Published by: Tom Morgan (an editorial byline, not an individual reporter). Last updated: October 6, 2026. Editorial process: Drafted with AI assistance. Every statistic was checked against the primary source or the named secondary sources listed at the bottom, and anything we could not verify was left out. This article has not been reviewed by a licensed clinician. It is informational and is not medical advice.
The short answer: The FDA’s public AI-enabled device list had reached 1,524 entries through March 30, 2026, about 76% of them in radiology, and most were cleared through the 510(k) “substantial equivalence” route, which does not require proof of better patient outcomes. The AI tool that has spread fastest in everyday clinical work is the ambient scribe: 62.6% of US hospitals running Epic had adopted one by mid-2025. And the best randomized evidence on how doctors behave with LLMs points to a risk that benchmark scores hide: automation bias.
How many AI medical devices has the FDA authorized?
The FDA maintains a public AI-enabled medical device list. The most recent complete tally we could verify is 1,524 authorizations through March 30, 2026, up from 1,451 at the end of 2025, 882 in March 2024, and just 33 in 2016. The first AI-enabled device, a cervical-cytology screening tool, was authorized in 1995.
| Checkpoint | Cumulative authorizations on the FDA list |
|---|---|
| 2016 | 33 |
| March 2024 | 882 |
| End of 2025 | 1,451 |
| Through March 30, 2026 | 1,524 |
Confidence: Established for the direction and order of magnitude. Probable for exact counts, because trackers disagree at the margins (see below).
Three caveats change how you should read that table:
- The list keeps growing. The FDA’s own page shows entries dated as late as June 29, 2026, so the live total is higher than the March snapshot. Check the FDA page for the current count before quoting a number.
- The list is not a complete census. The FDA says it identifies AI-enabled devices mainly through AI-related terms in authorization summaries, so devices whose summaries don’t use that language can be missed. The agency also says it hopes to identify LLM-based functionality in a future update.
- Trackers disagree on single-year counts. For 2025 alone, published figures range from 295 to 333 clearances (295 in Imaging Wire coverage, 331 in MedTech Dive, and 333 in Axis Intelligence and Voronoi analyses). If you cite one number, say which source it comes from.
Why radiology is about three-quarters of the list
Radiology accounted for 1,104 of the 1,451 entries at the end of 2025 (76%). In 2025 specifically, radiology made up 254 of 333 clearances, versus 27 for cardiovascular and 19 for neurology.
Part of this reflects genuine technical fit: imaging is standardized, high-volume data. Part of it reflects regulatory mechanics, since a new triage or detection algorithm can often be compared with an existing predicate device under 510(k) without a new clinical trial. And part of it may reflect how the list is compiled, because term-based identification tends to catch imaging software more reliably than other specialties.
What FDA clearance does and doesn’t prove
Most AI devices reach the market through the 510(k) pathway. It requires showing substantial equivalence to a device already on the market. It does not require showing that using the tool improves patient outcomes.
That is the central point most “AI in healthcare” content blurs: FDA-cleared and clinically proven to improve outcomes are different claims.
One secondary analysis of the list reports that recalls of cleared AI devices were concentrated soon after clearance and mostly involved products that lacked clinical validation. Treat that as a pattern reported by that analysis, not a settled finding.
Large language models: one narrow clearance, most use outside review
Until recently, the answer to “has the FDA cleared an LLM-based device?” was no. That changed in 2026, narrowly. UpDoc announced on June 25, 2026, a 510(k) clearance (K253281, decided December 23, 2025) that it describes as the first software as a medical device using patient-facing large language models. It is a bounded product for insulin management in adults with type 2 diabetes, working within clinician-defined protocols. Two details matter:
- FDA’s decision summary describes a “conversation service” and doesn’t use the term “large language model,” and the device was classified as a drug-dose calculator. The “first LLM clearance” label comes from the company and secondary reporting, not from FDA language.
- It was cleared by substantial equivalence to an earlier, non-conversational predicate. It is not a general-purpose “AI doctor.”
Confidence: Established that the clearance exists and what it covers. The “first of its kind” framing is the company’s.
Meanwhile, the general-purpose chat tools many clinicians use for differential diagnosis, literature summaries, and documentation mostly sit outside device clearance. Any accuracy claim made for them has not been through FDA premarket review.
The regulatory ground shifted in 2026.
