Speaker: David Talby, PhD
Affiliation: CEO, John Snow Labs and Pacific AI
Date: December 3, 2026

Join link on the Fred Hutch event page

More than 600 laws, regulations, standards, and frameworks now apply to organizations deploying artificial intelligence. David Talby, PhD, examines whether today’s medical AI benchmarks can show that clinical AI systems—and increasingly, AI agents—are safe, reliable, fair, and ready for real-world use.

Abstract

The number of laws, regulations, standards, and frameworks relevant to organizations deploying AI grew from more than 250 in March 2026 to more than 600 by September, including over 200 measures from U.S. state legislatures. Recent laws increasingly require specific system behaviors, not merely disclosure that AI is being used. A model may perform well on a medical knowledge benchmark while still failing to escalate a mental health crisis, route urgent symptoms appropriately, disclose that it is not a physician, or remain free of discrimination after a vendor update.

The first part of this talk presents nine questions that healthcare institutions should be able to answer about AI testing and monitoring before and after deployment. These include which legal requirements a benchmark evaluates, what score constitutes a deployment threshold, how performance varies across the patients an organization serves, and who reviewed the test cases. The talk examines gaps in medical safety benchmarks published since 2024 and introduces PacificMed, a suite of more than 70 benchmarks covering medical AI bias, fairness, safety, and reliability.

The second part turns to AI agents that do more than answer questions—for example, completing prior authorization tasks, coordinating a specialist consultation across many tool calls, or auditing a medical record against a standard. Results from three agentic benchmarks in the open-source MedHELM project illustrate the difficulty of evaluating these systems. Agents completed 83% of individual subtasks, yet only 36% of workflows succeeded end to end. The best of 13 agents completed 46% of long-horizon clinical tasks, and all 16 systems in another evaluation produced answers even when the patient record did not support one.

The talk concludes with practical steps organizations can take now to evaluate and monitor clinical AI systems and agents, including benchmark suites available free for academic research.

Speaker bio

David Talby is CEO of John Snow Labs and Pacific AI, where he helps organizations apply artificial intelligence to real-world problems in healthcare and life sciences.

He has extensive experience building and leading web-scale software platforms and teams across startups, open-source projects, Microsoft Bing in the United States and Europe, and Amazon’s financial systems in Seattle and the United Kingdom.

Dr. Talby holds a PhD in Computer Science as well as master’s degrees in Computer Science and Business Administration. His honors include USA CTO of the Year from the Global 100 Awards in 2022, the Game Changers Awards in 2023, and the ACQ5 Global Awards in 2025.

Rights and attribution

This recording and accompanying text are presented for educational access with attribution to the speaker. Speaker views are their own. No open license is implied unless explicitly stated.

This is not an official Fred Hutch page; all views expressed here belong to the individual expressing them.

Back to the Fred Hutch AI Club archive