3.3.2. Responsible AI: Content Safety, Prompt Shields, and Groundedness
💡 First Principle: Responsible-AI controls map to distinct threat surfaces — harm-category filtering (the four categories at configurable severities) blocks harmful content, prompt shields block injection/jailbreak attempts, protected-material detection flags copyrighted output, groundedness detection catches ungrounded hallucinations, and PII detection flags sensitive data. Matching the threat in the scenario to the right control is the testable skill.
Content filtering screens four harm categories — hate, sexual, violence, self-harm — each with severity thresholds (Safe/Low/Medium/High) you tune per use case; content at or above the threshold is blocked or annotated. Filters are created at the resource/hub level as reusable configurations and associated with deployments (one active filter per deployment), screening both input and output. Beyond harm categories: Prompt Shields defend against direct jailbreaks and indirect prompt injection from documents; Protected Material detection flags copyrighted text/code; Groundedness detection identifies hallucinated, unsupported output; PII detection flags sensitive personal data in completions; and custom categories/blocklists let you block domain-specific content.
⚠️ Exam Trap: The four harm categories (hate, sexual, violence, self-harm) are not the same as prompt shields or groundedness. A scenario about users trying to jailbreak the system is Prompt Shields, not a harm-category threshold; a scenario about the model inventing facts not in the source is groundedness detection, not a safety severity setting. Reaching for harm-category sliders to solve an injection problem is the mismatch trap.
Reflection Question: Three scenarios — (a) users craft prompts to bypass safety rules, (b) the model cites a policy that doesn't exist in the provided docs, (c) the model emits violent content. Which responsible-AI control addresses each, and why aren't they interchangeable?