On January 6, 2026, the FDA issued revised final guidance on clinical decision support (CDS) software and on general wellness products, replacing earlier versions. Law-firm analyses describe the changes as softening the FDA’s approach: for example, easing enforcement for certain CDS software that gives a clinician a single diagnosis or treatment recommendation and removing most language that treats time-critical decision support as automatically a device. The revised guidance was issued without a public comment period. (Confidence: Established. Reported consistently by several law-firm analyses; read the guidance text before making compliance decisions.)
Model updates after clearance
In December 2024 the FDA finalized guidance on Predetermined Change Control Plans (PCCPs). A PCCP lets a manufacturer describe in advance certain planned model updates as part of its marketing submission, so those changes don’t each need a new submission. That speeds iteration, and it also means the model version a hospital validated at go-live may differ from the version running two years later. Buyers should ask vendors how updates are communicated and revalidated. (Confidence: Established.)
Ambient AI scribes: the use case that actually scaled
Ambient documentation software listens to a visit and drafts the clinical note. It doesn’t diagnose or prescribe, and that narrow scope is a plausible reason it spread faster than other clinical AI.
The best national figure comes from an American Journal of Managed Care study (Graetz et al., January 2026): among 2,784 US hospitals using Epic, 62.6% (1,744) had adopted an ambient AI tool as of June 2025.
| Group | Adoption among Epic hospitals |
|---|---|
| All Epic hospitals | 62.6% |
| Nonprofit | 70.2% |
| Government | 45% |
| For-profit | 28.8% |
| Metropolitan | 64.7% |
| Non-metropolitan | 54.3% |
Confidence: Established (primary study, consistent across the AJMC page and independent coverage).
Read that number carefully:
- It covers Epic hospitals only, which were 42.4% of the 6,561 US hospitals studied. It is not a figure for all US hospitals.
- It measures whether a hospital has adopted a tool, not how many clinicians use it on how many encounters.
- The study associates adoption with workload, financial performance, and ownership, and reports lower adoption in the Midwest than the South. That raises an equity question: better-resourced systems are getting these tools first.
- Three products (DAX Copilot, Abridge, and ThinkAndor) accounted for more than 80% of implementations among Epic hospitals.
What the benefits evidence shows
A multicenter quality improvement study in JAMA Network Open (Olson et al., October 2025) followed 263 ambulatory physicians and advanced practice practitioners across six US health systems that deployed the Abridge scribe. After 30 days, the share reporting burnout fell from 51.9% to 38.8%. The design was a before-and-after survey with voluntary participants recruited by the health systems’ digital health leaders and no control group, so it shows an association, not proof that the scribe caused the change.
A separate multi-site study of five academic medical centers, reported in JAMA, found more modest effects: roughly 16 minutes of documentation time saved per 8 hours of patient care on average, concentrated among heavy users. (Confidence: Probable. Figures come from a secondary summary; check the paper before quoting them.)
The consent problem
Two issues rarely appear in adoption stories.
Consent and disclosure. In April 2026, patients filed a proposed class action in the US District Court for the Northern District of California against Sutter Health and MemorialCare (Memorial Health Services). It alleges the systems used Abridge’s ambient scribe to record and process visit conversations without meaningful, informed consent, citing California’s wiretap statute, the federal Wiretap Act, and California’s Confidentiality of Medical Information Act. These are allegations, not findings, and the defendants dispute them. California requires all parties to consent to recording a confidential communication, which is why the case centers on how consent is obtained and documented.
Disclosure rules are also tightening. Texas’s TRAIGA (House Bill 149), effective January 1, 2026, requires healthcare providers using an AI system in relation to a patient’s service or treatment to disclose that to the patient, no later than the date the service or treatment is first provided. Texas SB 1188, effective September 1, 2025, adds disclosure duties for practitioners who use AI for diagnostic purposes. Requirements vary by state and are changing, so check your own state’s rules. (Confidence: Established for the Texas laws and the lawsuit’s existence; legal analyses agree.)
Review quality. Clinical notes feed coding, quality measures and risk models. If clinicians skim or skip review of AI-drafted notes, errors can flow into those downstream systems. We found little published research measuring that effect.
Diagnostic AI: strong on benchmarks, riskier in the room
What the benchmarks show
Two recent studies tested multimodal models on the NEJM Image Challenge, a set of published image-based diagnostic quiz cases.
