The paper's risk framework scores a generative AI function on two axes: what the tool does, ranging from providing non-directive information to taking action to acting fully autonomously, and the consequence of relying on an incorrect output. Test rigor would scale with that risk profile. The comment period closed shortly before the panel, and the agency has received more than 300 responses. Abramson did not say when proposed regulations will be issued, only that the agency needs time to review the comments. The AMA sent Abramson a letter on October 16 urging a risk-based approach that includes physician expertise in performance standards and post-deployment monitoring. The paper has not yet produced final rules, and it remains unclear how physician-style testing would work in practice.
The paper builds on earlier FDA work, including a November 2024 Digital Health Advisory Committee meeting on lifecycle considerations for generative AI devices, which the paper cites as its origin. FDA has not authorized any generative AI-based device for a clinical purpose. A November 2025 committee meeting flagged hallucinations, model drift, and missed clinical deterioration as risks in digital mental health devices. Attorneys advising device makers should track the comment docket and any proposed rule, since the test-rigor and risk-tier structure could shape premarket evidence requirements. Companies deploying generative AI tools now should also document how their products handle incorrect outputs, because those consequences drive the framework's risk scoring.