Using Copyrighted Works to Train AI in Germany (2026): Legal Risks, Licences & Compliance Checklist

The question of AI copyright Germany has moved from academic debate to active courtroom enforcement. A series of rulings by German regional courts in 2024–2026, most notably the Landgericht München I proceedings brought by GEMA against a major model operator, has confirmed that feeding copyrighted music lyrics, images, film scripts and literary texts into generative AI training pipelines can constitute infringement under the Urheberrechtsgesetz (UrhG). Simultaneously, the EU AI Act and Data Act are imposing new transparency, governance and data-provenance obligations on model developers and platform hosts.

For in-house counsel, rights holders and platform operators, the window for reactive compliance is closing: the practical priority now is to audit datasets, secure training-data licences and implement the governance controls that German courts and EU regulators expect.

Executive Summary and Quick Compliance Answer

Is it legal to train AI on copyrighted works in Germany? The short answer is: it depends. German copyright law protects original works created by human authors, and using those works to train AI models typically involves acts of reproduction that engage the right holder’s exclusive rights under the UrhG. Courts have delivered mixed outcomes, some uses of image data have been permitted where no commercial copying occurred, while the reproduction of song lyrics in large language models has been found to infringe. No blanket statutory exception reliably shields commercial AI training in Germany today.

The practical implication is straightforward: assume you need a licence unless you can demonstrate, with documentary evidence, that a specific statutory limitation applies to your particular use case. Three immediate steps will materially reduce your risk exposure:

• Stop training on contested datasets. If you cannot trace provenance and rights clearance for every data source in your pipeline, pause ingestion of those sources until clearance is obtained.
• Audit existing datasets. Inventory every copyrighted work currently stored in or used by your training infrastructure. Map each item to a rights holder and a licence status.
• Acquire licences and powers of attorney. Prioritise high-risk asset classes, song lyrics, music stems, film scripts, published literary works and photographic collections, and negotiate explicit training-data licences with rights holders or collective management organisations.

The sections below set out the statutory framework, enforcement landscape, ownership rules, licensing mechanics and a step-by-step compliance checklist designed for legal and product teams operating in Germany.

Legal Framework: AI Copyright Germany, Statutes, EU Acts and Core Principles

Three interconnected legal instruments define the compliance perimeter for anyone building, deploying or hosting AI models that process copyrighted material in Germany. Understanding how they interact is the foundation for any defensible dataset licensing strategy.

Key Statutes and EU Instruments

• Urheberrechtsgesetz (UrhG), German Copyright Act. The UrhG grants authors exclusive rights over the reproduction, distribution and communication of their original works. Sections 15–24 UrhG define these exclusive rights; Sections 44a–63a set out statutory limitations and exceptions. Critically, the text and data mining exception introduced by the DSM Directive (transposed into § 44b UrhG) permits reproductions for purposes of text and data mining, but rights holders may reserve their rights by machine-readable means, and many major publishers, music labels and image libraries have done exactly that.
• EU AI Act. The AI Act establishes governance, transparency and risk-management obligations for providers of AI systems. For copyright purposes, its most important provisions require general-purpose AI model providers to document training data, maintain sufficiently detailed summaries of copyrighted training content, and comply with Union copyright law, including the rights-reservation mechanism under the DSM Directive.
• EU Data Act. The Data Act governs contractual and technical access to data, including conditions under which data holders must make datasets available. For AI training, it affects the contractual routes through which developers may lawfully obtain access to datasets and negotiate re-use rights, particularly where data is generated by connected products or related services.

How Originality and Human Authorship Are Assessed in Germany

German copyright law requires that a work reflect the “personal intellectual creation” (persönliche geistige Schöpfung) of a human author. This threshold, codified in § 2(2) UrhG, has been interpreted consistently with the CJEU’s landmark ruling in Infopaq International A/S v Danske Dagblades Forening (Case C‑5/08), which established that copyright protection extends to any element that is the “author’s own intellectual creation.” The practical consequence for generative AI copyright Germany is twofold: works fed into a model are almost certainly protected (assuming they meet the originality threshold), while works produced by a model may lack copyright protection unless a human author contributed creative choices that shaped the output. This distinction is fundamental to both licensing and ownership analysis.

