An access rights gap arises when privacy law calls for an account of a person’s data, but a trained model cannot reliably retrieve it. It may return fragments, errors, or nothing—even when personal data shaped its training. section 5.5.2
Separate the model’s core weights from retrievable data stores:
- A model prompt may reveal a fragment, invent an answer, or produce no answer; it cannot reliably provide a complete list of a person’s data. section 5.5.2
- A RAG index may point back to documents such as emails, support tickets, chat logs, or HR files. Access workflows can use the underlying documents and explain the role of embeddings, metadata, and logs. section 5.5.2
Jane Smith
Asking “What do you know about Jane Smith?” may yield only a fragment, a plausible invention, or silence. A separate RAG index may instead locate documents associated with her, such as emails. section 5.5.2
