Get Full Government Meeting Transcripts, Videos, & Alerts Forever!
Get email alerts on the Ai Healthcare topic
No spam. Unsubscribe anytime.
Harvard expert tells CalPERS: AI can transform primary care but must be governed to avoid harm
Summary
Dr. Eric Schneider told the CalPERS board that AI shows promise for chronic‑disease management, documentation relief and expanded access, but generative models can 'hallucinate,' entrench bias, erode clinician skills and raise privacy and liability issues.
Dr. Eric Schneider, adjunct professor at the Harvard T.H. Chan School of Public Health, told CalPERS trustees that artificial intelligence presents real opportunities for primary care — and real risks that require governance.
Schneider distinguished two broad AI flavors: predictive/machine‑learning models (risk scores, image analysis) and generative large language models (LLMs). He said predictive tools can be mature and useful when validated against the population where they are applied. Generative AI (ambient scribes, chatbots, summarization) is newer, offers big productivity gains but carries unique safety risks because it can produce fluent but incorrect outputs (hallucinations).
Promises Schneider highlighted: better chronic‑disease monitoring and coaching; substantial reductions in clinician documentation burden via ambient scribing (studies showing clinicians recovered about an hour of work time per week after tool deployment); improved access in underserved areas using decision support that extends limited clinician expertise; and scalable behavioral‑health screening and adjunct support.
Perils he emphasized: confident but sometimes wrong model outputs that can cause omission errors; algorithmic bias that can worsen disparities because models reflect gaps in training data; cognitive surrender and skill erosion if clinicians over‑rely on AI; privacy and security exposure from large training data sets; workload paradox where excessive alerts and false positives erode time savings; and unresolved liability allocation among clinicians, vendors and developers.
Schneider urged careful governance: testing tools in real‑world clinical workflows, co‑design with clinicians, transparency about training populations, and human‑in‑the‑loop designs that preserve clinician oversight. He said several pilots already exist (some health systems use limited chatbot triage with human triage fallback), but the field is fast moving and safety monitoring frameworks are still emerging.
Board discussion covered pilots, regulation, equity implications, and whether self‑regulation by vendors can play a useful role before stable federal rules are in place. Staff said CalPERS is starting to coordinate with state partners and purchaser peers on contractual guardrails and monitoring.
Quote: "The promise is real, but there is a lot of uncertainty," Schneider said. "An AI tool that speeds up visits but undermines coordination or continuity won't be an improvement."
