The U.S. Food and Drug Administration is experimenting with a radically different approach to regulating generative artificial intelligence in medicine, allowing selected AI-powered medical devices to reach patients before receiving traditional marketing authorization. The initiative could provide regulators with valuable real-world experience while accelerating access to rapidly evolving technology, but it also raises important questions about patient safety and oversight.
The initiative is called TEMPO, short for Technology-Enabled Meaningful Patient Outcomes. Through the pilot, the FDA can exercise “enforcement discretion,” meaning it may temporarily choose not to enforce certain requirements—including premarket authorization—while participating companies deploy their technologies in controlled healthcare settings.
This does not mean the products have received FDA approval or clearance. The agency explicitly states that it has not yet evaluated the effectiveness of participating devices for the uses being tested. Instead, companies are expected eventually to seek normal marketing authorization, using real-world information collected during the pilot to help support their applications.
The experiment is particularly important for generative AI.
Traditional medical-device regulation was largely built around products that remain relatively stable after reaching the market. Generative AI can behave differently. Models may produce complex outputs, interact dynamically with clinicians and patients, and potentially evolve much faster than conventional medical technologies.
That makes evaluating them through traditional regulatory methods increasingly challenging.
According to STAT, products from Cadence and Limbic are among four devices recently accepted into TEMPO, giving FDA officials an opportunity to observe how emerging AI systems perform in actual clinical environments rather than relying exclusively on evidence generated before deployment.
The pilot is connected to the Centers for Medicare and Medicaid Services Innovation Center’s ACCESS model, which focuses on technology-supported treatment of chronic conditions. Eligible TEMPO devices must be intended for clinician-supervised outpatient treatment and address specific areas including cardiovascular, kidney and metabolic conditions, musculoskeletal disorders and behavioral health.
Patient risk remains an important limitation.
FDA says devices accepted into TEMPO cannot present the potential for serious risk to patients’ health, safety or welfare. Regulators also consider whether a product has a reasonable expectation of providing clinical benefit, how companies plan to monitor patients and whether meaningful real-world data can be collected.
The program could offer an important advantage over the conventional regulatory system: FDA officials can learn alongside developers while products are actually being used.
Generative AI poses regulatory questions that are difficult to answer in theory. Regulators need to understand how clinicians interact with AI recommendations, when users trust or ignore them, how systems behave across different populations and whether unexpected errors appear after widespread deployment.
Real-world monitoring could provide answers much faster.
For technology companies, TEMPO also offers an opportunity to bring products into clinical environments without waiting for the entire traditional authorization process to conclude. That could encourage innovation in a healthcare sector where AI technology is developing considerably faster than regulatory frameworks.
But the approach carries obvious risks.
Allowing medical devices onto the market before authorization effectively shifts some evaluation from the laboratory and regulatory-review stage into the real world. If an AI system provides inaccurate recommendations, behaves unpredictably or performs differently across patient populations, those problems could emerge while actual patients are using it.
FDA is attempting to manage that trade-off through monitoring, data collection and restrictions on which technologies can participate. Manufacturers generally need quality-management systems and must establish plans for collecting and reporting real-world evidence.
Participation is also limited. FDA currently expects to select up to approximately 10 U.S.-based manufacturers in each of four clinical areas, rather than opening the pathway broadly to the entire digital-health industry.
TEMPO therefore represents more than another regulatory pilot. It could become an early blueprint for how the government handles medical technologies that change faster than traditional regulation can comfortably accommodate.
The central challenge is balancing two competing risks. Moving too slowly could prevent patients and physicians from benefiting from useful AI innovations. Moving too quickly could expose patients to systems whose limitations regulators do not yet fully understand.
For the FDA, TEMPO is an attempt to find a middle ground: let carefully selected AI medical technologies enter real clinical settings earlier, monitor what actually happens and use that evidence to build a regulatory system better suited to the age of generative AI.
If successful, the experiment could fundamentally change how advanced medical software moves from development into American healthcare—and establish a model for regulating an increasingly powerful generation of AI systems without forcing innovation to wait for regulation to catch up.
