The Food and Drug Administration has cleared 1,357 AI-based medical devices in the U.S. Only three have been tested to determine if they improve patient health, a recent study found.
AI medical tools often lack real-world health testing
A review of all FDA-cleared AI-enabled devices as of December 2025 revealed that 99.8% lacked evidence of reducing mortality, complications, or hospital readmissions. Nearly 80% of the 1,357 devices were designed for radiology, with the rest split between cardiology and neurology.
The agency’s approval process requires manufacturers to demonstrate only that a device is similar to an existing one, not that it works effectively in clinical practice. This method creates a chain of unproven tools, where each new product relies on the last without independent checks.
Related: Training fees mount with few clear answers
Study authors stated, “Regulatory clearance has moved faster than clinical validation, building a system where innovation progresses without accountability. Patients end up bearing the risks of unproven technologies.” The three devices tested used small sample groups and limited geographic representation.
Previous failures show the problem. The Epic Sepsis Model missed two-thirds of sepsis cases while generating false alerts for nearly 20% of hospitalized patients. Such issues reveal the difference between regulatory approval and actual performance.
Global impact of U.S. approvals
FDA decisions influence markets worldwide. Many countries have historically accepted U.S. clearance as sufficient for their own approvals, though some are now imposing stricter reviews. The findings suggest untested AI tools may spread globally with little examination of their real-world effects.
The team did not identify the three tested devices but described the studies as limited. Most tools in the dataset assist with diagnostics, such as analyzing X-rays or MRIs, rather than guiding treatment directly. Even diagnostic aids can shape medical decisions, and errors at this stage may lead to broader care problems.
Related: A Very Low Calorie Diet – Benefits
Radiology leads the AI medical device market, but the authors warned against assuming these tools are safer. A misread scan can delay treatment just as a faulty sepsis alert can mislead doctors. The rapid growth of these devices points to a rush to market that often skips thorough testing.
This trend mirrors past patterns in medical devices. Pacemakers and surgical robots have long followed a regulatory path that values similarity over outcomes. AI’s complexity makes flaws harder to detect before they affect patients. Unlike a faulty implant, which may fail predictably, AI errors can be subtle and hard to trace.
The authors did not demand a halt to AI medical devices but pushed for closing the evidence gap. One idea involves requiring post-market studies for high-risk tools, similar to some drug regulations. Another proposal called for transparency about training data, including its diversity.
