b'reduce the error. 48As this process continues across millions and millions of words, the weights become increasingly precise, and the models output begins to sound more like the data it was trained on. 49Conversation and Document Storage GAI tools are typically built around traditional storage systems that are not significantly different from any other cloud-based service. In most GAI tools, these databases hold conversation history, documents, and other interactions with the system. Weve illustrated these components in Figure 2 above using orange shading. Following the flow shown in Figure 2, the user might enter text or a document into the user interface (such as, for example, asking ChatGPT a question). That text or document is then formatted into an AI-compatible form and sent to the transformer model to be processed. The transformer returns a human-readable response. The combination of the users input (orprompt ) and the transformers output now forms a conversation which is stored in the same type of cloud storage that any other text-based information might be stored. Retention settings for these databases determine how long content remains accessible after a user deletes it, and how long it is retained after an account is closed. Deleted items may also remain in backup storage for weeks or months before being permanently removed, depending on the providers policies. 50Based on our review of several major model providers, deleted conversations are generally removed from the providers systems within a fixed period (often about 30 days). Enterprise-grade systems may offer additional protection, such as: Administrativecontrolsthatallowtheorganizationtosetretentionlimitsforconversations, removing this discretion from individual users; Account configurations that give the customer direct, unilateral control over storage areas holding documents, conversation history, and other model outputs; and Encryption of traditional storage containers at rest within the providers platform. Because the traditional database layer often contains the most complete record of a users interactions with theGAItool,itiscriticalforlawyerstodeterminewhethertheseretentionperiodsandcontrolsare appropriate for the sensitivity of the data being stored. There are a few reasons we feel that lawyers should understand the distinction between AI components and traditional storage and networking components. First, having this understanding can help demystify the overall idea of using GAI tools. Its helpful to think of these systems as traditional computer and cloud services that have the addition of transformer-based processors. Second, the AI components used to process and store 48 Matthew Burtell & Helen Toner, Part 1 , Ctr. for Sec. & Emerging Tech., Georgetown Univ. (Mar. 8, 2024), https://cset.georgetown.edu/article/the-surprising-power-of-next-word-prTedhiec Stiuornp-rlaisrigneg- lPaonwguera goef -Nmeoxtd-eWlso-redx pPlraeidniecdti-opna:r Lt-a1r g(lea sLta vnigsuitaegde J uMnoed 1el8s,E2x0p2la5i)n. ed, 49 See ,e.g .,ChatGPT Consumer Terms of Use , OpenAI (Dec. 11, 2024), https://openai.com/policies/terms-of-use (providing that, unless users opt out, OpenAI may use content submitted through ChatGPT to improve its models). 50 AI Logs and Legal Holds: How to Build a Defensible Retention Strategy , Hanzo (Sept. 17, 2025), https://www.jdsupra.com/legalnews/ai-logs-and-legal-holds-how-to-build-a-7261821 (last visited October 1, 2025). Page | 27'