Google Research and Google DeepMind have introduced guardrailed-AMIE (g-AMIE), a diagnostic AI system designed to operate strictly under physician oversight. This new framework, detailed by researchers David Stutz and Natalie Harris, decouples patient history-taking from medical decision-making to ensure regulatory compliance and patient safety. Unlike systems that provide direct advice, g-AMIE generates comprehensive SOAP notes and draft patient messages for review by a licensed primary care physician via a 'clinician cockpit' interface.

In a randomized, blinded virtual OSCE study, g-AMIE's diagnostic performance and management plans were preferred by overseeing physicians and independent raters compared to nurse practitioners and physician assistants operating under the same guardrails. This research represents a significant step toward asynchronous human-AI collaboration in healthcare, prioritizing the established medical paradigm of clinician accountability while leveraging AI efficiency for information gathering. The study highlights the potential for AI to augment clinical workflows rather than replace them.