5.3.2. Visual Policy: Watermarks, Symbols, and Brand Enforcement
💡 First Principle: Beyond classifying unsafe content, vision pipelines often must enforce organizational policy on what is produced or ingested — applying watermarks to generated media, flagging prohibited symbols, upholding brand-usage requirements, and detecting potentially inappropriate (but not strictly unsafe) content. This is a governance layer distinct from the harm-category safety filters.
The harm-category filters answer "is this content unsafe?" Visual policy enforcement answers a different question: "does this content comply with our rules?" Applying a watermark to AI-generated media supports provenance and disclosure requirements. Flagging prohibited symbols and enforcing brand-usage rules protect against off-policy or off-brand output. Detecting potentially inappropriate content catches material that is not a clear safety violation but still violates organizational standards. These are configured as policy checks layered onto the generation or ingestion pipeline, separate from — and complementary to — the content-safety harm filters. The skill is recognizing that a brand-or-policy requirement is not the same as an unsafe-content requirement and reaches for the policy layer.
⚠️ Exam Trap: "Ensure every AI-generated image carries a watermark and contains no competitor logos" is a visual policy requirement, not a harm-category-filter task — the safety filters classify unsafe content, they do not apply watermarks or enforce brand rules. Map a provenance/brand/policy requirement to the policy-enforcement layer.
Reflection Question: A marketing team must watermark every generated image and block any that contain a competitor's trademark. Why is this a visual-policy-enforcement task rather than something the unsafe-content harm filters handle?
