OpenAI and Anthropic donate 2,000 licenses to let health agencies test AI

Ten public health departments across the U.S. are getting their hands on enterprise AI tools from OpenAI and Anthropic, thanks to a new program that starts trials in fall 2026. The licenses, enough for 2,000 practitioners, are a donation. The program, called PULSE, is being orchestrated by the Coalition for Health AI (CHAI) with Accenture handling onboarding and the eventual creation of instructional playbooks. The goal is to figure out if generative AI can actually help with core public health work.

The five use cases on the table are substantive: biosurveillance and predicting drug waves, mapping social determinants of health, analyzing community feedback, building a multilingual translation hub for public communications, and a fuzzy-sounding one involving automated clinical data retrieval and FHIR query engines. That last point is where things get especially murky. The announcement doesn’t specify if an AI model will be writing database queries, pulling records, summarizing them, or some combination of all three. For work that involves clinical data, that’s not a minor detail to leave out.

And the governance questions don’t stop there. CHAI hasn’t published its evaluation criteria, security requirements, or rules for human review. We don’t know if a human has to sign off on a translated press release or a biosurveillance alert before it’s used. We don’t know what kind of data—identifiable patient records, de-identified sets, or synthetic data—will be fed into these models. Dr. David Lakey, a former Texas health commissioner, said the program’s success hinges on “trust, governance and execution.” Right now, the governance part is just an outline. HIPAA will apply to some pilots and not others, a compliance patchwork that adds complexity.

The vendors’ default policies say they won’t train on business data, but the specific configurations for these pilots haven’t been set. Accenture is supposed to compile the lessons learned into playbooks by 2027, giving other agencies a template. But if the evaluation framework isn’t built first, those playbooks will be more like travelogues than instruction manuals. They’ll tell a story of what happened, but not necessarily whether it was safe, accurate, or a good idea to repeat.