Copyright Law: Training Data vs. Grounding Data

With AI being a relatively new technology, nomenclature can get muddy and confusing. Two different concepts can be commonly referred to by the same word in conversationally convenient but imprecise ways. Often we refer to an AI’s “dataset” as all the data used to train an AI model as well as the data that model has access to. A recent Sidley memo draws an important distinction in datasets between training data and grounding data. While both types of data are necessary for a functional AI tool, they present different legal considerations. The memo describes those considerations unique to grounding data:

“From a legal perspective, grounding data presents a distinct set of considerations because it is actively retrieved and potentially reproduced at the time of each query rather than absorbed into the model’s parameters during the development phase. This distinction also has practical consequences postdeployment: A trained model generally continues to function even if access to its training data is later lost or a license expires, because the learned patterns are encoded in the model’s weights. Grounding data, by contrast, must remain continuously accessible, and loss of access or limitations on access to support risk management needs (e.g., cybersecurity, privacy, or other data governance purposes) may directly impair the model’s outputs.”

Training data is what AI developers use to build their models. Grounding data is data we feed to an AI through a database that provides context for the tool. For example, a company may use an AI model for compliance training. That model’s training data likely includes literary works, encyclopedias, and other non-germane materials. These materials allow AI models to identify patterns and give numerical probability to human language. On the other hand, those materials don’t do much to help the AI answer users’ specific questions. That’s where grounding data comes in. Grounding data allows the AI model to retrieve relevant data and present that data to the user. In the case of compliance training, this would include a company’s compliance policies and manuals.

Understanding what data your AI tools use in certain contexts can help you better understand the risks. Grounding data is often highly proprietary and sensitive. Companies should ensure that their AI contracts don’t allow sharing of sensitive grounding data for future AI training purposes.