Practitioners working on recording and privacy compliance in Germany will recognise the pattern: German law layers sector-specific statutes on top of EU harmonisation, creating a compliance environment that demands careful, jurisdiction-specific analysis rather than reliance on broad EU-level generalisations.

Enforcement Landscape and Selected German Case Law

German courts have moved faster than most European counterparts to address AI training and copyright. The decisions rendered between 2024 and 2026 do not yield a single, clear rule, but they do establish that rights holders have viable claims and that injunctive relief is available. For platform operators and model developers, the enforcement risk is real and immediate.

Timeline of Notable Decisions

The GEMA v OpenAI proceedings, analysed by Friedrich-Alexander University Erlangen-Nürnberg (FAU), represent a critical signpost for the generative AI copyright Germany landscape. The FAU analysis emphasised that the Munich court’s reasoning focused on the act of reproduction during training, not merely on the similarity of model outputs to source material. This means that even where a model does not reproduce a copyrighted work verbatim in its outputs, the prior act of copying the work into the training dataset may independently infringe.

For rights holders, the enforcement toolkit is expanding. Industry observers expect that courts will increasingly entertain claims for injunctive relief (including interim orders), damages computed on licence-analogy principles, and potentially orders requiring model operators to demonstrate that infringing material has been removed from training data or model weights.

Ownership and Protectability of AI-Generated Content

Who owns AI-generated content under German law? Under the UrhG, only a natural person can be an author. A work must constitute a “personal intellectual creation”, which requires human creative choices that shaped the result. Content generated entirely by an AI system, without meaningful human creative input, does not attract copyright protection under current German doctrine. This position is consistent with the CJEU’s originality standard established in Infopaq and subsequent decisions.

The practical question, however, is rarely binary. In most commercial workflows, a human operator selects prompts, curates outputs, edits results or combines AI-generated elements with original material. Where those human contributions meet the originality threshold, the resulting work may qualify for copyright protection, with authorship vesting in the human contributor. The challenge is documenting and evidencing the nature and extent of human creative input at the point of creation.

Practical Contract Clauses for Ownership Allocation

Because the law does not automatically resolve ownership in many AI-assisted creation scenarios, contractual allocation is essential. Employment contracts, commission agreements and collaboration frameworks should address the following points:

• Ownership assignment. Specify that all rights in AI-assisted outputs, to the extent they are protectable, are assigned to the commissioning party or employer upon creation.
• Moral rights acknowledgement. German moral rights (Urheberpersönlichkeitsrecht) cannot be waived, but contracts should clarify that the employer may exercise exploitation rights without restriction, and that the author consents to non-attribution where legally permissible.
• Record of human input. Require contemporaneous documentation of the human creative choices made during each production cycle, prompt engineering records, editorial decisions, curation logs, to support any future claim of originality.
• Fallback licensing. Where a work is later determined not to attract copyright (because human input was insufficient), include a fallback licence granting the commissioning party exclusive use of the non-protectable material.

Licensing Training Datasets: Types, Clauses and Examples

In the current enforcement environment, obtaining a robust training data licence Germany is the single most effective risk-mitigation measure available to model developers and platform operators. Licensing requirements vary by asset class, and the commercial terms are still evolving, but the core legal clauses are now well established.

Asset-Class Licensing Requirements

• Music (lyrics, compositions, recordings). High risk. The GEMA v OpenAI proceedings confirmed that lyrics are protected and that their ingestion into training pipelines engages exclusive rights. Licence directly from publishers, labels or collective management organisations. Include explicit grants for “use in training AI models” and specify whether the licence covers internal research, commercial deployment, or both.
• Film and television (scripts, clips, stills). High risk. Scripts and screenplays are literary works under § 2(1) No. 1 UrhG; audiovisual content engages multiple layers of rights (director, producer, performers). Negotiate with the relevant producer or distributor and ensure the licence covers the specific modes of reproduction involved in dataset curation.
• Images (photographs, illustrations, design works). Medium-to-high risk. Photographs enjoy protection under § 2(1) No. 5 or § 72 UrhG. Many stock-image libraries have reserved their text and data mining rights. Check licence terms carefully; “editorial use” licences typically do not cover AI training.
• Text (news articles, books, academic papers). Medium risk, but watch for rights reservations. Major publishers have widely deployed machine-readable rights-reservation notices under § 44b(3) UrhG, which removes the text and data mining exception for those works.

