Article 53 explicitly references Article 4 of DSM Directive in context of training general-purpose AI models (GPAI), confirming TDM exception applicable to AI training.
• Munich Court decision in GEMA v. OpenAI (November 2025): Court distinguished reproduction for data analysis (covered by TDM) and “memorization” in model (not covered).

For AI tool users (i.e., publisher using AI tools): TDM questions primarily concern providers of AI. Publisher using AI tools for writing texts is on output side — here different legal norms apply (authorship, originality, potential third-party rights infringement).

1.5 Transposition in Portugal

#

Portugal transposed Directive 2019/790 via Decreto-Lei n.º 47/2023 of June 19, 2023, effective July 4, 2023. Decree introduced amendments to:

• CDADC (Código do Direito de Autor e dos Direitos Conexos)
• Decreto-Lei n.º 122/2000 (transposition of Directive 96/9/EC on database protection)
• Lei n.º 26/2015 (regulation of collective rights management organizations)

Introduced TDM exceptions:

• Exception for bulk data mining
• Permission for parody, pastiche, and caricature
• Facilitation of cultural heritage preservation
• Permission for didactic illustration in online environment

2. AI Authorship and Copyright Protection

#

2.1 Human Authorship Requirement

#

Portuguese Copyright Law (CDADC):

Article 11 CDADC — Titularidade (Ownership):

“O direito de autor pertence ao criador intelectual da obra” (Copyright belongs to the intellectual creator of the work).

Article 1(1) CDADC — Definition of Work:

Works are intellectual creations in literary, scientific and artistic fields, in whatever form expressed, and as such protected by this Code.

Key Conclusion: Portuguese law unambiguously requires human authorship. AI cannot be “criador intelectual” (intellectual creator). This fully aligns with pan-European approach based on CJEU case law.

2.2 Infopaq Standard — EU Originality Criterion

#

CJEU decision in Infopaq International A/S v. Danske Dagblades Forening (C-5/08, 2009) established unified originality standard for all EU:

Work is original if it represents “author’s own intellectual creation” and reflects their “personal contribution”.

This standard applies in Portugal through:

• Supremacy of EU law (primacy of CJEU interpretation)
• Consistency with Article 1 CDADC

2.3 AI Participation Spectrum

#

2.4 European Parliament Position (2025)

#

Draft report of European Parliament “Copyright and Generative AI” (June 2025):

“AI-generated content (i.e., content created without human authorship) should remain ineligible for copyright protection” — AI-generated content without human authorship should remain outside copyright protection.

However, Parliament clarifies for AI-assisted works:

• Simple prompt input insufficient for authorship claim
• Actions of selecting results, iterating prompts, editing, creatively combining AI-generated content may qualify as intellectual creation

2.5 Moral Rights (Direitos Morais)

#

Article 56 CDADC and following establish author’s moral rights:

Composition of Moral Rights:

1. Right of attribution (direito de paternidade) — right to be named as author
2. Right to integrity of work (direito à integridade da obra) — right to prevent alterations distorting work
3. Right of first disclosure (direito de divulgação) — right to decide when and how work first disclosed
4. Right of withdrawal (direito de retirada) — right to withdraw work from circulation

Characteristics:

• Inalienable (inalienáveis) — cannot be transferred or alienated
• Irrevocable (irrenunciáveis) — cannot be waived
• Perpetual (perpétuos) — survive even after economic rights expire

Consequences for AI Content:

If AI-generated content is not a work under CDADC (i.e., has no human author):

• Moral rights do not arise — no one to be their bearer
• Creates legal vacuum: no moral rights, but also not public domain in traditional sense

If publisher substantially reworked AI content:

• Moral rights arise for publisher as creator of derivative/reworked work
• Attributing authorship with minimal contribution may qualify as fraud or misrepresentation

3. Open-Source Licenses and AI Content

#

3.1 Fundamental Problem

#

Open-source licenses (MIT, Apache-2.0, CC) assume existence of licensor — person holding copyright in licensed material. If AI-generated content not protected by copyright:

• No copyright → nothing to license → license potentially invalid
• Material de facto in public domain
• Any person may use it without complying with license conditions

3.2 MIT and Apache-2.0 for AI-Generated Code

#

MIT License

#

Text: “Permission is hereby granted, free of charge, to any person obtaining a copy of this software…”

Problems:

• License contains copyright notice — “Copyright (c) [year] [name]”
• If code generated by AI without substantial human contribution, licensor may have no copyright
• Without copyright, license becomes legally meaningless (but not harmful)
• Third parties may use code without complying with MIT conditions (but unlikely anyone will contest)

In practice: MIT license on AI-assisted code where human made substantial contribution (architecture, logic, selection, review) is legally valid for human contribution.

Apache License 2.0

#

Additional conditions: Patent licensing (patent grant) and requirement to note modifications.

