How Many AI Devices Has the FDA Authorized? A 2025–2026 Breakdown

Through the end of 2025, the FDA-authorized artificial intelligence device landscape stood at roughly 1,451 AI-enabled medical devices, with 295 clearances in 2025 alone. Radiology still accounted for about three quarters of the cumulative list and 71.5% of the 2025 clearances, so the shortest honest answer is: FDA artificial intelligence authorization is now large and growing quickly, but it remains mostly an imaging-device story.[1][2]

That count is useful, but it is not as clean as a trial enrollment figure or a registry denominator. The FDA's public AI-enabled device list is updated periodically and identifies devices largely from AI-related terms in public summaries; the agency itself says the list is not intended to be a comprehensive resource.[3] Innolitics' 2025 analysis also used AI-assisted identification from FDA public summaries and acknowledged possible omissions, while arguing that the directional patterns remain reliable.[2] For a mid-2026 operating baseline, those caveats do not make the numbers unusable. They make the denominator worth watching.

The Count Depends On What Is Being Counted

The 1,451 figure is cumulative through 2025. It is not the number of products currently installed in hospitals, not the number of unique algorithms actively running in clinical workflows, and not evidence that each authorized device has demonstrated improved outcomes in routine care. It is a regulatory authorization count drawn from the FDA AI-enabled device landscape as reported after the agency's 2026 list update.[1]

The 295 figure is annual: AI/ML medical device clearances in 2025. Innolitics described it as a record year and tied those clearances to 221 unique manufacturers.[2] That is the part that should catch attention. Even if the public list misses some devices or classifies some entries imperfectly, a record annual volume on top of a four-digit cumulative base says the FDA artificial intelligence device pipeline is no longer a small specialty queue.

Those rows answer different questions. A procurement team asking whether a category is mature should care about cumulative authorization volume. A regulatory team forecasting submission load should care about annual clearances. A hospital governance committee deciding whether to prioritize radiology AI review should care about specialty concentration. Treating all of those as the same number is how a clearance list turns into a sales deck.

Radiology Is Still The Center Of Gravity

Radiology's dominance is not a historical footnote. It is the present structure of the FDA-authorized AI device market. In 2025, radiology accounted for 71.5% of AI/ML medical device clearances in Innolitics' review.[2] On the cumulative FDA list reported in March 2026, radiology represented about 76% of the total.[1]

The reason is not mysterious. Radiology produces standardized digital images, depends on pattern recognition, already has well-developed workstation workflows, and has decades of regulatory precedent for imaging software. An AI tool that flags a suspected finding on a CT, prioritizes a worklist, segments anatomy, or measures a lesion fits more naturally into existing device categories than an AI system that proposes a treatment plan across a messy longitudinal record.

Cardiovascular and neurology are visible, but they have not displaced radiology. Cardiovascular devices accounted for 8.8% and neurology devices for 4.7% in the reported landscape.[2] Those are meaningful shares if the question is whether FDA-authorized AI is moving beyond radiology at all. They are not meaningful enough to support the claim that the center of gravity has changed.

This is where the distinction between authorization and adoption matters. A device can be FDA-cleared and still fail to win budget, clinician trust, integration support, reimbursement, or measurable local value. Clearance means the device passed a regulatory pathway for its intended use. It does not mean a hospital should automatically buy it, turn it on, or advertise it as clinically beneficial without local review.

Most Devices Still Move Through 510(k)

The pathway mix is almost as important as the specialty mix. Industry analysis cited by Censinet estimates that 95–97% of AI/ML medical devices clear through the 510(k) pathway, compared with about 2.9% through De Novo and less than 0.5% through PMA.[4] That statistic should be handled as an industry-derived estimate, not an official FDA-published aggregate, but it matches the practical shape of the field: most authorized AI devices are being reviewed as substantially equivalent to predicate devices, not as high-risk first-of-kind technologies.

That pathway dominance has consequences. A 510(k)-heavy market tends to reward devices that can be framed within existing intended uses, existing clinical workflows, and existing performance-comparison logic. It also helps explain why image-analysis products have scaled faster than broader clinical decision systems. The regulatory path is not the only reason, but it is part of the reason the landscape looks familiar even as the word “AI” makes it sound new.

Review time is also becoming easier to benchmark. Censinet reported a median 510(k) review time of 142 days for AI/ML medical devices and said about half of AI/ML 510(k) submissions receive an Additional Information request, which can pause review for 30–60 days.[4] That does not make review friction disappear. It does mean regulatory planning can be discussed in months and submission cycles rather than in vague warnings about an undefined AI exception.