Sample Clause Bank

The following clauses are starting-point templates. They require adaptation to your specific transaction and review by qualified legal counsel before use.

• Permitted uses clause. “Licensor grants Licensee a non-exclusive, worldwide licence to reproduce and store the Licensed Works solely for the purpose of training, validating and testing Licensee’s AI model(s), including any fine-tuning or transfer-learning processes. This licence extends to [internal research only / commercial deployment, select as applicable].”
• Indemnity clause. “Licensor represents and warrants that it holds all rights necessary to grant this licence and shall indemnify Licensee against any third-party claims arising from a breach of this warranty, including reasonable legal costs.”
• Audit and deletion clause. “Licensee shall maintain auditable records of all Licensed Works ingested into training datasets. Upon termination or expiry, Licensee shall delete all copies of Licensed Works from its training infrastructure within [30] days and certify deletion in writing.”
• Royalty and reporting clause. “Licensee shall pay Licensor a royalty of [€X / percentage] per [model trained / commercial deployment / revenue milestone], payable [quarterly], accompanied by a written report detailing usage metrics.”
• Provenance and chain-of-title clause. “Licensor warrants that it has obtained all necessary clearances, sub-licences and consents from original rights holders for the Licensed Works and shall provide documentary evidence of chain of title upon reasonable request.”

Negotiation Checklist for Dataset Licensing Germany

• Confirm pricing model: flat fee, per-work royalty, revenue share, or hybrid.
• Determine exclusivity: exclusive training licence vs. non-exclusive (affects pricing and competitive risk).
• Address retro-clearance: if you have already trained on unlicensed material, negotiate retrospective authorisation and document the scope covered.
• Specify sublicensing rights: can the licensee sublicense to downstream model users or API customers?
• Set territorial scope: Germany-only, EU-wide, or worldwide.

Platform and Intermediary Liability for AI Copyright Germany

Platform liability AI Germany is not a hypothetical risk, it is an active enforcement reality. Different actors in the AI value chain face different legal exposures, and no blanket safe-harbour provision currently shields platforms from copyright claims arising from AI training activities. The following table maps likely exposure by entity type.

Reporting and Takedown Obligations

When a platform receives a copyright complaint regarding an AI-generated output, the following process should be triggered immediately:

1. Preserve evidence. Log the complaint, the disputed output, and all associated metadata (timestamps, model version, input prompts if available, dataset identifiers).
2. Suspend the disputed endpoint. Where the complaint is credible and the risk of ongoing infringement is high, consider suspending the model endpoint or disabling the specific output functionality pending review.
3. Initiate legal review. Route the complaint to designated legal counsel within 24 hours. Assess whether the output reproduces protected expression, whether a licence covers the source material, and whether a statutory limitation may apply.
4. Notify the rights holder. Acknowledge receipt of the complaint and communicate the platform’s response timeline.
5. Remediate or defend. If infringement is confirmed, remove the offending content, implement technical measures to prevent recurrence, and assess whether retrospective licensing or settlement is appropriate.

Practical Compliance Checklist for Creators, Producers and Platforms

The following checklist is designed for joint use by legal, product and engineering teams. Each item specifies the operational step, the responsible function and the evidence to collect. A downloadable one-page PDF version of this checklist is available, contact our team to request a copy.

