Data Retention and Deletion for AI Systems

Objective

Define and enforce retention and deletion of AI-related data (training datasets, fine-tuning data, embeddings, prompts, completions, model artifacts, logs) per a documented schedule aligned to legal, regulatory, and operational requirements.

Applicability

AI types

LLM, Agentic AI, Traditional ML, Computer Vision, Generative AI, Multi-modal, Recommender

Deployment models

SaaS, Self-hosted, Hybrid, Edge

Lifecycle stages

Operation & Monitoring, Retirement

Risk domains

Data

Regulatory regimes

GDPR, PIPEDA, HIPAA, EU AI Act, Banking, ISO 42001

Company size

SMB, MidMarket, Enterprise

Rationale

AI systems create new categories of data that traditional retention schedules do not cover: embeddings of personal data, prompts containing customer content, completions reflecting model behavior, intermediate checkpoints. Each category has distinct legal and operational considerations. Failure to apply retention and deletion creates unnecessary privacy exposure, complicates DSR responses, and exceeds the data minimization principles of GDPR and similar laws.

Control narrative

The organization maintains an AI Data Retention Schedule covering: training/fine-tuning datasets, model checkpoints and final weights, evaluation datasets, prompts and completions captured operationally, embeddings derived from organizational or personal data, conversation/session logs, and feedback (RLHF) data. The schedule specifies retention period, deletion verification method, exceptions (legal hold, regulatory required retention), and roles. Vendor-managed services have analogous configurations and contractual deletion guarantees with deletion certificates. Embeddings derived from personal data are treated as personal data and deleted when source data is erased or anonymized. Logs of deletion are themselves retained for audit.

Test of Design

Procedures

1. Obtain the AI Data Retention Schedule and confirm coverage of all AI-related data categories.
2. Confirm schedule is aligned to applicable privacy law, sectoral retention rules, and ISO/IEC 42001 requirements.
3. Confirm deletion verification methods are documented (e.g., cryptographic erasure, key destruction, attestation).
4. Confirm vendor configurations are aligned and deletion certificates are required contractually.
5. Confirm exception handling (legal hold, regulatory holds) and re-evaluation cadence.

Inquiries

• Who owns AI data retention?
• How are embeddings of personal data handled when source data is erased?
• How is deletion of training data on backup media or with third parties verified?
• How are vendor deletion certificates obtained and stored?

Inspections

• AI Data Retention Schedule.
• Vendor contracts with deletion guarantees.
• Sample deletion records and certificates.
• Configuration evidence for systems with automatic deletion.

Test of Operating Effectiveness

Procedures

1. From the inventory, sample AI systems.
2. For each, confirm retention configuration matches the schedule for each applicable data category.
3. For data scheduled for deletion in the audit period, confirm deletion occurred and was verified.
4. Sample vendor deletion certificates and confirm receipt and storage.
5. For systems with embeddings of personal data, confirm process for erasure-triggered embedding deletion.
6. Confirm legal-hold overrides were applied appropriately and lifted when no longer required.

Sample-size guidance

Evidence requirements

Required

• AI Data Retention Schedule PDF/Excel · At fieldwork
• Deletion records for audit-period deletions System export · Per sample
• Vendor deletion certificates PDF · Per sample

Supporting

• Retention configuration evidence Config screenshot · Per sample
• Legal-hold records affecting AI data Legal system export · Per applicable case

Retention: Deletion records: 7 years for regulated environments; 3 years otherwise

Framework mappings

Related controls

• AI-CTRL-002

  Training Data Provenance and Lineage

• AI-CTRL-013

  Personal Data Minimization in Training and Inference

• AI-CTRL-018

  AI System Decommissioning

References

• GDPR Article 17 — Right to erasure
• ISO/IEC 42001:2023, clause 7.4
• NIST AI RMF — MEASURE-2.10