Skip to main content

Pro audit program · v1.0

AI Privacy & Training Data Lawfulness

Quick check on the privacy + lawfulness side of using or building AI — training-data provenance, lawful basis, DSARs against models, and synthetic data alternatives.

  • General target area
  • GDPR / EU AI Act framework
  • 7 controls in this program
  • Cyentrix Cyentrix Trusted Author
Sign in to run → Back to library

About this program

Quick check on the privacy + lawfulness side of using or building AI — training-data provenance, lawful basis, DSARs against models, and synthetic data alternatives.

Risks addressed

  • Critical Training data scraped without a lawful basis
  • High Personal data memorised + leaked by the model
  • High Cannot satisfy a DSAR (erasure) against an embedded model
  • Critical Customer data fine-tuned into a model and made shareable

Controls (7)

  1. Lawful basis recorded for training / fine-tune data

    Critical

    This control requires organizations to document and maintain records of the lawful basis (e.g., consent, legitimate interest, contractual necessity, legal obligation) for collecting, processing, and using data in AI model training and fine-tuning activities. The organization must map each dataset…

    How to test + evidence

    Testing procedure: For every dataset used, the lawful basis is documented (consent, contract, legitimate interest).

    Evidence to collect: RoPA entry per dataset.

  2. Provenance + licence per training dataset

    High

    This control requires organizations to maintain a formal inventory of all training datasets used in AI/ML model development, documenting the origin (provenance) and licensing terms for each dataset. Provenance tracking includes source identification, data lineage, acquisition date, and custodial chain,…

    How to test + evidence

    Testing procedure: Source + licence + scraping legality for every dataset on file.

    Evidence to collect: Dataset register.

  3. PII minimisation pre-training

    High

    PII minimisation pre-training is a data governance control that ensures personally identifiable information is identified, assessed, and removed or anonymised from datasets before they are used to train machine learning or AI models. This control involves implementing automated scanning, manual…

    How to test + evidence

    Testing procedure: PII filtering / hashing applied before training; documented filter coverage.

    Evidence to collect: Pipeline + filter coverage report.

  4. No customer data in shared / global models

    Critical

    This control ensures that customer-specific data, including personally identifiable information (PII), protected health information (PHI), or proprietary business data, is never used to train, tune, or improve shared machine learning models or global AI systems that serve multiple customers or…

    How to test + evidence

    Testing procedure: Customer fine-tunes are tenant-isolated; no cross-customer leakage possible.

    Evidence to collect: Architecture + isolation tests.

  5. DSAR / erasure process for AI in scope

    High

    This control ensures that Data Subject Access Request (DSAR) and erasure processes explicitly account for personal data processed by or embedded within AI systems, including training datasets, model weights, embeddings, vector stores, and inference logs. When a data subject requests…

    How to test + evidence

    Testing procedure: Erasure workflow accounts for embeddings, indices and cached outputs.

    Evidence to collect: DSAR workflow doc.

  6. Synthetic / anonymised data preferred where viable

    Medium

    This control mandates the preferential use of synthetic, anonymised, or pseudonymised datasets in place of production data containing sensitive or personally identifiable information, particularly in non-production environments such as development, testing, staging, and analytics. Synthetic data is artificially generated to…

    How to test + evidence

    Testing procedure: Synthetic data used for testing and evaluation where it preserves utility.

    Evidence to collect: Pipeline evidence.

  7. Output filters for memorised / sensitive content

    High

    Output filters for memorized or sensitive content prevent systems—particularly AI models, chatbots, and automated response platforms—from inadvertently disclosing confidential data, training material, credentials, or personally identifiable information through generated responses. These filters employ pattern matching, redaction logic, keyword blocking, and…

    How to test + evidence

    Testing procedure: Guardrails block outputs containing PII / verbatim training data.

    Evidence to collect: Guardrail config + test results.