b'Any software provider can add AI functionality to their application using a transmission channel called an Application Programming Interface , orAPI . When developers connect their software to a model providers API, they create a two-way communication between your data and one or more underlying transformer model providers (such as OpenAI, Anthropic, or Google).Distinguishing between the underlying model and the products that integrate these models is important. Lawyers often interact with GAI tools throughbranded interfaces, but the actual transformermodel is typically managed directly by the underlying model provider. For example, consider a PDF application that has an AI feature allowing you to summarize PDF documents. Behind the scenes, that feature most likely works as follows: User clicks Summarize Document The PDF application sends the document content to a model provider (such as OpenAI) using that model providers API interface The model provider processes the document with their transformer model, and sends a summary back to the PDF application The PDF application then displays the transformer models output to the user Altho ughtheuserinteractsdirectlywiththeapplication,theunderlyingAIprocessingoccurson infrastructure controlled by the model provider and is subject to whatever privacy and security assurances are negotiated between the application provider and the model provider. In short, applications with AI integrations may give the appearance of operating in an isolated environment, but this is rarely the case. In most cases, these arrangements simply shift the relationship with the transformer model provider from that of your direct vendor to that of your vendors subprocessor. When attorneys evaluate whether a particular GAI tool is suitable for use in their practice, its essential for them to investigate what model is used to create the artificial intelligence experience. A reputable legal services provider should be transparent about the model vendors they use, and they should make that information readily available to lawyers prior to licensing the product. For example, and without endorsing any specific provider, at the time of this writing Thomson Reuters publicly states of its CoCounsel product: We process all interactions with OpenAI GPT and Google Gemini in the U.S. 28 (b) Self-Managed GAI Tools In a self-hosted environment, the transformer model is downloaded and deployed directly onto hardware or cloud infrastructure controlled by the user. This includes models such as GPT-OSS, Mistral, or Llama, which can be run entirely within a firms own servers or secure virtual environments, and in some cases directly onalawyersowncomputer.Self-hostedmodelsofferthehighestlevelofcontroloverdatastorage, processing, and retention, since no information is likely to be transmitted to any third party.28CoCounsel: The Industry-Leading GenAI Assistant for Professionals , Thomson Reuters, https://www.thomsonreuters.com/en/cocounsel (last visited Oct. 4. 2025). Page | 11'