AI

1,357 AI medical devices are cleared and just 3 were tested on actual patients

Key takeaways

  • A University of Toronto and MIT audit found only 3 of 1,357 FDA cleared AI medical devices had registered trials measuring patient outcomes.
  • 1,059 of the devices, about 78 percent, are radiology tools.
  • Clearance shows a device meets regulatory requirements; it is not evidence that patients do better with it.

Last updated: 5 October 2026

Three out of 1,357. That is how many FDA cleared AI medical devices had registered trials measuring what happened to patients, according to an audit from the University of Toronto and MIT published in PLOS Digital Health.

It works out to about 0.2%, or 3 divided by 1,357. The other devices were cleared, which is a different thing from being shown to help. The audit looked only at registered trials, so it counts studies that were formally set up in advance to measure patient outcomes.

What did the audit find?

Of the 1,357 cleared devices, 1,059 are radiology tools, roughly 78%. These systems look at scans and flag findings, so most were likely cleared on the strength of how accurately they read images against a reference.

That leaves about 298 devices outside radiology, covering everything else that regulators have let onto the market. For all of them, the audit found just 3 with a registered trial that measured patient outcomes.

Accuracy on images is a useful measure, but it is not the same as a better outcome. A tool can spot a shadow reliably and still change nothing about whether a patient is treated sooner, recovers faster or avoids harm.

The numbers at a glance

MeasureFigureWhat it means
FDA cleared AI devices audited1,357The full set examined by the University of Toronto and MIT team
Devices with registered outcome trials3The only ones with a formal study of what happens to patients
Share with outcome trialsAbout 0.2%Roughly one device in every 450
Radiology devices1,059Tools that read scans and flag findings
Radiology share of the totalAbout 78%Most of the market is image analysis

What is the difference between clearance and proof of benefit?

Clearance is a regulatory decision about whether a device meets the bar to be sold. In plain terms, it asks whether the device is safe and similar enough to something already on the market, not whether patients do better with it.

That is a sensible test for a product regulators need to move quickly. It was built for devices whose effects are well understood, such as a new design of a blood pressure cuff, and it leans on comparison with an earlier product. It was not designed as a verdict on whether a piece of software improves care.

A registered trial asks a harder question. It states in advance what will be measured, whether that is time to treatment, recovery or complications, and then reports the result whichever way it falls. Registration also makes it possible for others to see which studies were planned, so a missing result cannot quietly vanish.

Why does clearance get misread?

The risk is that hospitals, buyers and patients read cleared as proven. When the paperwork says authorised and the marketing says AI, the missing outcome data is easy to overlook.

Radiology is where this matters most, because it is where most of the devices sit. A hospital deciding which tool to buy will often see an accuracy figure and a regulatory badge. It will rarely see a trial showing that patients under the tool's care did better than patients without it.

Accuracy studies are cheaper and faster to run than outcome trials. A developer can test an algorithm on a stored set of scans in weeks. Following patients through treatment, with a comparison group, takes far longer and costs far more. That gap in effort helps explain why the first kind of evidence is plentiful and the second is rare.

None of this means the devices are useless or unsafe. The audit does not say that, and it would be a mistake to read it that way. It says the evidence that would show whether they help has mostly not been gathered, or at least not registered.

Who pays if the benefit is not proven?

Payers have started to ask the same question from the money side. Our report on insurers saying AI healthcare tools are pushing costs up covers one version of the worry, that tools adopted on the promise of savings can add activity without improving results.

An outcome trial is the natural way to settle that argument. If a scan-reading tool leads to earlier treatment and fewer complications, a hospital or insurer can weigh that against the price. If it simply produces more flagged findings to follow up, the case looks different.

For patients, the stakes are more direct. A finding flagged by software can lead to further tests, appointments and worry. Whether that is worth it depends on evidence about what happens next, and that is precisely the evidence the audit found missing.

How does this fit the wider debate on AI oversight?

Questions about how much testing AI should face before use are not limited to medicine. Industry leaders have been arguing over rules and control, as we covered in how AI industry leaders clash over regulation. Healthcare is an unusually clear test case because the outcome being measured is a person's health.

Much of the AI coverage we run is about scale, in gigawatts, valuations and parameters, as in our look at how much power AI data centres use. This story sits at the opposite end, with a handful of trials.

The same pattern of deployment running ahead of testing shows up elsewhere, including the OpenAI agent incident. In both cases, the technology reached real settings before the evidence about how it behaves there was complete.

Consumer health technology faces a similar question. Companies such as Neko Health, the body-scanning startup we covered in Daniel Ek's 700 million dollar funding round, as Radiology Business reported it, sell the idea that more data about the body leads to better health. Whether it does is again a matter for outcome evidence rather than for product launches.

What should hospitals and patients look for?

The practical question to ask of any clinical AI tool is simple: where is the study showing patients did better? A few follow-ups help sort a strong answer from a weak one.

  • Was the study registered before it began, so the planned measures are public?
  • Did it measure patient outcomes, such as treatment times or recovery, rather than only the accuracy of the software?
  • Did it compare patients treated with the tool against patients treated without it?
  • Was it run in settings similar to the one considering the purchase?

If the answer to the first question is no, that does not make the tool bad. It means the buyer is relying on the regulatory decision and on accuracy data, and should know that is what they are relying on.

What happens next?

The audit points to a gap rather than a fix. Closing it would take more registered trials of cleared devices, and a clearer line between what clearance shows and what it does not.

Developers, hospitals and regulators each have a part. Developers can run and register outcome studies. Hospitals can make them a condition of purchase. Regulators can make the limits of clearance plain to buyers, so that the word does not stand in for evidence it was never meant to supply.

For now, the figure to remember is the small one. Three of 1,357 devices had a registered trial measuring patient outcomes, and the other 1,354 rest on a decision that they met regulatory requirements.

More from Future Technology HQ