1. Dataset audit (Legal + Engineering). Inventory every data source currently ingested or stored in your training pipeline. For each source, record: origin, format, date of acquisition, volume and the identity of the rights holder. Produce a rights map showing licence status (licensed / unlicensed / status unknown).
2. Risk triage (Legal). Classify datasets by asset class and risk level. Highest risk: song lyrics, music stems, film scripts, published literary works, professional photographic collections. Medium risk: news articles, blog content, user-generated content with clear authorship. Lower risk: public-domain material, works with expired copyright, freely licensed (e.g., Creative Commons) content with terms that permit AI training.
3. Licence acquisition strategy (Legal + Commercial). For high-risk and medium-risk material, identify the relevant rights holders or collective management organisations. Pursue direct licences where possible; engage aggregators for pooled licensing; confirm whether any statutory exception (e.g., § 44b UrhG text and data mining) applies and whether rights have been reserved.
4. Contract clauses (Legal). Ensure every training-data licence includes: explicit permitted-use grant, indemnity, provenance warranty, audit and deletion rights, royalty and reporting obligations, and sublicensing terms. Use the sample clause bank above as a starting point.
5. Model governance (Engineering + Legal). Implement documentation protocols for data lineage: which datasets were used to train which model versions, when, and under what licence terms. Maintain version control for both datasets and model weights.
6. Monitoring and remediation (Product + Engineering). Deploy output-similarity checking tools to detect cases where model outputs reproduce substantial portions of copyrighted training data. Establish user-facing reporting channels for rights holders to flag potentially infringing outputs.
7. Litigation readiness (Legal). Preserve all training datasets, licence agreements and correspondence with rights holders under legal hold. Prepare template response letters for cease-and-desist notices and court proceedings.
8. Insurance and indemnity (Legal + Finance). Evaluate IP liability insurance products that cover AI-related copyright claims. Ensure that indemnity provisions in upstream data-supply agreements are robust and enforceable.
9. Transparency and end-user disclosures (Product + Legal). Under the EU AI Act, providers of general-purpose AI models must publish sufficiently detailed summaries of training data. Ensure your public-facing disclosures comply with these requirements and with any additional national transparency obligations.
10. Periodic review schedule (Legal). Assign an owner (typically the General Counsel or Head of IP) to review dataset compliance quarterly. Update the rights map after every new dataset acquisition, model retraining cycle or relevant court decision.

Practical Drafting Aids and Short Templates

Disclaimer: The following clause templates are provided as starting points for negotiation and drafting. They must be adapted to the specific facts and commercial terms of each transaction and reviewed by qualified legal counsel before execution.

• Permitted uses (music). “Licensor grants Licensee the right to reproduce, store and process the Licensed Recordings and associated Lyrics for the sole purpose of training, fine-tuning and evaluating AI models operated by or on behalf of Licensee.”
• Indemnity (general). “Each party shall indemnify the other against all losses, damages and reasonable costs arising from a material breach of its representations, warranties or obligations under this Agreement.”
• Audit and deletion. “Licensee shall permit Licensor (or its appointed auditor) to inspect Licensee’s training infrastructure and records upon [30] days’ written notice, no more than [once] per calendar year, to verify compliance with the terms of this licence.”
• Royalty reporting. “Within [15] business days following the end of each calendar quarter, Licensee shall deliver to Licensor a written report specifying the number of models trained using Licensed Works, the volume of Licensed Works processed and any revenue attributable to models trained using Licensed Works.”

Conclusion: Next Steps and Recommended Governance Timeline

The AI copyright Germany enforcement landscape will continue to evolve rapidly as additional court decisions are handed down and the EU AI Act’s operational provisions take full effect. Platforms, labels, studios and content producers should adopt the following six-month governance roadmap:

• Month 0–1: Complete the dataset audit and rights map. Identify and quarantine unlicensed high-risk material.
• Month 1–3: Initiate licensing negotiations with priority rights holders and collective management organisations. Execute training-data licences for the highest-risk asset classes.
• Month 3–6: Implement model governance documentation, output-monitoring tools and IP liability insurance. Publish AI Act-compliant transparency disclosures.
• Ongoing: Maintain quarterly review cycles, monitor new case law and regulatory guidance, and update licensing and compliance documentation accordingly.

Taking these steps now, before the next wave of enforcement actions and legislative refinements, is the most effective way to protect your organisation, your creative partners and the long-term viability of your AI-powered products in Germany.

Sources

1. German Copyright Act (Urheberrechtsgesetz), Gesetze im Internet (Bundesministerium der Justiz)
2. CJEU, Infopaq International A/S v Danske Dagblades Forening (Case C‑5/08)
3. European Commission, Regulatory Framework for AI (AI Act)
4. European Commission, Data Act
5. Friedrich-Alexander University Erlangen-Nürnberg (FAU), AI Training and Copyright: A Landmark Ruling in Munich