DeepMind's AI co-clinician clears 97 of 98 primary-care queries
Original: Enabling a new model for healthcare with AI co-clinician View original →
The headline in medical AI is usually about models replacing clinicians. Google DeepMind's new AI co-clinician work points in a different direction: AI as a tightly supervised teammate that can handle evidence lookup, medication reasoning, and parts of a telemedicine conversation without pretending to be the doctor. That framing matters because the World Health Organization still expects a shortage of more than 10 million health workers by 2030, and health systems need leverage before they need autonomy.
DeepMind's strongest result came on clinician-facing tasks. In blind evaluations across 98 realistic primary-care queries, physicians consistently preferred the system to leading evidence-synthesis tools. The company says the model made zero critical errors in 97 of those 98 cases, and it also showed strong gains on difficult medication questions from the OpenFDA RxQA benchmark. For hospital groups and primary-care networks, that is the sort of result that makes ambient research assistance look more real and less like demo theater.
The more interesting constraint is what the system still cannot do. DeepMind also tested the model in 20 synthetic telemedical scenarios with 10 physician patient-actors, using live audio and video capabilities built on Gemini and Project Astra. Physicians outperformed the AI overall, especially when it came to spotting red flags and directing critical physical exams. Even so, the model matched or exceeded primary-care physicians in 68 of 140 assessed areas, and DeepMind says it could guide tasks such as correcting inhaler technique or walking through shoulder maneuvers in real time.
That mix of progress and restraint is what makes the release worth watching. DeepMind is explicit that this is research, not a diagnosis or treatment product, and says future evaluation will happen with collaborators across the U.S., India, Australia, New Zealand, Singapore, and the UAE. If the next phase holds up outside simulation, AI in medicine may advance less like a chatbot replacing visits and more like a clinical copilot quietly reducing the time doctors lose to search, recall, and routine follow-up.
Related Articles
Google says joint research with Imperial College London and the UK’s NHS found that an experimental AI system identified 25% of interval cancers missed by conventional screening. The studies also suggest AI could reduce screening workload, while highlighting trust and calibration challenges in real clinical workflows.
An AI system that assesses burn-wound healing potential in roughly 20 to 25 seconds has cleared a key US regulatory gate. Spectral AI can now begin commercial distribution of DeepView for burn centers, trauma centers, and emergency departments.
Google-backed UC San Diego researchers plan to build a low-carbon cloud platform from 2,000 retired Pixel phones. The design strips devices to motherboards, groups 25-50 phones into Kubernetes-managed clusters, and targets teaching, grading, and research workloads.