A randomized study published in Nature found Google's diagnostic AI system outperformed or matched doctors on nearly every metric evaluated.
Google’s medical AI system just passed a test that most people assumed was still years away. In a double-blind crossover study published in Nature in April 2025, the company’s Articulate Medical Intelligence Explorer, known as AMIE, matched or outperformed primary care physicians across nearly every diagnostic metric thrown at it.
What the study actually found #
The study involved 159 simulated case scenarios conducted across Canada, the UK, and India. Twenty primary care physicians participated in the randomized trial, going head-to-head with AMIE on complex diagnostic conversations.
Specialist doctors evaluated both the AI and human physicians across 32 performance metrics. AMIE outperformed or matched the PCPs on 30 of those 32 measures.
Patient-actors who interacted with both AMIE and the human doctors rated the AI system superior on 25 out of 26 metrics.
The research was conducted by teams at Google Research and Google DeepMind, building on foundational work that first introduced AMIE in a January 2024 blog post. That initial version relied on simulated dialogues and reasoning techniques to train the system.
Beyond text: AMIE goes multimodal #
A separate paper, also published in Nature, evaluated a multimodal extension of AMIE. This version can assess images, electrocardiograms (ECGs), and other clinical data types that physicians routinely use to make diagnoses.
Still, a critical caveat: all of this remains in the research phase. Feasibility studies are currently underway to evaluate how AMIE might perform in real clinical environments.
Why this matters beyond healthcare #
For investors tracking AI’s trajectory, AMIE’s results offer a concrete data point rather than a vague promise. Google isn’t just claiming its AI is smart. It’s publishing peer-reviewed evidence in one of the world’s most prestigious scientific journals showing that the system performs at physician-level competence across diverse geographic and clinical contexts. The risk side of the equation is worth watching too. Deploying AI in clinical settings introduces liability questions that don’t have clean answers yet. Who’s responsible when an AI misdiagnoses a patient? The hospital that deployed it? Google? The physician who relied on its recommendation? These aren’t hypothetical concerns anymore. They’re questions that need answers before AMIE or anything like it sees real patients at scale.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our