Google DeepMind announces AI co-clinician research to augment physician care

Google DeepMind has announced a new research initiative focused on developing "AI co-clinician"—systems designed to work as collaborative members of care teams alongside physicians, rather than replacing them.

The project responds to a critical healthcare challenge: the World Health Organization projects a shortfall of more than 10 million health workers by 2030. DeepMind's hypothesis is that the next phase of healthcare delivery will involve "triadic care," where AI agents support patients under the clinical authority of their physician.

Building trustworthy clinical AI#

The research emphasizes safety and evidence quality. DeepMind adapted the "NOHARM" framework to test for both errors of commission (incorrect information) and errors of omission (missed critical data). In blind evaluations against leading evidence synthesis tools, physicians consistently preferred AI co-clinician's responses.

On a benchmark of 98 realistic primary care queries, the system recorded zero critical errors in 97 cases, outperforming two widely-used AI systems. The team also evaluated the system on OpenFDA's RxQA medication knowledge benchmark, finding significant progress in handling complex drug-related reasoning.

Real-time multimodal capabilities#

Beyond text-based clinician support, DeepMind tested AI co-clinician in simulated telemedical settings with real-time video and audio. Working with researchers at Harvard and Stanford, they conducted a randomized study with 20 synthetic scenarios and 10 physician "patient-actors."

The system demonstrated capabilities like guiding patients through physical examinations in real time and correcting inhaler technique. However, expert physicians still outperformed the AI across 140 assessed consultation skills—particularly in identifying "red flags" and guiding critical examinations. AI co-clinician matched or exceeded primary care physicians in 68 of 140 areas, suggesting its current best use is as a supportive tool rather than a clinical replacement.

Safety architecture and next steps#

To ensure clinical-grade safety, AI co-clinician uses a dual-agent architecture: a "Planner" module continuously monitors the "Talker" agent to verify it stays within safe clinical boundaries.

DeepMind is advancing phased evaluations with academic and research collaborators across the US, India, Australia, New Zealand, Singapore, and the UAE. The company notes these research collaborations are not yet intended for diagnosis, treatment, or medical advice.