Mental model
Data minimization is an architectural constraint: reduce collection, context, telemetry, training use, and retention at every boundary.
Learning outcomes
- Explain the mechanism and ownership boundaries behind AI Data Classification and Minimization.
- Compare the main design alternatives and their operational trade-offs.
- Diagnose common failures and select evidence for a production decision.
Theory
Create allowed-data rules for prompts, retrieval, tools, traces, caches, evaluations, and exports. Redact or tokenize identifiers before provider calls where possible and document unavoidable sensitive fields.
Trade-offs
More context can improve quality but expands privacy, breach, and retention exposure. Aggressive redaction reduces risk but can remove information necessary for correct decisions.
Failure modes and misconceptions
Classifying only databases; logging full prompts by default; copying secrets into evaluation sets; indefinite caches; assuming provider settings cover application logs; and collecting data without a purpose.
Decision scenario
A healthcare assistant needs age and medication history but not full identity. Design a data flow that preserves utility while minimizing exposure.
Why must minimization be applied independently to prompts, traces, caches, and evaluation datasets?
Primary sources
nist-privacy-frameworkopenai-data-controls