What it asks for. The organisation must develop and implement policies and procedures for using personally identifiable information in internal testing, training and research, limit or minimise the amount used, authorise its use where it cannot be avoided and review and update the policies periodically.
In practice. This means using synthetic, masked or anonymised data in test and development environments, with documented approval and extra safeguards when real personal data is genuinely needed. Auditors look for the policy, data masking tools or processes and approval records for exceptions. This aligns with GDPR data minimisation and the expectations for training data in AI projects.
Common gaps. Production database copies in test environments with weaker controls are a long-standing problem. Data used to train AI models is often not assessed for personal data.