FDA-cleared AI lacks proof it helps patients, study finds

By late 2026, the central problem was no longer whether healthcare AI could win clearance, but whether clearance had been mistaken for proof of benefit. Where Tech Meets Bio reported that ARISE’s 2026 synthesis found that among 691 FDA-cleared AI/ML medical devices cleared between 1995 and 2023, fewer than 1% reported patient outcomes, while FDA summaries often omitted basics such as study design and sample size; ARISE also said more than 95% went through the 510(k) pathway, which relies on equivalence and can leave devices “sometimes being cleared on suboptimal evidence.”

That concern intensified in Q3 and Q4 as the installed base kept expanding faster than outcome validation. The Clinical Trial Vanguard noted that as of April 2026, over 1,500 AI-enabled medical devices had received FDA authorization and that 68% of those authorizations occurred since 2022, while ARISE separately counted 1,200+ FDA-cleared tools but said only a minority had undergone peer-reviewed evaluation; together, those findings showed a large, recent wave of cleared products whose prospective clinical evidence had not kept pace, sharpening clinician resistance to tools without robust real-world proof.

Healthcare AI’s bottleneck is not raw capability but the mismatch between what gets measured and what matters at the bedside. HIT Consultant notes that “Most evaluations rely on basic machine learning metrics (AUROC, F1 scores, AUPRC) that measure accuracy, precision, and recall,” yet “accuracy measured in retrospect is necessary but not sufficient,” and, citing Bedi, Liu, Orr-Ewing et al, says 95.4% of studies focused on accuracy while fairness, deployment readiness, and calibration were rarely assessed; “Further, only 5% of the trials studied used real patient care data for evaluation.” That concern also fits a broader performance framing: it cites “a paper published in Nature” and “a competing paper on Arxiv” that “came to the exact same conclusion” about frontier versus vertical performance, even as it emphasizes that “Nature 2 papers sit on opposite ends,” underscoring how benchmark narratives can diverge from bedside utility.

That gap becomes an adoption constraint because benchmark success can still produce unsafe or unusable care. HIT Consultant cites Jabbour, Fohey, Sheppard et al showing “diagnostic accuracy worsened by 11.3% when clinicians were shown biased AI model predictions,” while AI Health Uncut captured the trust test bluntly: “If the CMO doesn’t like the first five cases, you can have 99.999% accuracy and that’s not going to help you”; with data drift and constantly evolving models making “static” evaluations obsolete, the answer increasingly looks like continuous, quality-style monitoring akin to value based purchasing, where “80%” of government payments are tied to tracked quality.

The measurable problem is not that healthcare AI lacks clearances; it is that most cleared products still lack prospective proof that patients do better because they are used. Researchers evaluated all 1,357 AI- and machine learning-enabled medical devices cleared by the U.S. FDA through Dec. 5, 2025 and found only 34 (2.5%) devices were linked to registered prospective trials, while only three devices (0.2%) had been evaluated for patient-centered outcomes such as mortality, morbidity or readmissions—meaning, as TechTarget summarized, “Only three of 1,357 FDA-cleared AI devices tested patient outcomes.” The study ties that evidence gap in part to the FDA’s 510(k) pathway, which does not require independent demonstration of safety and effectiveness via patient outcomes, instead asking vendors to show a tool has “substantial equivalence” to an existing device.

Even that thin layer of prospective evidence is weak and uneven: of the 34 devices linked to registered prospective trials, only 12 (0.9%) posted results and 12 (0.9%) progressed to peer-reviewed publication. TechTarget reported that 32 of the 34 trials, or 94%, were industry-led, with designs “heavily focused on accuracy metrics rather than patient outcomes,” leading the authors to warn the evidence base is “incomplete, systematically biased, and tilted toward optimism”; the gap is especially visible in radiology, which accounts for 78% of cleared devices but had only 1% linked to prospective trials. As the article explains, 510(k) clearance “does not require the vendor to demonstrate independent evidence of safety and effectiveness,” helping explain why so many cleared products reach the market without proof of patient benefit.

Clinician uptake is proving uneven rather than broad: Ipsos found that 41% of healthcare professionals already use AI diagnostic tools weekly, yet trust is highest in lower-friction tasks and remains lower in areas such as mental health assessment and surgical guidance. That pattern fits what WION described as resistance once AI moves beyond narrow, easily checked uses, with a 2026 survey finding 74 per cent of clinicians worried about deskilling and the same share worried about AI inventing false information, fears that compound when doctors must catch failures in higher-stakes decisions.

The trust problem is not abstract: IDC and SAS found employees override AI recommendations because of insufficient explanation (29.9%), factual errors (22.7%), and bias or fairness concerns (22.7%), while SAS CTO Bryan Harris said state-of-the-art agents can have error rates exceeding 25% on complex tasks, which is unacceptable in high-stakes decision-making. Even as integrated or transformative AI programmes rose to 48.7% in 2026 from 38% a year earlier, SAS and IDC said trust has not kept pace, and workforce safeguards weakened too, with ethics, compliance and risk training falling to 26.9% and regular audits and impact assessments dropping 9.1 percentage points.