SafeScreen: A Safety-First Screening Framework for Personalized Video Retrieval for Vulnerable Users

A team of researchers has introduced SafeScreen, a novel framework designed to protect vulnerable users like children and dementia patients from harmful video content. Unlike traditional platforms like YouTube that prioritize engagement, SafeScreen treats safety as a non-negotiable prerequisite. The system works by creating a pipeline that sequentially approves or rejects candidate videos using three core components: extracting individualized safety criteria from user profiles, performing evidence-grounded assessments via adaptive question generation and multimodal VideoRAG (Retrieval-Augmented Generation) analysis, and finally employing LLM-based decision-making to verify safety, appropriateness, and relevance in real-time.

In a detailed case study focused on dementia-care reminiscence therapy, the researchers evaluated SafeScreen using 30 synthetic patient profiles and 90 test queries. The results were striking: SafeScreen's safety-first approach caused its recommendations to diverge from YouTube's engagement-optimized rankings in 80% to 93% of cases. The framework demonstrated high levels of safety coverage, sensibleness, and groundedness, as validated by both automated LLM evaluation and human domain experts. This proof-of-concept shows it's possible to screen uncurated video repositories without relying on precomputed safety labels, offering a more explainable and trustworthy alternative for sensitive applications.

The research, currently under review for ACM ICMI 2026, addresses a critical gap in how AI handles content for at-risk populations. By flipping the script from 'maximize watch time' to 'enforce safety first,' SafeScreen provides a blueprint for building responsible AI systems in healthcare, education, and caregiving. Its multimodal approach combining vision, language, and user profiling represents a significant technical advancement in personalized content safety.