A city must stop unsafe or disallowed images from entering the 311 photo-upload path. Which responsible-AI control should they implement?

Responsible AI for multimodal civic applications includes filters that detect and block unsafe or disallowed visual content before it enters production systems. A city securing the 311 photo-upload path should apply those visual safety filters so resident submissions that contain prohibited or harmful imagery are stopped at intake rather than stored, routed, or shown to agents unnecessarily. Filtering is a preventive control aligned with multimodal responsible-AI practice.

Using optical character recognition to detect and block unsafe visual content in resident uploads does not detect or block unsafe or disallowed visual imagery at intake because OCR only extracts text, so prohibited images without readable text still enter and the control requirement fails. Applying image captioning to detect and block unsafe or disallowed visual content in resident uploads does not implement the automated filtering control the stem asks for because captioning only describes content and does not block unsafe or disallowed imagery before it reaches the system. Using face detection to detect and block unsafe or disallowed visual content in resident uploads fails because detecting faces only locates faces and does not check for unsafe or disallowed imagery, so prohibited content still reaches the system unblocked.

Exam caveat: filters reduce risk; they can also false-positive benign images (for example, medical or protest contexts depending on policy). Cities need appeal or manual-review paths so legitimate service requests are not silently dropped. Operational check: enable unsafe-content filters on 311 photo intake, log blocks with reason codes, sample false positives weekly, and keep a supervised override for agents when a blocked image is required for emergency response.