SuperML
← All issues

FDA Wants to Credential AI Like a Doctor. The Infrastructure Doesn't Exist.

FDA's competency-based framework for generative AI medical devices closes for public comment on October 19 — but the governance infrastructure to actually run it doesn't exist yet, and most manufacturers haven't mapped their products to the two-axis risk framework.

Hi there,

FDA published a discussion paper in August proposing to regulate generative AI medical devices like physician credentialing. The concept is sound. The infrastructure to run it doesn't exist. And the comment window closes October 19 — seven days from now. If your team builds or deploys clinical AI, this is the framework that will govern you for the next decade, and right now you can still shape it.

🔥 Featured Post

FDA Wants to Credential AI Like a Doctor. The Infrastructure Doesn't Exist.

  • FDA's August 2026 discussion paper introduces a two-axis risk framework and ten-element competency benchmark for generative AI devices — but none of this is binding yet
  • The "shared ecosystem responsibility" model distributes obligations across manufacturers, health systems, clinicians, and payors without defining who enforces what
  • Foundation Model Master Files are voluntary, creating no incentive for model developers to disclose training data, architecture, or update schedules
  • Agentic clinical AI gets exactly one benchmark element (A.1), which is wildly underspecified for multi-step autonomous systems
  • Health systems will inherit monitoring obligations they haven't budgeted for and can't fully scope until FDA finalizes the framework

Read the full post →

📚 In Case You Missed It

Step-Level Agent Monitoring: Halting Bad Runs at 1ms — A new arXiv paper on streaming trajectory monitoring shows agent runs can be scored and halted step by step at roughly 1 ms overhead, but only if you have the reference data to compare against. Here is how to design the monitoring layer.

The EU AI Act High-Risk Deferral Is a Trap, Not a Reprieve — The EU AI Act Digital Omnibus deferred the Annex III high-risk deadline to December 2027 — but that's exactly one conformity assessment cycle, and Article 50 transparency obligations are already live. Banks that filed this as a win are already behind.

Compound AI Failures Live at the Seams, Not in the Model — A 150-incident study of production compound AI systems finds 51% of failures are silent, take 4.2 days to detect, and originate at component boundaries. Circuit breakers, quality gates, and typed interfaces cut recovery time 71%.

More posts dropping every week. Stay sharp.

— Bhanu @ superml.dev