311 operations notice abusive and unsafe messages hitting the city’s public chatbot. Which operational practice should they emphasize?

311 operations should emphasize reviewing safety-event dashboards and alerts for the public-facing municipal chatbot so abusive and unsafe messages are detected, triaged, and used to tune filters. The practice works because public civic agents attract jailbreaks, harassment, and policy-violating prompts; safety telemetry (blocked categories, severity, volume spikes, false-positive rates) gives operators evidence to respond without waiting for social-media screenshots. Continuous review turns content safety from a one-time checkbox into an operating control for the city’s public AI channel.

Auditing groundedness scores and source citations for the public-facing municipal agent fails because it measures factual accuracy and hallucination, not abuse; strong citations do not detect, triage, or block harassing or jailbreak prompts hitting the live 311 bot. Monitoring token consumption and cost analytics for the public-facing municipal agent fails because it tracks spend and usage volume, not safety; cost dashboards do not surface blocked categories, severity, or attack patterns needed to respond to abusive messages. Tracking retrieval relevance and search recall for the public-facing municipal agent fails because it measures search quality and answer completeness, not content safety; high recall does not detect or mitigate harassment, jailbreaks, or policy-violating prompts.

Exam caveat: dashboards without on-call ownership or playbooks create alert fatigue; also watch for over-blocking that silences legitimate service requests after a spam wave. Operational check: wire safety events to a monitored channel, define severity thresholds and response owners, sample blocked and allowed transcripts weekly, and document filter changes with before/after metrics for the 311 AI ops cadence.