5.3. Responsible AI for Multimodal Content
💡 First Principle: Visual content carries safety risks that text-only pipelines do not — unsafe imagery, prompt-injection text hidden inside an image, and brand/policy violations — and the platform provides controls aimed specifically at these. Responsible AI for vision is a named exam objective, not a generic afterthought.
Why care: the official outline lists three distinct multimodal responsible-AI tasks — classifying unsafe or disallowed visual content, detecting and mitigating indirect prompt injection via embedded text in images, and enforcing visual policy (watermarks, prohibited symbols, brand usage, inappropriate content). Each maps to a different control, and the exam tests whether you reach for the right one.
⚠️ Common Misconception: "Content safety is for text; images are just pixels." Images carry the same safety-category risks as text (the filters classify visual content too), and they introduce a vision-specific attack: an attacker embeds instructions as text inside an image, which a naive multimodal pipeline reads and obeys. Treating images as safe-by-default is the gap.