- A 2026 analysis by Sheng and colleagues tested GPT-4o, Claude 3.7 and Doubao on 272 consecutive cases and compared them with 16,401,888 physician responses from NEJM’s published results. It concluded that multimodal testing substantially exceeded physician performance and that low physician-AI agreement suggests the models reason differently from people. We could confirm this as a preprint on Research Square but not a final peer-reviewed publication, so check for the journal version before citing it as such.
- A January 2026 Frontiers in Medicine study compared four models (o4-mini-high, Claude 4 Opus, Gemini 2.5 Pro and Qwen 3) on 200 cases against human evaluators and tested the effect of prompt engineering. A leading model beat both its peer models and the human participants. The authors note that Image Challenge images are high-quality and don’t reflect the noise and variation of routine imaging.
Be careful what “physician performance” means in these comparisons. An earlier analysis of 945 Image Challenge cases found the majority vote of respondents answered 90.8% correctly, ahead of every multimodal model tested at the time, while individual respondents scored far lower. A claim that AI “beats physicians” depends on whether the comparison is against the average individual or a collective. Specific percentages from the 2026 preprint are not reproduced here because they could not be verified against a final version.
What happens when the AI is wrong
The most important randomized evidence is a trial reported in NEJM AI: “Automation Bias in Large Language Model-Assisted Diagnostic Reasoning Among Physicians Trained in AI Literacy.”
- Who: 44 physicians in Pakistan, all of whom had completed a 20-hour AI-literacy training program.
- Task: six clinical vignettes (264 cases in total), with optional consultation of ChatGPT (GPT-4o).
- Design: one group received error-free AI suggestions. The other received suggestions with deliberate, clinically significant errors in three of the six vignettes.
- Result: the error-exposed group averaged 73.3% on the primary outcome, a composite diagnostic reasoning score, versus 84.9% for the control group, an adjusted difference of 14.0 percentage points (95% CI -18.9 to -9.1). About 70% of participants in both groups consulted the AI.
Confidence: Established (trial registry entry, preprint and journal summary agree on the headline numbers).
Two details matter for interpretation. The comparison group saw accurate AI suggestions, not no AI, so the study measures the cost of erroneous advice, not the net benefit of AI. And the vignettes were fictitious, the sample small, and all physicians were from one country, so the size of the effect elsewhere is unknown. The finding that AI-literacy training did not prevent the effect is still a strong caution against assuming trained clinicians will catch AI errors. A June 2026 preprint tests mitigation approaches in a follow-up trial; it is not peer-reviewed and its results are not summarized here.
How many doctors actually use AI?
The AMA’s 2026 Physician Survey on Augmented Intelligence (1,692 physicians, fielded January 15 to February 2, 2026) reports that 81% of physicians use AI in their practice, up from 38% in 2023. The average number of use cases per physician rose from 1.1 to 2.3. Summarizing medical research was the most common use at 39%, up from 13% in 2024. In coverage of the survey, 88% of physicians said they were at least mildly concerned about skill loss from AI use.
Confidence: Established for the AMA’s headline figures. Note the AMA’s report describes the 81% as physicians reporting awareness or use of AI in their practice, so avoid reading it as “81% use AI for every patient.”
What the evidence still can’t tell you
- Whether cleared AI devices improve outcomes. Equivalence-based clearance doesn’t test this, and large outcomes datasets are scarce.
- How model updates under PCCPs change real-world performance over time.
- How far LLM-based tools will go through formal clearance, and what evidence the FDA will require beyond narrow, bounded products.
- How AI-drafted notes alter the quality and risk data built from them.
- Whether the adoption gap (nonprofit vs. for-profit, South vs. Midwest, large vs. small) translates into differences in care.
A practical checklist for evaluating clinical AI
| # | Question to ask | Why it matters |
|---|---|---|
| 1 | Is it FDA-cleared, and via which pathway? | 510(k) shows equivalence, not outcomes benefit |
| 2 | Is there prospective or outcomes evidence in a setting like ours? | Benchmarks and vignettes don’t transfer automatically |
| 3 | Who reviews AI output before it enters the chart? | Unreviewed notes can contaminate downstream data |
| 4 | How are model updates communicated and re-validated? | Your go-live validation may not describe the current version |
| 5 | Does the workflow make clinicians form their own judgment first? | Training alone did not prevent automation bias in the trial above |
| 6 | How is patient consent or disclosure handled, and does it meet our state’s rules? | Disclosure laws and litigation risk are evolving |
| 7 | Is the tool cleared, or does it sit outside FDA review as decision support? | The evidence standard differs |
Frequently asked questions
How many AI medical devices has the FDA approved?
