A data subject access request (DSAR) lets an individual ask whether an organization processes their personal data and obtain a copy or description of it and how it is used. Personal data from public sources is still personal data under GDPR. section 5.5.2
For generative AI systems, fulfilling a request can be harder than exporting a customer record. Personal information may appear in source records, retrieval indexes, training datasets, or outputs; its influence may also be distributed across model weights. A mathematical embedding is not a human-readable copy of the underlying information. section 5.5 section 5.5.2
Prepare an access-request workflow that can:
- Track, at least broadly, the sources and types of personal data entering and leaving the system. section 5.5.2
- Identify the records and categories of information associated with the requester, including data used for fine-tuning where possible. section 5.5.2
- Coordinate with model vendors through contracts or cooperation; using a third-party model does not remove the deployer's responsibility for customer data it uses. section 5.5.2
- Explain what can be provided and where technical limits prevent a complete account. Handle any separate erasure request without promising selective removal from model weights. section 5.5
A claim with a hallucinated collision
After a claims assistant invented a second collision, Elena requested access to her data, an explanation of its use, and deletion of non-essential information. The compliance team could export her CRM record and claims file, but the retrieval index held unreadable vectors and her accident details had influenced fine-tuned model weights. Records and vector-store chunks could be removed; residual model influence could not be selectively removed without retraining from scratch. section 5.5