Problems:

• Same as MIT, plus: AI patent license impossible (AI cannot own patents)
• If code substantially reworked by human — license valid for reworked part

3.3 Creative Commons for AI Content

#

Creative Commons Position (as of 2025-2026)

#

Creative Commons took following positions:

1. AI-generated content and CC licenses:

   • CC licenses require copyright holder as licensor
   • If content not protected by copyright (purely AI-generated), CC license technically inapplicable
   • CC supports principle that “copyright and related rights are unwarranted for AI-generated outputs”

2. Using CC content for AI training:

   • CC believes “the use of works to train AI should be considered non-infringing by default”
   • Strictly from copyright perspective, special permission from licensor for AI training not required to extent such permission is necessary at all

3. ShareAlike and AI training:

   • Research by Open Future (2025): most ShareAlike/CopyLeft licenses ineffective when licensed materials used for AI training
   • Reason: if training falls under TDM exceptions, license conditions don’t apply to this use

Specific CC Licenses and AI Content

#

3.4 Attribution Requirements for AI-Generated Text

#

Current situation:

• No legal requirement in EU/Portugal to indicate AI as “author”
• EU AI Act requires labeling AI-generated content (but obligation lies with AI providers, not users)
• Some CC licenses (CC BY) require attribution — but if content AI-generated, attribution to AI legally meaningless (AI cannot be rights holder)

Practical recommendation: Indicating AI participation is matter of ethics and transparency, not legal copyright requirement. However, EU AI Act may create such requirement through AI regulation line (not copyright).

4. Copyright Infringement Risks

#

4.1 Risk of Reproducing Protected Material in AI Output

#

Issue: LLMs trained on massive text corpora, including protected works. Risk exists that AI output will contain fragments substantially matching protected works of third parties.

Munich Court decision in GEMA v. OpenAI (42 O 14139/24, November 11, 2025) established:

• “Memorization” of protected works in AI model is reproduction under copyright law
• TDM exception does not cover reproduction through memorization in model output
• OpenAI ordered to cease such actions, pay damages, provide information on usage volumes

For AI user (publisher):

Reproduction risk exists when:

• Asking AI to paraphrase/summarize specific sources
• Generating text on narrow topic with few source materials
• Copying AI output without checking for plagiarism

4.2 Liability for Publishing AI Content Infringing Third-Party Rights

#

Who is liable?

Under EU and Portuguese law, liability may fall on:

1. AI provider — for training model on protected content and for “memorization”
2. AI user (publisher) — for publishing content infringing third-party rights

Articles 195-199 CDADC qualify unauthorized use of work as:

• Usurpação (usurpation) — using another’s work under own name
• Contrafacção (counterfeiting) — unauthorized reproduction

Both forms entail civil and criminal liability.

For publisher: Publishing AI-generated content that reproduces protected works may lead to:

• Copyright infringement lawsuit
• Obligation to remove content
• Compensation for damages
• In theory — criminal liability (though in practice unlikely for good-faith publisher)

4.3 Safe Harbors and Defenses

#

E-Commerce Directive (2000/31/EC) and Digital Services Act (DSA, 2022/2065):

Safe harbor for hosting providers: liability exemption if provider:

• Did not know about infringement
• Acted expeditiously after receiving notice (notice-and-takedown)

Applicability to individual publisher:

• Publisher publishing own content (including AI-assisted) — not a hosting provider
• Safe harbor does not apply — publisher bears direct liability for published content
• Ignorance of AI reproduction does not exempt from liability (strict copyright liability)

Possible Defenses:

• Good faith — mitigating circumstance, but not exemption
• De minimis — if matching fragment insignificant
• Independent creation — if can prove AI created text independently (difficult to prove)
• TDM exceptions — apply to training, not to publishing output

5. Recent Case Law

#

5.1 Germany: GEMA v. OpenAI (Munich Regional Court, November 11, 2025)

#

Essence: GEMA (German music rights management organization) sued OpenAI for using song lyrics in ChatGPT training.

Decision:

• “Memorization” of linguistic works in AI model is reproduction under copyright — both in model processing and in output to users
• TDM exception covers reproduction when creating training data, but does NOT cover the training process itself
• GEMA validly implemented opt-out per §44b UrhG (transposition of Art. 4(3) DSM Directive)
• OpenAI ordered to: cease violations; pay damages; provide information on usage volumes and revenues received
• Penalty for non-compliance: up to €250,000 or imprisonment

Significance: First decision establishing “memorization” of works in AI model is copyright infringement. Critical distinction between “data analysis” (TDM) and “memorization”.

5.2 Germany: Kneschke v. LAION (OLG Hamburg, December 10, 2025)

#

Essence: Photographer Robert Kneschke sued LAION e.V. for including his photos in LAION 5B dataset for AI training without consent.

First instance decision (Hamburg, September 2024): In favor of LAION. Photo use falls under TDM exception for scientific research (§60d UrhG, transposition of Art. 3 DSM Directive).

