Doximity Ask ranks first in Stanford-Harvard clinical AI safety study

Doximity Inc.'s HIPAA-compliant clinical AI assistant outperformed frontier models from OpenAI and OpenEvidence in an independent safety evaluation conducted by physicians from Stanford and Harvard Medical Schools, marking a shift in how healthcare AI is measured.

The NOHARM (Numerous Options Harm Assessment for Risk in Medicine) benchmark, developed by more than 50 researchers with contributions from 29 board-certified physicians, tested AI systems on more than 1,100 simulated patient scenarios spanning 10 medical specialties. Doximity Ask ranked first among all evaluated systems on the study's real-world clinical sample — the portion designed to mirror how physicians actually use these tools in practice.

"Continuous physician review isn't a differentiator. It's a requirement," said Dr. Louis-Antoine Mullie, Head of Medical AI at Doximity. "This result reinforces the importance of combining advanced AI systems with rigorous clinical oversight and independent safety evaluations like NOHARM."

Across the broader automated evaluation, purpose-built clinical AI systems outperformed general-purpose frontier models by a wide margin. Doximity Ask's performance traces to its PeerCheck program, through which more than 11,000 cited physician experts have evaluated and improved the platform's outputs. The company's approach — embedding physician review into the AI training loop rather than treating it as a post-hoc step — produced higher F1 scores that balance precision and recall, according to the study's methodology.

Why Physician Oversight Matters for Clinical AI

The NOHARM results highlight a structural advantage for domain-specific AI over general-purpose models in healthcare settings. While frontier models from OpenAI and others excel at broad knowledge tasks, clinical decision support requires handling nuanced trade-offs where incorrect answers carry real patient risk. The benchmark specifically measured harm potential — not just accuracy — making it a more relevant gauge for hospital deployment decisions.

Doximity's Clinical AI Suite, including Ask, has been reviewed and deployed across more than 150 health systems, including eight of the nation's top 20 hospitals. The platform includes end-to-end encryption, role-based access controls, audit logging, and session isolation — security features that healthcare organizations require before deploying AI in clinical workflows.

The company's network already includes more than 85% of U.S. physicians across all specialties, giving it a distribution advantage that pure AI vendors lack. Doximity was founded in 2010 and has built its platform around verified clinical membership, providing tools for collaboration, medical news, career management, and virtual patient visits.

What This Means for the Clinical AI Market

The independent study from Stanford and Harvard positions Doximity Ask as a benchmark for clinical AI safety at a time when hospitals are increasingly cautious about deploying generative AI. OpenEvidence, a direct competitor in the clinical AI space, was among the models Doximity outperformed, as were general-purpose frontier models that lack domain-specific training.

For investors, the study creates a measurable differentiation point. Doximity (NYSE: DOCS) now has third-party academic confirmation that its AI platform produces safer clinical outputs than both specialized competitors and general-purpose frontier models. The company's existing footprint across 150 health systems provides a distribution channel that could accelerate adoption as safety benchmarks become standard procurement criteria for hospital AI purchasing decisions.

The study was conducted by ARISE, a clinical AI research team led by physicians from Stanford and Harvard Medical Schools. The full methodology and results are available on the arXiv preprint server.

This article is for informational purposes only and does not constitute investment advice.