There are now more than 1,000 AI-enabled medical devices cleared by the U.S. Food and Drug Administration. That number has roughly tripled since 2020. And yet a surprisingly uncomfortable question hangs over the entire enterprise: Does any of this stuff actually help patients?
The answer, according to a growing body of research and reporting, is that almost nobody knows. Not the regulators. Not the hospitals buying these tools. And often not even the companies selling them.
A sweeping examination by MIT Technology Review lays bare the fundamental disconnect between the proliferation of AI health tools and the evidence base supporting their real-world effectiveness. The piece, published March 30, 2026, documents how the FDA’s clearance process — which most AI medical devices pass through via the 510(k) pathway — doesn’t require companies to demonstrate that their products improve clinical outcomes. It requires them to show that a new device is “substantially equivalent” to something already on the market. That’s a much lower bar.
Think about what that means in practice. A company can get an AI-powered radiology tool cleared by showing it performs comparably to an existing tool — even if that existing tool was itself cleared without rigorous outcome data. The chain of equivalence can stretch back years, sometimes to devices that predate modern AI entirely. It’s turtles all the way down.
The consequences aren’t abstract. Hospitals are spending millions on AI diagnostic systems, workflow automation, and clinical decision support tools. Health systems are integrating these products into emergency departments, radiology suites, and primary care workflows. Patients are being triaged, scanned, and flagged by algorithms that were validated on curated datasets but may perform very differently on the messy, heterogeneous populations that actually walk through hospital doors.
As MIT Technology Review reports, independent evaluations of AI health tools have repeatedly found performance gaps between what companies claim in regulatory filings and what happens when the software encounters real patients. Some tools that performed well in studies conducted by their manufacturers showed significantly degraded accuracy when tested by outside researchers using different patient populations. Racial and demographic biases have been documented across multiple product categories. And the post-market surveillance infrastructure — the system that’s supposed to catch problems after a device hits the market — remains thin and largely voluntary.
This isn’t a new concern. But it’s intensifying.
The sheer velocity of AI device approvals has outpaced the capacity of regulators, hospitals, and independent researchers to evaluate what’s actually working. The FDA cleared more AI-enabled devices in 2025 than in any prior year, and the pace shows no sign of slowing. Radiology accounts for the largest share — roughly 75% of all cleared AI medical devices — but the technology is rapidly expanding into cardiology, pathology, ophthalmology, and mental health screening.
So what does the FDA actually require? For the 510(k) pathway, which covers the vast majority of AI medical devices, the standard is predicate equivalence. The agency reviews technical specifications, bench testing data, and sometimes limited clinical data. But it doesn’t mandate randomized controlled trials. It doesn’t require head-to-head comparisons with standard care. And it doesn’t require post-clearance studies to confirm that the device delivers measurable patient benefit. The De Novo pathway, used for genuinely novel devices, involves somewhat more scrutiny — but still doesn’t consistently demand the kind of evidence that would satisfy a skeptical clinician.
The FDA has taken steps to modernize its approach. Its Digital Health Center of Excellence, established in 2020, has published guidance on predetermined change control plans — frameworks that allow AI developers to update their algorithms without seeking new clearance for every modification. The agency has also signaled interest in a “total product lifecycle” approach to AI regulation, one that would involve ongoing performance monitoring rather than a single gate-check at the point of clearance. But these frameworks are still largely aspirational. Implementation has been slow, and enforcement mechanisms remain underdeveloped.
Meanwhile, the market keeps growing. Grand View Research estimated the global AI in healthcare market at over $26 billion in 2025, with projections exceeding $150 billion by 2030. Venture capital continues to flow into health AI startups. And the largest technology companies — Google, Microsoft, Amazon — are all making aggressive plays in clinical AI, from ambient documentation tools to diagnostic imaging algorithms.
The result is a strange kind of paradox. There’s more AI in medicine than ever before. And there’s less certainty about its impact than the investment levels would suggest.
Some products do have solid evidence behind them. IDx-DR, an autonomous AI system for detecting diabetic retinopathy, was the first device to receive FDA De Novo clearance for making a clinical decision without physician oversight. It was supported by a prospective, multicenter clinical trial. Viz.ai’s stroke detection platform has published real-world data suggesting it reduces time to treatment. And a handful of radiology AI tools have demonstrated measurable improvements in diagnostic sensitivity in peer-reviewed studies. But these are exceptions. According to research published in Nature Medicine and other journals, the majority of cleared AI health tools have never been evaluated in a randomized clinical trial, and many have no published peer-reviewed evidence of clinical effectiveness at all.
The problem compounds itself at the hospital level. Health system leaders are being pitched dozens of AI products, often with glossy marketing materials and impressive-sounding accuracy metrics. But accuracy on a test dataset is not the same thing as clinical utility. A tool might correctly identify a finding on an image 95% of the time — but if radiologists were already catching 94% of those findings without AI assistance, the marginal benefit is tiny. And if the tool generates a high volume of false positives, it could actually slow down workflows and increase unnecessary follow-up procedures.
