Clinical AI Oversight Lags as FDA Device List Tops 1,600 Entries

The gap in one picture

American hospitals have not waited for regulators to settle the rules before bringing artificial intelligence into clinical work. Most oversight is now happening inside hospitals themselves, through internal committees, vendor reviews, bias checks and local validation. The coverage also agrees that this oversight is uneven and thin in many places.28 One governance analysis found that half of surveyed U.S. healthcare organizations were using generative AI by April 2026. Only 10% to 15% of large health-system boards had formal structures for overseeing AI.21

Regulators have not been idle. The FDA is still reviewing imaging algorithms one at a time, and states are passing disclosure laws in large numbers. But the federal approach has two tracks that don't line up well. The FDA has kept premarket review for diagnostic imaging AI. Meanwhile, revised federal guidance and a reliance on existing hospital rules have left much of the newer clinical AI to institutions that, by their own account, are not ready to police it. The rules are not missing. They cover the tools that matter least to the current wave of adoption.

The device pipeline: radiology dominates

The best-mapped part of clinical AI is the FDA's public list of AI-enabled devices. One analysis puts the list at 1,614 entries as of its September 4, 2026 update, with 181 decisions dated in 2026 through June 29. Radiology was the lead review panel for 135 of those, or 75%.1 A separate directory says the FDA had authorized more than 1,600 AI-enabled devices by September 2026. It also notes the agency's warning that the list is incomplete.6

The totals differ depending on when each snapshot was taken. One account counted 1,451 authorizations at the end of 2025, with yearly totals rising from 221 in 2023 to 253 in 2024 and 295 in 2025.10 A first-quarter 2026 snapshot showed 1,524 entries, 1,164 of them radiology devices. That quarter's 92 authorizations included 69 for radiology.5 Some trade guides still use the older phrase "more than 1,000" and flag their own numbers as needing verification.2 Every source agrees on the trend: growth is speeding up, and about three-quarters of it is in imaging.

That imaging share is spread across scanner makers and specialist software companies. GE HealthCare holds 130 radiology AI authorizations, Siemens Healthineers 95 and Philips 58, according to one vendor comparison.5 Pure-play triage companies have also grown large. Aidoc is reported to run at roughly 2,000 hospitals and Viz.ai at more than 1,700.10 In January 2026 Aidoc received clearance for what was described as the first clinical AI device built on a foundation model. It is a body-CT triage tool covering 14 conditions, which reported 97% mean sensitivity and 98% mean specificity in its pivotal study.10 Oncology is following the same path. Recent examples include RevealAI-Lung, cleared in January 2026 to score lung nodules for malignancy risk, and an AI-assisted optical coherence tomography system for checking breast-surgery margins, which received premarket approval in March.3

The main caveat is about how these tools get to market. Nearly all radiology AI arrives through the 510(k) pathway, which means it was shown to be "substantially equivalent" to a device already on the market. It is not "approved" in the strict sense.5 One policy handbook says more than 90% of AI devices come through 510(k), and only about 30% have published validation.23

The FDA draws a line on imaging

The clearest federal signal of 2026 came in radiology. Australian company Harrison.ai petitioned the FDA to let manufacturers that already hold a clearance release more detection, diagnosis and triage software without new 510(k) reviews, as long as they kept post-market monitoring, transparency and training measures in place.34 The agency denied the request by letter on April 1. It published a final order in the Federal Register on September 17, effective that day.34

The reasoning matters more than the result. The FDA said one past clearance doesn't show a manufacturer can safely build its next product.39 It also rejected the claim that radiologists would notice if an algorithm's performance drifted. For a possible cancer, the agency said, a reader often can't tell whether the software was right without a biopsy, and there is no standard way to link final diagnoses back to the device.33 Most of the more than 45 public comments opposed the exemption. They raised concerns about patient safety, how fast AI is changing, and post-market surveillance that is still unproven.36 The agency pointed developers toward predetermined change control plans, which let a company get pre-approval for future model updates.3539

