A public-health chatbot must avoid echoing Social Security numbers that residents sometimes paste into chat. Which privacy-preserving control best addresses that risk?

Applying Amazon Bedrock guardrails or equivalent privacy filters so sensitive identifiers are not returned in outputs is the privacy-preserving control that stops a public-health chatbot from echoing Social Security numbers residents may paste into chat. Guardrails configured for sensitive information filter or block SSN-like patterns in model responses, so even if an SSN reaches the context window, the assistant does not politely repeat it back into the resident transcript, screenshots, or downstream CRM notes. Public-health deployments typically pair output filters with input redaction for defense in depth.

Instructing residents to type SSNs in all caps so the model can highlight them increases visibility of the sensitive value instead of suppressing it—highlighting is the opposite of privacy preservation. Storing every pasted SSN in a public metrics dashboard for fraud analytics creates a centralized, high-value disclosure surface and violates minimization. Disabling all output filters and relying on residents to self-censor after the fact shifts responsibility onto stressed users and leaves echo risk unmitigated when pastes already occurred.

Exam caveat: distinguish input handling from output guardrails. Both matter, but this stem targets not returning identifiers in outputs—Bedrock Guardrails (sensitive-information filters) is the named AWS-aligned pattern. Do not confuse “make PII more visible for the model” with privacy controls.

Operational check: paste synthetic SSNs into a non-production public-health bot, verify responses redact or refuse to echo them, confirm guardrail/filter metrics register blocked sensitive content, and ensure support workflows never re-inject raw SSNs into outbound messages.