FDA Cleared 1,357 AI Devices. Only 3 Proven.

A sepsis alert fires on your monitor. The badge next to the tool's name says FDA cleared. You act on it, because clearance sounds like proof. It usually isn't.

The number behind the clearance

A study published August 19, 2026 in PLOS Digital Health, led by Rawan Abulibdeh at the University of Toronto, reviewed every FDA authorized AI and machine learning medical device on record as of December 5, 2025. The count: 1,357 devices.

Of those, 34 were linked to a registered clinical trial. Twelve had posted results. Twelve had a peer reviewed publication. Only three, 0.2 percent of the total, had been tested against a patient centered outcome: death, stroke, hospitalization, quality of life.

Radiology accounts for 78 percent of all cleared devices, 1,059 tools, and fewer than 1 percent of those have a prospective trial behind them. Among the 34 registered trials that do exist, 94 percent were industry led and nearly three quarters enrolled fewer than 500 participants, often from a single, well resourced health system.

What "cleared" actually means

Most of these devices went through the FDA's 510(k) pathway, which asks a narrower question than doctors assume it does. It asks whether a new device is substantially equivalent to one already on the market, not whether it makes patients better off. Equivalence is a paperwork bar. Outcome improvement is a clinical one. The study is a reminder that clearance answers the first question and stays silent on the second.

The researchers also flagged who gets left out of the studies that do exist: pregnant patients, adults over 75, non-English speakers. The subgroup most likely to be in front of you on a busy ward is often the one least represented in the evidence behind the tool.

Two cautionary tales, already public

The paper situates its findings against tools whose real-world record is already known. The Epic Sepsis Model, deployed across more than half of U.S. hospitals at its peak, missed 67 percent of sepsis cases in an external validation while firing alerts on 18 percent of all hospitalized patients, a flood of false positives dressed up as vigilance. IBM Watson for Oncology, once marketed as a second opinion for cancer treatment, matched local tumor board decisions only 33 percent of the time in a Danish hospital and 12 percent of the time for gastric cancer in China.

Neither tool was fraudulent. Both were cleared or adopted on the strength of technical performance, not outcome data. The gap the study measures is exactly the gap that let both slip through.

Why this matters on rounds

None of this means AI has no place on the ward. It means the badge on the box tells you less than it looks like it tells you. A tool can be cleared, deployed in thousands of hospitals, and still have never been checked against whether patients did better because of it. The doctor at the bedside is the one left holding that uncertainty, usually without knowing it's there.

That is the case for treating "verifiable" as a design requirement, not a marketing word. A clinical AI tool should show its sources, flag what it does not know, and make it easy for the physician reviewing its output to check the underlying reasoning against the chart, the same way you would check a colleague's differential. Confidence without a trail to verify it is the exact failure mode this study documents at scale.

The fix is not fewer AI tools on the ward. It is fewer tools that ask for trust they have not earned, and more that make it cheap to check their work before you act on it.

Sources

• Abulibdeh R, et al. "Clinical evidence behind FDA-cleared artificial intelligence and machine learning medical devices." PLOS Digital Health, August 19, 2026.
• Most AI medical devices cleared for use were not tested on patient outcomes, Medical Xpress, August 19, 2026.