We're sharing our framework for building safer and more helpful AI for health.
Google published a framework to evaluate AI tools for health, developed with experts, assessing accuracy, safety, and usefulness to promote trustworthy AI in healthcare.
Artificial intelligence is increasingly used in healthcare, raising questions about the safety and reliability of AI-driven health tools. To address these concerns, Google collaborated with health experts Dr. Elizabeth McGlynn and Dr. Dana Gelb Safran to create a structured evaluation approach. The framework is designed to assess AI models across three critical dimensions: accuracy in delivering correct information, safety in minimizing risks, and usefulness in meeting real health needs. This method combines automated assessments, expert reviews, and human oversight to evaluate both routine and high-risk health scenarios.
The framework was developed to guide Google’s own AI model assessments, ensuring responses are medically sound and safe for users. It incorporates closed-question benchmarks, automated rating systems, and expert human evaluations to test AI performance. The goal is to identify potential weaknesses and areas for improvement before deployment in healthcare settings. Google states it continues to use this framework internally to refine its AI tools for health applications.
Google is making the framework publicly available to foster transparency and collaboration within the health ecosystem. The methodology is intended to help developers, researchers, and clinicians build AI tools that users can trust. By sharing the approach, Google aims to set a standard for evaluating AI in health contexts, encouraging broader adoption of safe and effective practices. The move reflects a commitment to responsible AI development in a field where accuracy is critical.
The whitepaper outlines the framework’s structure and its application to real-world health scenarios, including both common and high-risk cases. It emphasizes the importance of rigorous testing to ensure AI tools do not mislead or harm users. Google’s initiative seeks to align with health experts’ priorities, promoting tools that are not only advanced but also reliable and patient-centered. The framework is positioned as a resource for the entire healthcare community to adopt and adapt.