Copyright © Media Law International 2026. All Rights Reserved. Privacy Policy Contact Us Unsubscribe
Guide to the Global Leaders in Media Law Practice
REGIONS WE COVER
Middle East and North Africa
North America
Asia-Pacific
Western Europe
Central and Eastern Europe
AI and International Licensing Agreements – Key Considerations for Contract Design
At the same time, a growing licensing market has emerged: rights holders across all industries are entering into licensing agreements with AI companies to monetize and legally secure the use of their content.
However, as new business models emerge, new legal challenges arise. How can a licensor control the use of its content, and how can it prevent its licensee’s AI system from cannibalizing its own core business? The following article addresses these questions by examining selected AI licensing models and highlighting important considerations for drafting contracts in an international context.
From Litigation to Licensing – Reasons for AI Licenses
Before examining the various AI licensing models in more detail, it is first necessary to examine the motives behind the interest of both licensees and licensors in AI licenses. On the licensee's side, it is no longer only major AI model providers, such as OpenAI or Meta, seeking such licenses, but also companies of all sizes.
One reason for entering into such license agreements is legal certainty. Licensees obtain a contractual basis for their use and no longer need to grapple with the scope of statutory exceptions of copyright, such as the text and data mining exception at EU level or fair use in the US.
Given the territorial nature of copyright, the use of protected content for AI purposes can be classified differently depending on the applicable jurisdiction, making contractual arrangements with worldwide enforceability attractive. Another reason is that licensees can demand warranties from their licensors that using the licensed content within the agreed scope will not infringe third-party rights.
In addition to legal considerations, there are also practical reasons. Licensees often gain access to (otherwise inaccessible) data and/or data of specific quality only through respective contractual arrangements. For example, license agreements typically provide access via a programming interface, or API. This provides the licensee with significantly faster and more convenient access to high-quality, pre-curated data than scraping or scrolling a platform or even the internet would.
Furthermore, the regulatory requirements of the EU AI Regulation (Regulation (EU) 2024/1689) have led companies to tighten their compliance standards. Any company offering an AI-based system to users in the EU is subject to the transparency and compliance requirements of the EU AI Regulation. Additionally, customers of AI systems are becoming increasingly demanding. Questions about trustworthy training materials and the legality of AI systems are being raised more frequently, particularly in the B2B sector, where contractual warranties about the origin and licensing of training data are expected.
Ultimately, an AI license allows the licensee to adapt their business model to their specific needs. For example, if the licensee wishes to generate output containing parts of original works, such as text snippets, they can do so legally via a license.
For licensors and rights holders of legally protected content, participation in the licensing market offers the opportunity to appropriately monetize the use of their content in connection with AI and negotiate customized remuneration models depending on the licensed purpose.
Rights holders also gain the ability to control the scope of use through license agreements. They can secure contractual rights that protect their core business. They can impose corresponding obligations on the licensee to prevent the licensed AI system from ultimately cannibalizing the exploitation of the original content.
Key Considerations in Contract Design
The following section illustrates the possibilities and challenges of contract design and drafting in an international context using two examples of AI licensing models.
First, there are AI training licenses.
These licenses are characterized by the fact that the licensed content is intended to serve as input for training an AI model. Rather than reproducing the licensed content in the output, these licenses make the AI model more capable of handling queries of all kinds. They train the “brain” of the AI, regardless of the type of output.
On the other hand, there are models where the IP is intended to improve specific content generation. These models include Retrieval Augmented Generation (RAG) licenses. In these models, content is stored in a separate knowledge database. When queried, an independent AI system searches the knowledge database for relevant information, which is then transferred to the AI model in vectorized form. The AI model generates a precise, context-sensitive response based on this information.
1. Input: From copyright protection to copyright and data protection
Use of content in connection with AI contains a variety of different stages. Each stage must be assessed differently when assessing the legal implications and statutory protection by copyright law may provide the rightsholder with different levels of protection. This directly affects the strategy for contract design.
The first stage involves using the original work. In both AI licensing models, a series of reproduction and processing operations take place. For example, the original works are first reproduced and stored in databases, either alone or with other data. Then, the works are converted from their original format (text or image) into a numerical format (so-called 'vectorization').
These uses are relevant under copyright law and must be addressed in a license. At this stage of processing, the licensor and original rightsholder can rely on its exclusive copyright in the content. However, legal views differ between countries globally as to whether and under which circumstances, such copyright relevant act of reproduction and adaptation fall within exceptions of copyright. Therefore, licensors are advised to explicitly state in the license that any reproduction or other use of the original work beyond what is allowed under the license is strictly prohibited.
The result of this initial stage of processing is a data set which represents the second stage. The original works are no longer recognizable as copies, but exist as vectorized data with numerical weights and parameters. However, the original work's value has been fully incorporated into the dataset and might even be reproducible from such data.
This presents the licensor with a key challenge: While they agree that the vectorized data may be used for the licensed purposes - such as for the specific AI system in training licenses or for output generation in RAG - they have a significant interest in ensuring that the datasets are not used beyond the agreed purpose or resold to third parties.
In most cases, the licensor will not be able to prevent the further use of the datasets on the basis of their copyright in the original works, unless the use results in full reproduction. Copyright protection covers the rights holder against unauthorized use of the original work or adaptations of the work, provided that the original work is still recognizable in the adaptation. Courts differ on whether trained AI models contain copyright relevant reproductions.
