How long you keep things and what happens at the end. Every category of data should have an answer, and ‘forever, just in case’ is a decision with consequences rather than an absence of one. For AI systems this includes prompts, outputs and logs, which people often forget are records.
Technical view
Requires defined periods per data category, tied to purpose and legal basis, with automated deletion and evidence that it ran. Complications specific to AI systems include prompts and completions held by providers, embeddings and vector indexes derived from deleted source records, model checkpoints trained on data now subject to erasure, and backups. Deletion from a trained model is not achievable by deleting the source, which is the practical problem unlearning attempts to address.
Related
Part of AI Terminology Index, section: Data.