Procurement decisions for AI tools are frequently made without the kind of rigorous evaluation that hospitals would demand for a new drug or surgical device. As MIT Technology Review notes, many hospitals lack the internal data science expertise to independently validate AI products before deployment. They rely on vendor-provided performance data, which may not reflect the demographics, imaging equipment, or clinical protocols of their specific institution. It’s a trust-based system operating in a domain where trust should be earned through evidence.
There are efforts underway to change this. The Coalition for Health AI (CHAI), a multi-stakeholder group that includes academic medical centers, technology companies, and federal agencies, has been developing standards for AI evaluation and deployment in clinical settings. The American Medical Association has called for greater transparency in AI performance reporting. And several academic medical centers have established internal AI governance committees tasked with vetting tools before they’re integrated into clinical workflows.
But standardization remains elusive. There’s no universally accepted framework for evaluating clinical AI. Different hospitals use different metrics, different validation methods, and different thresholds for what constitutes acceptable performance. Some institutions run prospective pilot studies before deploying a new tool. Others plug it in and see what happens.
The regulatory picture outside the United States isn’t much clearer. The European Union’s AI Act, which began phased implementation in 2025, classifies most medical AI tools as “high-risk” and imposes requirements around transparency, data governance, and human oversight. But the specifics of how these requirements will be enforced in clinical settings are still being worked out. The UK’s Medicines and Healthcare products Regulatory Agency (MHRA) has published its own AI guidance, emphasizing the importance of real-world performance monitoring. And Health Canada has been exploring adaptive regulatory pathways for AI-based medical devices. None of these frameworks have yet produced the kind of consistent, rigorous evidence base that clinicians and patients deserve.
Part of the difficulty is technical. AI systems in healthcare don’t behave like traditional medical devices. A hip implant doesn’t change after it’s manufactured. An AI algorithm can — and often does. Models can be retrained on new data, updated to reflect new clinical guidelines, or modified to address performance issues identified after deployment. This creates a moving target for regulators. How do you evaluate something that’s designed to change over time?
The FDA’s predetermined change control plans are one answer, but they’ve drawn criticism from some researchers who worry they could allow significant algorithmic changes without sufficient independent review. Others argue that the alternative — requiring new regulatory submissions for every model update — would be so burdensome as to discourage improvement entirely.
Then there’s the data problem. AI health tools are only as good as the data they’re trained on, and healthcare data is notoriously messy, fragmented, and biased. Electronic health records vary wildly across institutions. Imaging protocols differ between hospitals. Patient populations served by an academic medical center in Boston look nothing like those at a rural clinic in Mississippi. A model trained primarily on data from one setting may perform poorly in another — a phenomenon known as distribution shift that has been documented repeatedly in the medical AI literature.
Addressing this requires diverse, representative training data and rigorous external validation. But assembling such datasets is expensive, time-consuming, and fraught with privacy concerns. Many AI companies rely on data partnerships with a small number of large health systems, which may not reflect the broader patient population. And once a product is on the market, there’s limited infrastructure for ongoing performance monitoring across diverse clinical environments.
The financial incentives don’t help. Companies are under pressure to get products to market quickly, and the 510(k) pathway offers a relatively fast route to commercialization. Conducting large-scale clinical trials takes years and costs millions. For a startup burning through venture capital, that timeline can be existential. So the temptation is to do the minimum required for clearance and let the market sort out effectiveness after the fact.
Hospitals, for their part, face their own pressures. Staffing shortages, rising patient volumes, and the relentless push to improve efficiency create a strong pull toward any technology that promises to help. AI vendors know this. Marketing materials emphasize time savings, workflow optimization, and competitive advantage. The evidence question often takes a back seat.
Not everyone is pessimistic. Proponents of health AI argue that demanding the same evidentiary standards as pharmaceuticals would be impractical and counterproductive. Software isn’t a drug. It can be updated, improved, and customized in ways that physical products can’t. And the potential benefits — earlier disease detection, reduced diagnostic errors, more efficient care delivery — are real and significant. The question isn’t whether AI should be used in healthcare. It’s whether the current system for evaluating and deploying it is adequate to ensure it does more good than harm.
Right now, the honest answer is that the system isn’t keeping up. The gap between what’s being sold and what’s been proven is wide. And the patients on the receiving end of these tools — the ones being scanned, scored, and sorted by algorithms — mostly have no idea that the evidence base is this thin.
That’s the part that should bother everyone. Not the technology itself, which holds genuine promise. But the collective willingness to deploy it at scale before doing the hard work of proving it works. Medicine has a long and painful history of adopting interventions that seemed like good ideas but turned out to cause harm — from routine episiotomies to aggressive PSA screening. AI could be different. But only if the people building, buying, and regulating these tools insist on a higher standard of proof than the market is currently demanding.
The tools are here. Over a thousand of them, and counting. The evidence? Still catching up.
WebProNews is an iEntry Publication