Generative AI: Key IP considerations from a German litigation perspective

IP Infringement Risks in Training GenAI Systems

GenAI systems often utilise extensive datasets that may include copyright-protected material or other IP-protected data, such as databases. Training a GenAI system with copyrighted works without the copyright owner's consent, especially if it involves copying a substantial part of the work, can constitute an infringement in many jurisdictions, unless specific defences or exceptions apply.

Relevant Defences and Exceptions 

Whether a defence or exception applies will depend on the jurisdiction in which the training occurs. In the EU, the Digital Copyright Directive (EU) 2019/790 allows text and data mining of lawfully accessible works for any purpose unless the rights holder has expressly reserved their rights. The German Copyright Act (Urheberrechtsgesetz – UrhG) provides specific exceptions:

• Section 44b UrhG: Allows reproductions for automated analysis unless the author has expressly reserved their rights. The work must be lawfully accessed and deleted after training.
• Section 44a UrhG: Permits temporary reproductions that have no individual economic value, for example data mining. AI training can only be based on this exception if no training corpus is created, and the AI learns from works stored in the working memory, which is deleted again immediately afterward.
• Section 60d UrhG: Allows privileged research organizations to reproduce works even if the author has reserved their rights.

In a landmark judgment with far reaching ramifications, a German court recently held that the copying of images by Large-scale Artificial Intelligence Open Network (LAION) – a nonprofit organization that provides datasets, tools and models to liberate machine learning research – did not infringe copyright law. Based on the arguments in the first hearing of the case in July 2024, it was widely expected that the case would be decided based on the general text and data mining (TDM) exception to copyright infringement in Section 44b UrhG.  However, the court instead dismissed Kneschke’s claim on the basis of the “text and data mining for scientific research purpose” exception in section 60d UrhG.

Risk of Loss of Confidentiality in GenAI Prompts

Using public GenAI systems risks losing control over confidential prompts, as providers can reuse this data without restriction, making it no longer confidential. Similar to internet search engines or translation websites, organizations should have policies and training programs to manage confidential information in public GenAI systems.

Reviewing Supply Agreement Restrictions

For private or enterprise deployments, supply agreements should include terms preventing the provider from reusing or disclosing inputs and the deployer's data used for training. It is crucial to review the scope of these restrictions, the provider’s liability limitations, and the governing law and jurisdiction.

Intellectual Property Rights in GenAI Outputs

General Risks regarding IP rights in GenAI Output

There are several general risks to consider regarding the IP protection of GenAI outputs:

• If the GenAI output is not protected by IP rights, deployers should take care of how the output is used as it might be unprotected against third-party copying without substantial human revision.
• Where IP rights may apply, it should be ensured that the supply agreement clarifies who owns the IP rights in the output and does not unduly restrict usage rights.
• Using GenAI for invention development may risk maintaining confidentiality, especially in public deployments, potentially affecting the novelty and patentability of the invention.

Generally, GenAI output is not protected by German copyright law due to the lack of human authorship. However, exceptions include:

• When a user makes specific creative decisions, the user is deemed the author.
• When the AI's output is part of an overall project used as an artistic element.

Neighbouring rights protecting investments can protect AI-generated outputs, applicable to sound recordings, databases, press products, and moving images.

Patent protection in GenAI Output

German patent law is neutral with regard to technical aids used in the process of making an invention. Therefore, computer-generated inventions are patentable under applicable law. The inventor of a computer-generated invention could be the user of the AI system. In addition to the user, the provider or the producer of the training data may be a co-inventor if he or she has contributed significantly to the generated invention by designing the AI.

However, AI cannot be named as inventor on patent applications. The Legal Board of Appeal of the European Patent Office has confirmed in cases J 8/20 and J 9/20 (AI DABUS designated as inventor) that, under the European Patent Convention (EPC), an inventor designated in a patent application must be a human. Both the naming as inventor and the transfer of the rights to the invention (and in particular, the right to the patent conferred on the applicant) are considered not possible for an AI system because it lacks legal personality.

The German Federal Court of Justice (Federal Supreme Court, June 11, 2024 – X ZB 5/22) confirmed this in its most recent decision and furthermore stated that the designation of a natural person as inventor is also possible and necessary if a system with artificial intelligence has been used to find the claimed technical teaching.

The German Federal Patent Court (Federal Patent Court, November 11, 2021 – 11 W (pat) 5/21) also did not permit the sole naming of the AI system but allowed a co-naming of the AI system (without having to answer the question of the legal capacity of an AI system).

Infringement Risks in Creation and Use of GenAI Outputs

When a GenAI system's output closely resembles a third party's copyrighted work, the following considerations are critical:

• Whether the copyright work was copied should be determined.
• If it was copied, whether this constitutes infringement.
• Whether there is an AI actor who is responsible for the infringement.

Determining the Production Process

Understanding whether the AI system was trained with or accessed the copyrighted work and how this influenced the output is essential. Developers have not disclosed detailed operations, making GenAI systems largely a "black box" to deployers and users.

Evidence Requirements

To consider the output as copyright infringement, it must be shown that the output incorporates and reproduces a third party's work recognisably. According to sections 16 and 23 of the German Copyright Act, this depends on whether the work is identifiable in its unique qualities (reproduction) or not (free use). Evidence must demonstrate that the original work was used to create the output and is still recognisable as a reproduction.

Potential Liability

Deployers may be exposed to primary liability for copyright infringement and secondary liability for possession of an infringing GenAI system, provided they know or have reason to believe the system includes infringing content.

Under German copyright law, infringement not only requires temporary reproduction of a copyrighted work in the training process, but the person initiating and executing the reproduction (carrying out the training) would be held liable for any infringement in relation to the act of training.

Conclusion

Generative AI systems introduce various copyright issues for rights holders and users. German courts have yet to rule definitively on AI authorship and the use of protected content for training. As regulatory debates continue, stakeholders must assess risks and verify AI product rights to avoid infringements.

Balancing IP protection with AI development involves ensuring fair compensation for copyright owners, ethical data sourcing, and safeguarding data value. The ongoing regulatory debates underscore the challenges in harmonising AI investment with existing IP rights, highlighting the need for an international solution.