Question

Your team deployed a generative AI customer-support chatbot built on a foundation model served via a Vertex AI endpoint. During pilot testing the model mostly behaves well, but occasionally produces toxic, biased, or otherwise harmful responses to certain user phrasings. Leadership requires an ongoing responsible-AI control that catches harmful generations before they reach customers and produces auditable metrics on unsafe output rates over time. Which approach best meets these requirements?

Accepted Answer

Add a runtime safety filter (e.g., content-moderation/toxicity classification on each response) that blocks or rewrites unsafe outputs, and log filter decisions to compute unsafe-rate metrics and dashboards.

Answer

Rely on periodic offline evaluation of a static prompt test set before each model version release, and skip runtime checks since the base model is already safety-tuned.

Answer

Enable feature attribution drift monitoring on the endpoint's input embeddings to detect when prompts change distribution.

Answer

Configure endpoint autoscaling and p99 latency alerts so harmful responses are throttled under load.