A user’s personal information may exist in stored prompts, training datasets, search indexes, evaluation fixtures and learned model behavior. Removing one database row may leave other copies and exposures intact.
Locate each representation
Build an inventory of source records, preprocessing caches, embeddings, fine-tunes, backup stores, provider transcripts, moderation logs and model checkpoints. Distinguish data held directly from information that may be inferable from a model. Each layer can have different legal responsibilities, deletion capabilities and retention exceptions.
Determine the applicable right
The UK ICO explains how individual rights may apply to personal information embedded in machine-learning systems and why removal can involve retraining or even deleting a model. Under EU GDPR, the legal basis, exemptions, identity verification and controller role also matter; not every request requires or permits exactly the same remedy.
Create an auditable response path
Log receipt, identity verification, relevant systems, exemptions assessed, action owner and completion evidence without needlessly storing the person’s sensitive history. Provide accurate information about what was deleted, withheld or retained and why. Avoid promising immediate removal from all derivatives when the system cannot demonstrate it.
Test future reinsertion
A deleted record can reappear if a nightly upstream sync or outdated backup reseeds the training corpus. Use suppression lists where lawful and appropriate and versioned deletion events in ingestion. Verify the next retraining run excludes affected data and that caches expire as represented.
Example
A customer requests removal of chat content used for a fine-tune. The controller deletes retained transcripts, flags the dataset, assesses whether retraining is technically and legally necessary and documents unresolved influence. It does not tell the customer that historical model weights changed simply because a file disappeared.
Sources & Further Reading
- ICO: Ensuring individual rights in AI systems — UK Information Commissioner.
- EDPB Opinion 28/2024: Data protection aspects of AI models — European Data Protection Board.
- Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations — NIST.
Evidence checked 25 October 2026. Applicability: EU/UK legal illustration; other jurisdictions distinct. This material is general education, not legal, medical, investment or other professional advice. No vendor performance test is implied.