Software Dominates, But Not All Software Changes The Same Way

In 2025, 62% of AI/ML clearances were Software as a Medical Device, and 63% were diagnostic rather than therapeutic, according to Innolitics.[2] Those proportions fit the broader pattern: the authorized field is weighted toward software that analyzes patient data and supports detection, triage, measurement, or interpretation.

The more interesting 2025 development is not only how many software devices cleared, but how some of them are being allowed to change. Innolitics found that 10.2% of 2025 AI/ML clearances included an authorized Predetermined Change Control Plan, or PCCP.[2] A PCCP lets a manufacturer specify, in advance, certain future modifications and the method for controlling them, rather than treating every meaningful algorithm update as an entirely separate regulatory surprise.

That is a procedural signal, not a clinical-outcome signal. A PCCP does not prove that an updated model will help patients, and it does not excuse a hospital from validating how a tool performs in its own environment. But it does show that iterative software change is being pulled into a planned regulatory mechanism. For AI devices, that matters. Static clearance language has always been an awkward fit for models that manufacturers expect to refine over time.

What The Devices Actually Use As Inputs

A deeper taxonomy published in npj Digital Medicine in July 2025 looked at 1,016 FDA authorizations through September 27, 2024, so it should not be blended into the end-of-2025 total of 1,451.[5] Its value is different: it helps describe what authorized AI devices were actually built to process before the late-2024 and 2025 clearance surge.

In that taxonomy, 84.4% of FDA-authorized AI devices used images as their core input, while 14.5% used signals.[5] That finding is the technical mirror of the specialty distribution. The authorized market is not merely “healthcare AI” in a broad abstract sense. It is largely image-centered AI, with a smaller signal-processing segment and a much thinner presence for other data types.

The same study reported that no large language model-based medical devices were identified in the FDA list as of September 2024.[5] That is notable because public attention has moved quickly toward foundation models, clinical copilots, and generative interfaces. The authorized device landscape had not moved at the same pace by that cutoff. As of that analysis, the regulated AI device base was still dominated by systems reading images and signals, not by LLMs reading and generating clinical text.

That may change. It should be measured when it does. Until then, putting LLM expectations onto the current FDA-authorized device list blurs two different markets: the highly visible world of generative AI pilots and the more regulated world of authorized medical devices.

The 2025 Acceleration Is Real, But It Is Not A Deployment Mandate

A record 295 clearances in a single year is not a rounding error.[2] It changes the workload for compliance teams, clinical governance committees, health technology assessment groups, and procurement offices. More devices mean more intended-use statements to read, more predicate relationships to understand, more model-update language to track, and more local questions about data fit, workflow burden, monitoring, cybersecurity, and liability.

The mistake is to treat clearance volume as a proxy for clinical benefit. The FDA list can show that a device was authorized. It can show specialty, pathway, and sometimes the public summary language around performance. It does not tell a hospital whether the device will reduce missed diagnoses in its own patient population, save radiologist time after integration, reduce downstream testing, or create new alert fatigue. Those are adoption questions, not authorization counts.

For a hospital AI governance committee, the practical sequence is still sober: confirm the authorization and intended use, identify whether the product is SaMD or embedded in hardware, check whether a PCCP affects future update governance, compare the submitted claims with the proposed local use, and decide who will monitor performance after go-live. The FDA count can start that conversation. It cannot finish it.

A Mid-2026 Baseline

The most defensible mid-2026 reading is balanced. FDA-authorized healthcare AI is large, fast-growing, and increasingly procedurally mature. The 2025 clearance record and the appearance of PCCPs show a regulatory system getting more accustomed to software that evolves. At the same time, the authorized landscape remains concentrated in familiar imaging-heavy, 510(k)-cleared device categories.

That baseline is useful precisely because it is narrow. If later FDA updates show a larger PCCP share, a sustained rise in non-radiology specialties, more devices built around non-image inputs, or actual LLM-based medical device authorizations, those changes will be easier to recognize. Through 2025 and into mid-2026, the authorized AI device landscape was expanding quickly, but its center was still radiology, images, software, and 510(k).

References

1. FDA Updates AI List with New Clearances, Imaging Wire, Mar 2026.
2. 2025 Year in Review: AI/ML Medical Device 510(k) Clearances, Innolitics, Dec 2025.
3. Artificial Intelligence-Enabled Medical Devices (FDA AI-Enabled Device List), FDA.
4. AI Medical Devices: FDA Approval Process, Censinet, Jun 2026.
5. How AI is used in FDA-authorized medical devices: a taxonomy across 1,016 authorizations, npj Digital Medicine, Jul 2025.