The FDA’s AI-Enabled Medical Device List had 1,524 entries through March 30, 2026, and has been updated since. About 76% are in radiology. Technically, most are “cleared” through 510(k) rather than “approved” through premarket approval.
Does FDA clearance prove an AI device improves patient outcomes?
No. The 510(k) pathway requires substantial equivalence to an existing device, not evidence of better outcomes.
Has the FDA cleared a large language model device?
Narrowly, yes. UpDoc announced a 510(k) clearance in June 2026 for an insulin-management product using patient-facing LLMs. It is a bounded product, and the company’s “first LLM clearance” framing is not language from FDA’s own decision summary. General-purpose chat tools used by clinicians mostly remain outside device clearance.
How widely are ambient AI scribes used?
Among hospitals using Epic, 62.6% had adopted one by June 2025, per an AJMC study. That measures hospital adoption, not how often clinicians use the tool.
Can AI make doctors less accurate?
In one randomized trial of 44 AI-trained physicians, those shown deliberately erroneous GPT-4o suggestions scored about 14 percentage points lower on diagnostic reasoning than those shown correct suggestions. The study was small and used fictitious vignettes, but it shows the risk is real.
Do patients have to be told when AI is used in their care?
It depends on the state. Texas, for example, requires providers to disclose AI use in relation to a patient’s care under TRAIGA, effective January 1, 2026. Check the rules where you practice.
Is this article medical advice?
No. It summarizes published data for informational purposes. Clinical decisions should involve a licensed professional.
Sources and confidence
Confidence legend: Established = checked against the primary source or consistent across independent sources. Probable = reported by a secondary source we could not check against the primary document.
Regulation and devices
- FDA, List of Artificial Intelligence-Enabled Medical Devices
- FDA final guidance on Predetermined Change Control Plans for AI-enabled device software functions (December 2024), summarized by Foley and Ropes & Gray
- FDA revised CDS and general wellness guidance (January 6, 2026), summarized by King & Spalding, Polsinelli and The FDA Law Blog
- UpDoc 510(k) K253281: McGuireWoods analysis, Innolitics, IntuitionLabs
- Secondary trackers of the FDA list: MedTech Dive, Voronoi, IntuitionLabs, Imaging Wire coverage
Adoption, benefits and consent
- Graetz et al., “Ambient AI Tool Adoption in US Hospitals and Associated Factors,” American Journal of Managed Care, January 2026: ajmc.com
- Olson et al., “Use of Ambient AI Scribes to Reduce Administrative Burden and Professional Burnout,” JAMA Network Open, October 2025: PMC12492056
- Sutter Health / MemorialCare class action (N.D. Cal., April 2026): Alston & Bird analysis, TechTarget
- Texas TRAIGA (HB 149) and SB 1188: Norton Rose Fulbright, Texas Medical Liability Trust
Clinical evidence and surveys
- “Automation Bias in Large Language Model-Assisted Diagnostic Reasoning Among Physicians Trained in AI Literacy,” NEJM AI; preprint: medRxiv; registration: NCT06963957
- Sheng et al., “Multimodal Large Language Models Challenge NEJM Image Challenge” (preprint): Research Square
- “Diagnostic performance of large language models on the NEJM image challenge,” Frontiers in Medicine, January 2026: PMC12823889
- Earlier analysis of 945 NEJM Image Challenge cases: PMC11306783
- American Medical Association, More than 80% of physicians use AI professionally (March 2026) and the 2026 Physician Survey on Augmented Intelligence
Update ledger
| Date | Change |
|---|---|
| October 6, 2026 | Full re-verification against primary and named secondary sources. Corrected the description of the NEJM AI trial’s control group, the scope and date of the AJMC adoption figure, and when the FDA list crossed 1,000. Replaced the “no LLM device has been cleared” claim with the June 2026 UpDoc clearance and its caveats. Corrected the consent-litigation description to the actual California case. Removed statistics that could not be verified. Replaced an unsupported reviewer credential with an accurate editorial-process note. |