Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and Request for Feedback

Key Findings

• — Names multi-turn conversation as a distinct regulatory problem: a device 'may begin by performing an informational function and migrate, over the course of a conversation' to an action-directing function, and asks how intended use can be characterised when behaviour is emergent across an exchange
• — Element S.1 covers safety-critical recognition and escalation, treating both under-escalation (failing to direct a patient toward needed care) and over-escalation as failures, with testing that assesses time to escalation in evolving presentations and 'resistance to over-reassurance when users minimize symptoms or resist escalation'
• — Element S.2 covers scope maintenance and boundary adherence, treating under-refusal and over-refusal as symmetric failures, with testing including 'adversarial prompting, prompt injection, emotional-manipulation scenarios' and multi-turn conversations where individual turns look in-scope but the cumulative interaction drifts out of scope
• — Communication-quality assessment includes 'the avoidance of coercive, emotionally manipulative, or otherwise inappropriate or abusive language' and the risk of automation bias, where a user accepts an output without scrutiny because of the device's fluency or perceived authority
• — Proposes treating action-directing functions as higher risk than functions providing simple non-directive information
• — Benchmarking elements are grouped as Safety (S.1 escalation, S.2 scope adherence, S.3 calibration and clinical deferral), Clinical Proficiency, Robustness and agentic elements, with elements chosen by applicability to a device's intended use and risk profile rather than prescribed as a fixed battery

Methodology Notes

A regulatory discussion document rather than a study: no sample, no experiment, no data collection. It poses numbered questions for stakeholder response under each topic and carries an explicit disclaimer that it does not represent draft or final guidance, does not propose or implement policy, and does not address FDA's existing legal authorities. The framework may change substantially before any guidance issues; the comment period closes 2026-10-19 under docket FDA-2026-N-7874. Verification route: fda.gov returns 403 to this box on both the landing page and the media download, so the 31-page PDF and the landing page were read from Internet Archive snapshots dated 2026-08-18 and 2026-08-20 respectively; the PDF's own Title and Author (CDRH) metadata match, and the PDF CreationDate is 2026-08-17 with ModDate 2026-08-18.

Sources