Research Summary: A Safety-First Gateway Architecture for Trusted Public Health Resource Navigation

A research initiative proposes a Safety-First Science Gateway architecture to enhance access to public health information, specifically for maternal and child health (MCH), using conversational AI. This system integrates large language models (LLMs) and retrieval-augmented generation (RAG) with a multi-layered safety framework. Key features include emergency handling, domain screening, source attribution, anonymous session management, and audit logging, all while restricting information retrieval to curated institutional resources.

Why it matters

This development is crucial for organizations looking to leverage advanced AI for public information dissemination, particularly in sensitive sectors like health. It highlights the critical need for integrating comprehensive safety protocols to ensure responsible and trustworthy AI deployment, mitigating risks of misinformation and inappropriate advice.

Key insights

• Conversational AI can significantly improve access to public health information.
• Public-facing healthcare AI requires robust safeguards against inappropriate medical guidance and unsupported information generation.
• The proposed architecture combines LLMs and RAG with a multi-layer safety framework to address these risks.
• Safety features include emergency handling, domain/scope screening, explicit source attribution, anonymous session management, and operational audit logging.
• Information retrieval is explicitly restricted to curated institutional resources to maintain accuracy and reliability.
• The research includes a prototype implementation and functional verification of selected workflows for this gateway.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2607.13038