b'____________________________________________ Technical Addendum In the main body of this Guide, we presented the Core Framework for evaluating GAI tools for use with different levels of data sensitivity. That framework is sufficient to responsibly choose and implement GAI toolsinmostlegalpractices.However,somelawyersmaywishtoexploretheseissuesinmoredetail (particularly when considering tools that will store or process highly sensitive or regulated data).This Technical Addendum illustrates how typical GAI systems work and provides a deeper analysis of key GAI safeguards that comprise the Core Framework. The concepts outlined here apply broadly to most GAI tools, including ChatGPT, Claude, Gemini, Copilot, and many AI-driven legal industry tools. At a high level, these systems share the same essential building blocks; thus, understanding this model can help lawyers more easily evaluate and compare the tools they are likely to encounter in practice. Figure 2 illustrates the major components of a typical GAI system: Figure 2: Functional Diagram of a Basic Legal AI System. This model shows how user prompts interact with stored documents and conversation history, using embeddings, retrieval augmented generation, and a transformer-based language model to generate responses.Most of the lawyers experience takes place in the user interface, which is where we log into the application, manage settings, engage in AI conversations, and upload documents for use within the GAI tool. Behind the scenes, however, theres a lot going on.Model Training The Core Framework introduced the transformer model and explained why it is the essence of any GAI tool. We also cautioned lawyers about the risks associated with model training and identified the ability to disable model training as a principal safeguard for protecting client information. Page | 25'