Coverage agrees that nothing changes in practice. These devices needed a 510(k) before the order and still do.31 Industry accounts call it a non-event for the regulatory path, while radiology-focused outlets treat it as an important statement of principle. The more telling point is the FDA's own admission that post-market monitoring isn't strong enough yet to replace premarket review. That doubt applies just as much to hospitals, which are now expected to do that monitoring themselves.33

Where federal oversight loosened

The other track runs the opposite way. In January 2026 the FDA revised its guidance on clinical decision support software. Under the revision, software that informs a clinician who can independently review its reasoning may fall outside device rules, and the agency plans not to enforce rules for some tools that give a single recommendation.11 Compliance trackers say this means more AI decision-support tools will reach clinicians without FDA premarket review, leaving the deploying organizations to vet them.11 A public-health handbook goes further. It says generative AI tools for diagnosis or history-taking could now reach clinics without FDA review.23

Not everyone reads the guidance that way. One law-firm summary says AI products that directly shape clinical judgment will likely still need FDA review.15 The more detailed accounts point to more flexibility, not less, and the weight of the coverage supports them. The guidance moves responsibility toward hospitals.

Ambient documentation is the most visible example. As of 2026, AI scribes count as administrative tools, not medical devices, so the FDA does not oversee them. The signing clinician is responsible for any errors, and no ambient-AI vendor accepts clinical liability.28 The FDA has started looking at generative systems more broadly. In August it opened a public docket with a discussion paper on risk frameworks, premarket "competency assessment" and options for post-market monitoring.16 Its broader draft guidance on managing AI devices across their lifecycle, which recommends monitoring plans to catch drift, was still a draft as of August 2026.11

CMS has not created a separate AI requirement in its Conditions of Participation for hospitals. Surveyors instead apply existing quality, staffing and patient-rights standards to ask whether AI use is defined in policy and whether harms are investigated.24 An HHS request for input on speeding up clinical AI adoption closed in February, and no proposed rule had followed as of mid-year.13

States fill in, but not evenly

State legislatures have been far busier. One tracker counts more than 240 healthcare-related AI bills across 43 states in 2026.16 Texas's TRAIGA, in effect since January 1, requires providers to give clear written notice when AI is used in diagnosis or treatment.17 California's AB 489 bars AI systems from implying that they hold a healthcare license.18 Several states now require a licensed clinician or human reviewer before an insurer can deny coverage based on AI.15

The state picture is also changing fast, and secondary coverage hasn't always kept up. Some accounts still describe Colorado's original AI Act, with its impact assessments and bias controls, as heading toward enforcement in mid-2026.19 Other trackers report that Colorado repealed that law in May and replaced it with a narrower one built on notice and human review, with developer duties starting in 2027.14 Most of the newest state laws also target insurer decisions, mental-health chatbots and disclosure. Few address diagnostic accuracy, and none of the bills on liability for AI-caused harm had passed as of late July.13

Why hospitals are the weak link

With responsibility shifting to hospitals, how ready they are matters most. The numbers are not encouraging. A Black Book survey of 182 hospital leaders found that only 29% had put in place and enforced policies covering AI model inventories and sign-offs. Only 22% were highly confident they could produce a full AI audit trail within 30 days. Small hospitals planned to spend a median 2.3% of their IT and safety budgets on AI governance, compared with 6.8% at large systems.26 Other data show that 61% of hospitals test AI tools on their own patients before launch, but fewer than half check deployed tools for bias.21

A systematic review published this month names four regulatory gaps: weak post-market monitoring, little assurance on equity, unclear accountability, and governance capacity concentrated in well-funded institutions.22 Proposed fixes increasingly borrow from pharmacy practice. One framework would have hospitals "prescribe," monitor and "deprescribe" AI systems, using renewal dates and conditions for suspending them.27

The overall picture is consistent. The FDA has kept premarket review for the imaging algorithms that make up most cleared devices. For the faster-growing categories, including decision support, generative tools and scribes, oversight has shifted to hospital committees and state disclosure rules. Many of those committees lack the staff, budgets or audit trails to do the job. The agency's own case against Harrison.ai rested on the point that drift is hard to see from the reading room. That point applies to every hospital now being told the monitoring is its job.