In the German GEMA v. Open AI case (Munich Regional Court I, final judgment of November 11, 2025 – 42 O 14139/24), the court found reproductions existed in the trained model because the AI system reproduced the original work, having memorized the original content. In the UK, Getty Images vs. Stability AI (Getty Images (US) Inc & Ors v. Stability AI Limited [2025] EWHC 2863 (Ch).), the court ruled the model is merely a dataset of numerical parameters and can therefore not include a reproduction of the original work.
This is why it is a key requirement to restrict and control the disclosure and further use of any datasets derived from the permitted processing of the original work, regardless of the legal classification at the location where they were created. Contractual arrangements thus assume a central protective function.
Possible contractual restrictions include:
- “The licensee is only permitted to use the created datasets internally for the agreed contractual purpose (RAG or model training). In particular, the licensee shall not redistribute, transfer or disclose the created datasets or use them commercially in any other way.”
In the case of an AI training license, the data sets are then used in a third stage for training an AI model. The result of the training is a dataset that contains the trained AI model. As this model represents a new result independent of the original dataset, it must be regulated contractually separately and independently of the second processing stage.
Possible contractual restrictions include:
- “The licensee is permitted to utilize the created AI model exclusively for internal business operations.”
- “The licensee may deploy the AI model as part of and integrated in the licensee's agreed customer-facing products under the conditions set out in this agreement; however, sub-licensing the trained AI model as a standalone product is strictly prohibited.”
2. Output Limitations
Both license models use the licensed data to generate output upon a specific user prompt into the AI system.
Licensees may want to use parts of the original work, either in original or edited form, in their output. For example, RAG systems in academia often provide users with snippets of the original content, which serve as references for the information contained in the output. Even with AI training licenses, there may be an interest in using the original work within the output, e.g. to generate images incorporating elements of the original work.
The licensor, on the other hand, also wishes to retain control over the scope of use in the output. The licensor has a vested interest in ensuring that the original work is not substituted by the licensee's AI system and thus cannibalized. If the license permits the use of the original work in the output (including derivatives), the license should regulate the scope of such use in the output.
Even if the original work is not used directly in the output, there may be an interest in agreeing to further contractual obligations of licensee. For instance, if the output consists of short summaries of scientific articles, the licensor has a legitimate interest in ensuring that the licensee does not establish a business model that would hinder the distribution of the original articles.
Contractual restrictions regarding the output are frequently established through contractually defined guardrails, i.e. restrictions that the licensee must implement technically within their AI or RAG system, for example through filters. Possible guardrails include:
- “The output shall not be used for commercial purposes, including but not limited to redistribution.”
- “The original work must not be reproduced in the output.”
- “Verbatim reproductions of the original work in the output must be limited to a specific number of words.”
- “The output must not serve as a replacement for or substitution of the original work and must not compete with the distribution of the original work.”
In addition, licensee's obligations often include attribution and labeling obligations. The licensee must then provide details of the licensor and/or author when displaying the original work in the output, in particular to ensure that the end user can access the original source content independently.
3. Removal of Content
In certain circumstances, licensors may need to withdraw content. For example, content may need to be withdrawn if the rights revert to the original author. If the licensor wishes to ensure this withdrawal with the licensee, this must be contractually agreed. This can also safeguard the licensor from potential claims by the licensee and third parties.
In the RAG license model, withdrawal is technically feasible. The licensed content is stored separately from the AI system and is only forwarded to the AI system upon request. As a result, the pertinent data can be deleted from the database, and it will no longer be accessible in the RAG system in the future.
However, if content is used to train AI models, withdrawal becomes more challenging. In this case, the original work is no longer a "tangible" dataset but has been incorporated into the AI model's training process. It is essential to consult with the licensee to assess the technical feasibility of deletion and whether appropriate technical solutions can prevent the material from being used in output.
Outlook
In the context of AI licensing, the scope of protection of original works under copyright law may not cover all stages of AI technical processing. Contractual arrangements thus assume a central protective function.
One should also monitor the legal and technical landscape: case law on the copyright classification of AI training processes will continue to evolve, and the EU AI Regulation will further tighten licensing requirements. Technological advances, such as "machine unlearning," may improve the enforceability of contractual content removal rights in the future. Rightsholders as well as AI companies that proactively design their AI licensing arrangements to take into account all stages of AI processing will be best positioned to participate in these evolving opportunities.
Few areas of law are evolving as dynamically as the law surrounding artificial intelligence, with lawsuits worth billions and licensing deals worth hundreds of millions. Copyright lawsuits concerning the training of AI models are ongoing worldwide, initially started in the United States and the United Kingdom, now spanning around the world including India, China, Japan, Singapore and Germany.
Helena Jochem, LL.M.
AI and International Licensing Agreements - Key Considerations for
Contract Design
Written by Helena Jochem, LL.M.
LAUSEN
FEATURED ARTICLE
GERMANY
Helena Jochem, LL.M. advises and supports clients from the film, television, publishing and merchandising sectors in all copyright related matters and legal aspects of Artificial Intelligence. One focus of her work is providing support in setting up and negotiating licensing agreements and in implementing new business ideas and complex licensing structures on a global level. After her graduation in Germany, she obtained a Master of Laws at University of Wellington (New Zealand) and has been with LAUSEN since 2016.
Biography
GLOBAL
RECOGNITION