Public works keeps raw PII on-premises while an AI analysis job receives only hashed cohort features. What technical exposure control is this?

Keeping raw PII on-premises while an AI analysis job receives only hashed cohort features is a technical exposure-minimization control: the AI pipeline never sees authoritative identifiers, only derived features needed for the analytical task. Public-works teams reduce blast radius if the AI environment is compromised and shrink unnecessary third-party processing of personal data. Separation of PII stores from model inputs is a practical privacy-engineering pattern for civic analytics.

Uploading the full PII warehouse to every notebook kernel for convenience fails because convenience maximizes exposure and undermines least privilege. Printing Social Security numbers on the shop-floor dashboard fails by concept and basic safeguarding—raw SSN display is not required for cohort feature analysis. Disabling encryption because hashed features are already public fails because hashes are not a license to drop transport/at-rest protections, and hashed features may still be sensitive in context.

Exam caveat: Simple hashes of identifiers can be reversed with dictionaries if the identifier space is predictable—prefer salted keyed pseudonyms and irreversible features where policy requires. Cohort features can still enable discrimination risk; ethics review remains relevant.

Operational check: Architect a feature store that exports only approved hashed/aggregated cohorts to the AI job, keep raw PII in a restricted on-prem system, verify encryption in transit and at rest, and audit notebooks for accidental raw joins. Test that the AI environment cannot query the PII vault directly.