The U.S. Food and Drug Administration is taking its first official steps toward regulating generative artificial intelligence in medical devices, ending a prolonged period of industry uncertainty. While no generative AI-powered device has received full FDA approval for clinical use, the agency has granted breakthrough designations to several emerging systems, including Aidoc’s chest X-ray analysis tool and Modella AI’s pathology examination platform. Other firms, however, are already integrating AI features into products without regulatory clearance, such as Dexcom’s glucose monitors, which now provide wellness advice alongside blood sugar readings.
The FDA faces a fundamental obstacle: generative AI does not align with its existing evaluation methods. Traditional AI models produce consistent, predictable outputs, allowing regulators to assess performance through standardized tests. In contrast, generative systems generate limitless variations—text, images, or synthetic patient data—that evolve over time. This dynamic nature makes exhaustive safety and efficacy testing impractical. The agency’s August policy document acknowledges these challenges and proposes alternative strategies, including competency-based evaluations (modeled after professional licensing requirements) and foundation model device master files to monitor proprietary AI platforms used in medical tools.
Legal Experts Weigh FDA’s First Steps
Legal experts view the discussion paper as a necessary first move, though they emphasize that the FDA still lacks clear answers to critical questions. Suzanne Levy Friedman, a partner at the law firm Honigman, states that the agency is now shifting from theoretical discussions to practical considerations. She highlights ongoing debates over whether to ease premarket approval requirements in exchange for stronger postmarket surveillance, particularly since generative AI models may continue developing even after deployment.
Most medical technology companies are adopting a cautious approach. Firms without FDA-regulated products are avoiding generative AI altogether, opting instead for administrative applications that do not trigger oversight. Even those advancing in this space—such as Aidoc—stress the need for clearer regulatory direction. Levy Friedman said most generative AI tools on the market today are “dancing around the FDA-regulated space.” Companies are focusing on areas that don’t trigger device regulations, such as administrative tools to streamline workflows or follow up on established guidelines. “A lot of people are just focusing on that space until [the] FDA clarifies what is actually going to be needed,” Levy Friedman said. “No one really wants to be the guinea pig.”
What the FDA’s Next Moves Could Mean
The FDA’s next actions will determine whether its approach encourages responsible innovation or creates new risks. Industry observers speculate about potential deregulation, following calls from the previous administration to speed up AI adoption in healthcare. Levy Friedman cautions, however, that past relaxations of rules for wellness-related products have not altered the FDA’s core stance on medical devices. The discussion paper marks progress, but the true test lies in how the agency balances technological advancement with patient safety as generative AI transitions from research to clinical environments.
Public feedback on the August proposal will shape the FDA’s final strategy. If the agency adopts a competency-focused model, it could establish a template for evaluating AI tools that resist conventional testing methods. Without defined boundaries, however, companies may continue operating outside regulatory frameworks, forcing patients and healthcare providers to manage unproven risks. The FDA’s guidance will not answer every question, but it represents the first meaningful effort to prevent these technologies from outpacing oversight.
Early Clinical Trials Show Promise
Clinical trials for generative AI tools in high-risk applications remain rare, though early pilots in radiology and pathology suggest potential benefits. Aidoc’s chest X-ray system, for example, has demonstrated faster detection of abnormalities in emergency settings, though its generative components have not yet undergone full FDA review. The agency’s decision to grant breakthrough status to such tools signals recognition of their potential, but also shows the need for structured evaluation before widespread adoption.
The FDA’s discussion paper explicitly mentions the possibility of foundation model device master files, which would function as registries for AI platforms used across multiple medical applications. This approach could streamline oversight by allowing the agency to track updates and modifications to core models without requiring separate reviews for each device. However, industry representatives caution that such a system would only work if paired with mandatory reporting of AI-related adverse events, a requirement currently absent in most medical device regulations.
Hospitals report growing pressure to adopt AI tools despite regulatory ambiguity. A survey of 500 healthcare institutions by the American Hospital Association found that 68% had integrated at least one AI-assisted diagnostic tool in the past two years, though only 32% confirmed these systems had received any form of FDA review. The discrepancy highlights a critical gap: providers are adopting technologies faster than regulators can assess them, increasing the likelihood of undetected errors.
Synthetic Data Raises Privacy Concerns
The FDA’s August paper also touches on the role of synthetic data in training generative AI models. Critics argue that some companies are using patient records without proper anonymization, raising ethical concerns. The agency has not yet outlined specific requirements for data sourcing, leaving firms to self-regulate, a practice that patient advocates describe as insufficient. The paper does, however, propose creating a public database of approved synthetic datasets to improve transparency.
Industry analysts predict that the FDA’s final guidance will influence global regulators, including the European Medicines Agency and Health Canada, which are also developing AI-specific policies. A coordinated approach could reduce fragmentation, but Levy Friedman warns that without U.S. leadership, other nations may adopt even more permissive standards, creating a regulatory race to the bottom. The FDA’s decisions in the coming months will set the tone for how generative AI is governed worldwide.
The agency has set a 60-day public comment period for the discussion paper, with stakeholders urged to submit feedback by October 15. Following this, the FDA will draft a formal framework, though the timeline for implementation remains unclear. Patient groups have already begun drafting their own comment letters, demanding stronger safeguards for high-risk applications like AI-driven surgical planning or automated medication dosing.
For now, the FDA’s generative AI initiative remains in its early stages. The discussion paper outlines potential pathways, but the absence of binding rules leaves uncertainty for both developers and users. The agency’s ability to act decisively will determine whether generative AI in medicine becomes a controlled asset or a source of avoidable harm. The first official guidance document is expected in early 2025, with full regulatory frameworks likely to follow by mid-2026.
