Key Takeaways
- Defining IP rights in AI contracts is critical for both deployment and development projects.
- Deployment agreements should secure customer control over inputs and outputs, while protecting proprietary data from unauthorized reuse.
- Development agreements should address ownership of customized models, the data supplied by each party to the project, and ensure secure destruction of models upon the contract’s expiration.
The deployment of AI, or implementing a trained model to support business activities, introduces complexities concerning the creation and ownership of IP.
Businesses that integrate AI into their operations must thoughtfully address IP rights in the AI agreements.
Organizations with maturing AI strategies are going to develop AI systems that are customized for particular business objectives and trained on their data. When an AI model is customized, whether through fine-tuning a foundation model or using a retrieval augmented generation (RAG) library, for example, the question of IP ownership becomes incredibly nuanced. As a company’s AI program grows, so do the IP considerations to be addressed in the underlying agreements.
1. Defining the AI IP Landscape
An effective AI contract begins with a clear description of the AI system and a thoughtful allocation of the IP rights in each component. These typically include the following terms:
Core Models and Algorithms
Core models and algorithms are the unique logic, architecture, and machine learning models that drive the AI. For AI development contracts, it is critical to differentiate between the pre-existing elements of the AI system brought to the project by the developer and the new algorithms, models, and data created during the development process. It is becoming increasingly common for customers to contribute to the customization of AI models that are trained on their data. In such a situation, AI development agreements should define the “Customer-Trained Model” and identify who owns the IP rights in it and any relevant licensing provisions.
Training Data
Training data is the lifeblood of AI performance. AI contracts should identify the datasets used for training, distinguishing between “Customer Training Data” and “Provider Training Data,” and set forth the ownership rights of each party.
Outputs
In the United States, content generated exclusively by AI cannot obtain copyright or patent protection. Nevertheless, AI contracts should explicitly assign ownership of any and all IP rights in AI outputs to avoid ambiguity. This ensures commercialization rights remain with the customer irrespective of copyrightability. It is common practice for the customer to own the AI output, especially if they use or are derived from the customer’s proprietary data as part of an AI development project.
In sum, it is essential for an AI contract to define and identify the owner of any “Background IP” – the IP that existed before the contract – and any IP created as a result of developing and using the AI system. This prevents disputes over proprietary elements such as pre-developed code libraries, proprietary datasets, or unique configurations.
2. IP Considerations in Agreements to Deploy AI Technology
Deployment agreements usually involve licensing an AI solution rather than building one. The provider typically retains ownership of the model, algorithms and training data, and may make in the agreement about its IP rights in the AI system. In the deployment situation, the customer’s primary focus is securing sufficient rights relating to the use of the AI system by its personnel. In that regard, AI deployment agreements should expressly identify the customer as the owner of all IP rights in the AI inputs (which the agreement may call “prompts”) and the output.
In connection with identifying the IP ownership rights in AI inputs and output, AI deployment agreements should limit how. AI contracts often strictly prohibit the provider from retaining or using inputs and outputs for any purpose outside the contracted service, including for training of its and other AI systems.
3. IP Considerations in AI Development Agreements
Companies are increasingly leveraging proprietary data to develop specialized AI systems. Custom AI development presents unique IP issues, and these projects require careful drafting to avoid ownership disputes.
A customer’s direct contributions to customization, such as proprietary datasets supplied for fine-tuning or documents used as the knowledge base for RAG, are almost universally considered the customer’s IP. An important contractual IP consideration arises from the effect of combining these elements, which leads to the following scenario. Is the fine-tuned model itself, which now incorporates the customer’s data and reflects its specific domain, 1) owned by the customer, 2) jointly owned between the customer and developer, or 3) does the customer merely obtain a license to use its version of the model from the developer? Clear contractual language is essential to define what is owned and by who, the scope of any licenses, as well as limitations on use, and whether the developer can use insights gleaned from the customer’s data to improve its general model or provide AI system services to third parties without violating the customer’s IP.
In addition, the ownership of IP rights in the outputs generated by a customized AI model is an important point of negotiation. Customers invariably want to own the outputs generated by the customized model, especially if the outputs are commercially valuable. While RAG and fine-tuning enhance model capabilities, the use of these techniques adds another layer of complexity to IP ownership. Specifically, a RAG library may contain more than just the customer’s information and include publicly available sources (e.g., laws, regulations, industry standards), or materials licensed from a third party. Likewise, an AI model can be fine-tuned with datasets from external sources, such as websites, APIs, patents, and third-party databases. Clauses in AI development agreements relating to the ownership of outputs and related IP have to account for this situation.
It is worth noting that a party may want to use technological measures to streamline the IP ownership of AI outputs. For example, a contract to develop a customized AI model may require that outputs include only the customer’s proprietary data and information derived solely therefrom/ This approach strengthens the customer’s claim to own the IP in the outputs by mitigating the risk of inadvertently co-mingling its proprietary information with other data. To the extent, however, that output includes any developer or third-party data, prospective users of the AI system should negotiate for a license to use such portions of the output for its internal business purposes.
Because custom AI models are likely to house a customer’s proprietary information, AI development agreements should spell out what happens to the model and its data libraries when the contract ends. AI development agreements must include explicit provisions for the secure destruction or deletion of the fine-tuned model and any associated RAG databases upon contract termination. This is vital for protecting the customer’s confidential and proprietary data embedded within the fine-tuned model’s parameters or residing in the RAG knowledge base, and preventing it from being disclosed to or used by third parties. Also, the contract should specify the methods of deletion (e.g., cryptographic erasure, physical destruction of storage media), and require a verifiable certification from the developer confirming that all customer data, including any derivatives within the AI model, has been irrevocably erased from their systems and backups.
The post Securing Innovation: Key IP Contract Clauses for AI Deployment and Development appeared first on Contract Nerds.