Google DeepMind launches AI co-clinician research program with multimodal AI agent

AIMPACT message, May 1 (UTC+8): On April 30, 2026, DeepMind announced the launch of the AI Co-Clinician research initiative, exploring how multimodal AI agents can assist medical decision-making under physician supervision. The system enables real-time interaction via video and audio to collect patient histories, guide physical examinations, perform preliminary diagnostic reasoning (e.g., rotator cuff injury or tendonitis), provide care recommendations, and clearly disclose that it is not a physician—所有建议需在医生监督下使用. Technical features: Real-time multimodal processing (analyzing gait, respiration, rashes, etc.), dual-agent architecture (Planner continuously monitors Talker to enforce safety boundaries), and the NOHARM safety framework. Test results: Among 98 primary care queries, 97 achieved zero critical errors; in 20 simulated scenarios across 140 evaluation dimensions, 68 dimensions matched or exceeded physician performance (particularly in triage), while human physicians retained clear advantages in identifying critical “red flag” signs and guiding examinations. The project is collaborating with Harvard Medical School, Stanford School of Medicine, and others, and is currently in the research phase, with plans to gradually expand clinical testing. DeepMind emphasizes that the AI is designed to act as a “collaborative member” to augment, not replace, clinical judgment. (Source: BlockBeats)