FDA's Own Device List Shows 96% of AI Medical Tools Cleared Without Clinical Trials, Most Never Tested by Sex

Since the JAMA Network Open cross-sectional study of 903 FDA-authorized AI devices was published in April 2025, the finding that fewer than one in three had been validated on sex-specific patient populations appears repeatedly as researchers dig into the FDA's own authorization data. A regulatory analysis published September 5, 2026 by IntuitionLabs shows why: the pathway most of these devices travel through was never designed for AI in the first place.

The 510(k) Problem

Of the 1,614 entries on FDA's AI-Enabled Medical Device List as of September 5, 2026, 96.2% (1,553 devices) carry a 510(k) submission number, according to IntuitionLabs' tabulation of the agency's own downloadable list. Only 2.5% (40) went through De Novo classification and 1.3% (21) through Premarket Approval, the two pathways that require more rigorous review.

The 510(k) route lets a device to market by showing it's "substantially equivalent" to an already-cleared predicate device. It does not require prospective clinical studies. That process was built for comparatively simple hardware, not adaptive machine-learning systems whose performance can shift across demographic subgroups in ways that aren't obvious until the software is already in use, as Tech Times reported citing the JAMA analysis.

The FDA authorized 168 machine-learning-enabled Class II devices in 2024, a record year, according to Tech Times' review of the underlying data. Only 15.5% of those devices provided any demographic performance breakdown. Just 29.2% reported both sensitivity and specificity, the two most basic accuracy metrics a diagnostic tool can offer. In almost a quarter of all 903 device submissions reviewed in the JAMA study, sponsors told the FDA no clinical performance study had been conducted at all.

A separate PLOS Digital Health analysis of 1,357 devices, cited in the IntuitionLabs report, found that only three devices, 0.2%, evaluated patient-centered outcomes like mortality or hospital readmission rather than raw technical accuracy. A peer-reviewed taxonomy tracking 1,016 authorizations from 1995 through 2024 covering 736 unique devices shows how much the landscape has grown even as the evidence bar stayed low.

What FDA Has Done About It

The agency isn't ignoring the issue entirely. FDA issued final guidance in December 2024 on Predetermined Change Control Plans, which let manufacturers pre-specify how an AI device may be updated after it reaches the market without a new submission each time. But independent research cited by IntuitionLabs found voluntary adoption of these plans "remains low" among devices authorized between 2023 and 2025, with unidentified barriers still discouraging manufacturers from using the tool.

A January 2025 lifecycle guidance document remains in draft form, nonbinding, and not for implementation, per FDA's own language. In August 2026, FDA's Digital Health Center of Excellence released a discussion paper, not guidance, on how oversight might need to change as generative AI enters medical devices, according to the National Law Review. The paper explicitly does not impose new requirements. It flags that generative tools "do not always provide the same response twice" and may keep changing after they hit the market through software updates supplied by third parties, complicating any fixed testing standard.

Separately, FDA's Office of Women's Health will have its Associate Commissioner, Dr. Kaveeta Vasisht, deliver the opening keynote at Women's HealthX 2026 in Boston this December, according to an announcement from event organizer Alpha Events. The summit will focus on sex-differentiated clinical trial design and evidence generation, alongside representatives from Novartis, Humana, Elevance Health and other payers and drugmakers.

The Fair Case for the Current Rules

Defenders of the 510(k) pathway point out it exists for a reason: speed and cost. Requiring every AI-enabled device update to clear a full prospective clinical trial could slow beneficial diagnostic tools from reaching doctors and patients. The pathway has cleared thousands of conventional, non-AI devices for decades without the current level of scrutiny. None of the sources in this reporting document a specific patient who was harmed because an authorized AI device lacked sex-specific validation. The studies establish a testing and transparency gap. They do not establish a proven injury.

A Wider Oversight Debate

The underlying question of whether regulators are keeping pace with AI isn't confined to medical devices. OpenAI disclosed six instances of its models acting outside intended behavior in recent months, including fabricating information and concealing mistakes in task summaries, according to Fox News. Geoffrey Hinton, the Nobel Prize-winning AI researcher known as the "Godfather of AI," told lawmakers in a closed-door briefing this week that regulators have "maybe a year" to get ahead of the technology, warning that major companies are racing to build increasingly capable systems "without putting enough focus on how the systems will behave."

Outside the U.S., some health systems are taking a more cautious line before AI diagnostics scale up at all. Professor Edward Kunonga, chair of the Healthcare Excellence Indaba, said ahead of the World Health Expo 2026 in Johannesburg that "local validation on South African and African patient populations is essential," and that a successful pilot elsewhere "does not guarantee that an AI diagnostic tool will work safely and equitably in routine practice."

Back in Washington, no legislation forcing sex-specific or demographic validation on AI medical devices has been introduced, and FDA's most recent guidance documents remain draft or discussion-stage. Whether the agency moves from discussion papers to binding rules before the AI device count climbs past its current 1,614 entries remains an open question.