Appeal decision (OLG Hamburg, December 2025): Confirmed. Additionally court established:

• Opt-out by plaintiff was invalid, as expressed in natural language (not machine-readable)
• Court considered state of technology at time of download (2021): plaintiff did not prove technologies of that time could reliably interpret language restrictions in ToS or source code
• For valid opt-out, machine-readable form necessary

Significance: Confirms TDM exception for scientific purposes applicable to AI training; opt-out must be machine-readable.

5.3 UK: Getty Images v. Stability AI (High Court, November 4, 2025)

#

Essence: Getty Images sued Stability AI for using millions of Getty images to train Stable Diffusion.

Decision:

• Court rejected Getty’s main copyright claims
• Getty admitted lack of evidence that Stable Diffusion training and development conducted in UK
• Before final arguments, Getty withdrew primary copyright and database rights claims
• Court established limited trademark infringement: Getty/iStock watermarks appeared in images generated by early Stable Diffusion versions
• Key legal holding: “article” for secondary infringement purposes may be intangible, but final model trained on protected works outside UK and not containing their copies is not infringing copy

Getty received permission to appeal.

Significance: First AI copyright decision in UK. Territoriality of training matters.

5.4 CJEU: Like Company v. Google Ireland (C-250/25)

#

Preliminary reference from April 3, 2025, Budapest District Court.

Essence: Hungarian publisher Like Company claims Google Gemini (formerly Bard) reproduces and summarizes substantial parts of copyright articles without authorization.

Key Questions to CJEU:

1. Is LLM training reproduction under Art. 2 InfoSoc Directive?
2. If training is reproduction, does it fall under TDM exception Art. 4 DSM Directive?
3. Questions on applicability of press publisher rights

This is first case where CJEU formally considers legal issues specifically concerning generative AI.

Expected decision: late 2026 — 2027. Decision will be binding for all Member State courts (including Portugal).

6. Practical Recommendations

#

6.1 For Content Creators

#

Copyright risk analysis of AI-generated content under CC license:

• Under current EU law: AI-generated content not protected by copyright without significant human contribution
• If content substantially edited by human → may be protected
• CC license on unprotected content = legal uncertainty
• Recommendation: Position content as “AI-assisted, human-edited” (which corresponds to actual process)

Verify AI content doesn’t infringe others’ copyright:

• Commercial AI tools generally don’t quote verbatim protected texts (per their usage policies)
• But control output data for matches with existing texts
• Especially for: scientific descriptions, rules, card interpretations

6.2 For Code Developers

#

AI-assisted code and copyright:

“Sufficient human contribution” criterion:

European Parliament (2025) established for AI-assisted works:

• Insufficient: simply enter prompt
• May be sufficient: selecting results, iterating prompts, editing, creative combining

For code, sufficient contribution may include:

• Defining architecture and design
• Writing specifications and tests
• Code review and refactoring AI output
• Integrating AI fragments into overall structure
• Debugging and fixing errors

6.3 Disclaimers

#

For encyclopedias and informational content:

1. General disclaimer on site (footer/sidebar):

   • “This project is informational resource created by community. Materials prepared using various tools and technologies, including automated data processing. All materials undergo editorial review and verification against available sources. Authors not responsible for accuracy, completeness or currency of published information. Use information at your own risk.”

2. Legal pages: Privacy Policy, Terms of Use, Licensing page

Sources

#

EU Legislation

#

• Directive (EU) 2019/790 — Copyright in the Digital Single Market (DSM Directive)
• Directive 96/9/EC — Legal Protection of Databases
• Directive 2001/29/EC — Information Society Directive (InfoSoc)
• Regulation (EU) 2024/1689 — EU AI Act

Portuguese Legislation

#

• Decreto-Lei n.º 63/85 — CDADC (Código do Direito de Autor e dos Direitos Conexos)
• Decreto-Lei n.º 47/2023 — DSM Directive transposition
• Decreto-Lei n.º 122/2000 — Database legal protection

Case Law

#

• CJEU, C-5/08, Infopaq International A/S v. Danske Dagblades Forening (2009)
• CJEU, C-250/25, Like Company v. Google Ireland (pending, expected 2026-2027)
• LG München I, 42 O 14139/24, GEMA v. OpenAI (November 11, 2025)
• OLG Hamburg, 5 U 104/24, Kneschke v. LAION (December 10, 2025)
• English High Court, Getty Images v. Stability AI (November 4, 2025)

Research and Expert Opinions

#

• European Copyright Society: Opinion on Copyright and Generative AI (January 2025)
• European Parliament: Generative AI and Copyright (July 2025)
• EUIPO: Development of Generative AI from a Copyright Perspective (May 2025)
• Creative Commons: Understanding CC Licenses and Generative AI
• Open Future: The Impact of Share Alike/CopyLeft Licensing on Generative AI

This document prepared for research purposes and is not legal advice. For legal decisions, consult qualified lawyer specializing in intellectual property in Portugal/EU jurisdiction.