Ant Group presents MedGuard to verify medical facts in AI diagnostics
jiqizhixin · x · 2026-08-29
Ant Group and Xiamen University present MedGuard, accepted at npj Digital Medicine (IF 18.0). Unlike typical AI consultation systems, MedGuard is a safety infrastructure layer embedded into clinical workflows. Built on a 7B LLM, it intercepts doctor diagnoses, AI responses, and medication plans before finalization, extracting declarative statements and verifying key medical claims against clinical guidelines and evidence chains to flag risks.
Results: Achieves SOTA on fine-grained dialogue-level risk detection, improving F1 score by 22.1% over baselines and statement extraction by 23.2%. Average F1 reaches 0.7 across 41 medical subcategories.
More from Safety
- Jan Kulveiter: AI Models Should Have a Direct Line to Developers — jankulveit · 2026-08-29
- Deep Dive into AI Defense Dilemma: Evaluating Against Non-Stationary Model Adversaries — ziv_ravid · 2026-08-29
- Critique of Current Alignment Research: Models Easily Bypass Safeguards, RL Breeds Cheating — voooooogel · 2026-08-29
- OpenMined's work makes 'Glass-Steagall for AI' framework buildable — iamtrask · 2026-08-29
- Richard McNgo criticizes EA for turning AI safety into a 'fake field' — voooooogel · 2026-08-29
- MIT Media Lab's AI Observatory reveals gap between platform reports and actual user behavior — ShayneRedford · 2026-08-29