AI in Healthcare 2026: FDA Tops 1,600 Devices as Diagnosis Lags

The headline number

The FDA's own count shows how much medical AI has grown. The agency says it had authorized more than 1,600 AI-enabled medical devices for U.S. marketing as of September 2026. Its examples include skin-cancer imaging systems, diabetic retinopathy detectors and software that automates insulin dosing.11 Analysts who downloaded the September 4 update of the list counted 1,614 entries, with the newest decision dated June 29, 2026.1215

The pace keeps rising. Counted by decision date, authorizations went from 226 in 2023 and 235 in 2024 to 335 in 2025.12 One analysis notes that the FDA authorized just six AI-enabled devices in 2015. It also counts 181 in the first half of 2026, which puts this year on track to match or beat last year.14

That device count is the most solid figure in this year's flood of "AI in healthcare statistics" roundups. The rest of the data is much less consistent, and the gaps say something about where clinical AI really stands.

Why the counts don't agree

Published totals for FDA-cleared AI devices in 2026 range from about 1,250 to more than 1,600. One widely shared roundup says the FDA had cleared about 1,250 AI or machine-learning devices by May 2025.21 Another research summary uses roughly the same number but dates it to May 2026.20 A device directory updated in February listed 1,357 entries.17 A July analysis was still using the year-end 2025 figure of 1,451.16

Some of this is just timing. The list is updated in batches, so snapshots fall behind. One analysis counted 1,524 entries in late March 2026 before the September update added more.12 Some of it is method. One annual series counted 295 AI/ML clearances in 2025, while the FDA's list dates 335 authorizations to that year.12 The FDA itself warns that its list is not a complete record of AI-enabled devices. It only reflects products found by searching public summaries for AI-related terms.1218 The list also counts authorizations, not separate products, so one device family can show up more than once.15

Readers should treat "more than 1,600, as of mid-2026" as the current figure. Any 2026 statistic below roughly 1,450 is probably out of date.

Radiology is most of the regulated market

Every dataset agrees that radiology dominates. Of the 1,614 authorizations, 1,230 fall under the radiology panel, about 76 percent. Cardiovascular is far behind at 154, and no other specialty reaches 100.14 Radiology also made up 75 percent of 2026 decisions through June.12 GE HealthCare leads radiology AI makers with 120 total authorizations, followed by Siemens Healthineers at 89 and Philips at 50.12

Almost all of these devices took the FDA's lightest route. Of the 1,614 authorizations, 1,553 were 510(k) clearances, which rest on showing "substantial equivalence" to a device already on the market. Only 40 were De Novo grants and 21 were full premarket approvals.14 An academic review of authorizations through December 2024 found that 84.4 percent of distinct devices use medical images as their main input. It found no evidence of large language models anywhere on the list.12

That last finding matters. Most public talk about AI in medicine is about chatbots and generative models. The FDA-regulated market is mostly narrow, task-specific image tools: software that flags a brain bleed, speeds up an MRI, or ranks a chest X-ray for review.

The radiology market forecasts disagree badly

Market-size estimates show how unreliable the commercial figures can be. Grand View Research puts the global AI-in-radiology market at $14.6 billion in 2025, rising to $20.1 billion in 2026 and $193 billion by 2033.31 Polaris Market Research has a similar starting point of $14.24 billion.33 Coherent Market Insights, however, sizes the 2026 market at $3.42 billion.39 Another roundup cites $2.85 billion for 2025.38

That is a gap of roughly five to six times for the same year. It most likely comes from different definitions of what counts as a radiology AI sale, such as embedded scanner software versus standalone tools. The growth rates in these reports should be read as signs of momentum, not as revenue forecasts. The FDA authorization data is the firmer evidence of real activity.

Hospital adoption is wide but not deep

Adoption numbers also vary with the wording of the survey question. The figure cited most often is that about 80 percent of hospitals use AI in at least one clinical or operational function.21 An Eliciting Insights survey of 120 health systems found that 75 percent were using at least one AI application in early 2026, up from 59 percent a year earlier.23 Physicians report similar growth. The American Medical Association's 2026 survey of 1,692 doctors found 81 percent using AI at work, compared with 38 percent in 2023.29

The growth is mostly not in diagnosis. Clinical note-taking was the most adopted application in the Eliciting Insights survey, at 68 percent, followed by AI-based clinical documentation improvement at 43 percent.23 The AMA survey found that only 17 percent of physicians use AI for assistive diagnosis.24 One review of real-world deployments concluded that fewer than one in five institutions report sustained, highly successful use of AI in core clinical diagnosis.21

Other research points the same way. A peer-reviewed survey of 43 large U.S. health systems found that every one was at least developing or piloting ambient documentation, but only 14 percent had fully deployed it.22 Imaging was the most widely deployed clinical category, with 90 percent of those systems using it in at least limited areas. However, only 38 percent of organizations working on clinical risk-prediction AI reported a high degree of success.24 Deloitte's 2026 outlook found about 30 percent of health systems running generative AI at scale in some areas and just 2 percent running it across the whole organization.29

Smaller hospitals are further behind. In the Eliciting Insights data, 73 percent of systems with 900 or more beds had adopted AI clinical decision support, compared with 13 percent of those under 100 beds.27

Diagnostic accuracy depends on the tool

The accuracy data follows the same divide. Narrow image models trained on labeled data reach specialist-level results, including about 96 percent for diabetic retinopathy detection. General-purpose generative AI averaged 52.1 percent across 83 studies of open-ended diagnosis.30 A multicenter study of brain-bleed detection across 67 organizations found that radiologists working with AI beat the AI alone. Sensitivity was 98.91 percent with AI assistance versus 95.91 percent standalone, and specificity was 99.83 percent versus 87.35 percent.32

The evidence behind authorized devices is uneven. A review of 691 devices found documented premarket safety assessments for only 28.2 percent of them.12 One trade commentary argues that hospitals are signing multi-year contracts for tools that were never tested in trials to see whether they improve outcomes.19

Regulators are easing oversight of non-imaging tools

Regulation has moved in two directions in 2026. On January 6, the FDA issued revised guidance on clinical decision support software, reissued January 29.6 Under the new policy, the agency will generally not enforce device rules against tools that give a single clinically appropriate recommendation, as long as clinicians can independently review the reasoning behind it.13 Earlier guidance pushed such tools toward device status, and some developers responded by offering several options even when only one made clinical sense.1 The FDA also withdrew its guidance on clinical evaluation of software as a medical device.10

The rules for imaging did not change. Software that analyzes medical images or continuous physiological signals to make diagnostic recommendations is still a regulated device.6 The FDA also turned down a petition from Harrison.ai, which kept premarket review in place for the radiology AI categories involved.8 In August, the agency released a discussion paper on generative-AI medical devices that asked for comments on competency-based evaluation, foundation models and agentic AI. It is exploratory and does not change current rules.7

States are filling some of the space. Texas now requires practitioners to tell patients when AI is used for diagnosis.7 California, Alabama, Georgia and Colorado have passed laws requiring humans to make or review AI-assisted coverage decisions.2

What it means for hospitals

Taken together, the data shows two kinds of medical AI moving at different speeds. Imaging AI is regulated, increasingly well tested and growing steadily through the 510(k) pathway. Text-based and generative tools are spreading faster in hospitals, and federal premarket review now covers fewer of them.

The practical result is that hospitals carry more of the burden of checking these tools. With fewer decision-support products going through FDA review, health systems have to evaluate them on their own, at the same moment adoption is speeding up.26 FDA authorization still means something for imaging AI. For the documentation tools and AI assistants now in most hospitals, buyers should not treat it as a sign of quality.