A policy chatbot that indexes an entire HR drive may expose résumés, disciplinary notes, and ID scans when a few de-identified policy files would suffice. Data minimization limits personal data collection and use to what is necessary for a defined purpose. A lawful basis for processing is necessary, but it does not justify collecting everything available. section 5.2.2
Minimization matters early in a generative AI pipeline: personal data can be memorized in model weights or exposed through outputs, and removing its influence may require costly retraining. Prompts, outputs, and telemetry retained for debugging can also become unnecessary stores of personal data or sources for later training. section 5.2 section 5.2.2
Apply the boundary at each data decision:
- Define the task and select only relevant records or fields. For retrieval-augmented generation (RAG), route a curated set of documents into the index rather than entire repositories. section 5.2.2 section 5.2.4
- Exclude unnecessary personal data from training and fine-tuning. If it is included, document the purpose and lawful basis and plan for access, correction, and deletion. section 5.2.4
- Justify how long prompts, outputs, and telemetry need to be retained, and check retention settings during acceptance testing. section 5.2.2 section 1.5
- Prevent unauthorized secondary use by obtaining vendor commitments to opt out of model training or other uses. section 5.2.4
Controls depend on deployment. An API integration can use a central gateway to redact identifiers or secrets, block sensitive fields, and sanitize traffic before it reaches a vendor. SaaS applications offer less control; where supported, sanitize inputs before submission, or explore data loss prevention. Both vendor-hosted and local RAG need controls over what enters the index and a legal assessment for personal data. section 5.2.4
Anonymization is not a guarantee of safety: indirect identifiers and combinations of attributes can identify people even after names are removed. Keeping information in external lookup tables rather than model weights can make it easier to manage, but retrieval documents and embeddings still require protection. Documents and embeddings can usually be updated or deleted without retraining; data incorporated into weights is harder to remove. section 5.2.2 section 5.2.4 section 5.7




