OpenAI: Astra Receives Safety Rating in 99% of Evaluated Medical Inbox Tasks
On November 17, according to The Defiant, OpenAI stated that in an evaluation of an AI agent designed to operate electronic health records software, 99% of the 325 medical intake tasks completed by GPT-6 Astra were rated clinically safe.
Evaluators examined the model's operations, as well as the medical records, orders, and patient messages it generated, to assess whether errors, omissions, or actions could cause patient harm. Karan Singhal of OpenAI stated that in cases that did not receive a safety rating, the model sometimes took action without appropriate approval or authorization.
The report notes that this figure represents the company's evaluation of task execution, rather than actual patient outcomes or independent clinical validation. OpenAI stated that these medical operational capabilities are accessible through APIs, ChatGPT for Healthcare, and ChatGPT for Clinicians.