# AI cleared a thousand devices to read your scan. The one that changed medicine listens to the room. > The public imagines diagnostic AI catching what doctors miss. The actual deployment win is documentation, and meanwhile half the cleared diagnostic devices were never shown to work in a clinical study. - **Pillar:** Healthcare - **Author:** Aditya Marin Gasga (Founding Editor) - **Published:** 2026-06-24T15:35:00.000Z - **Tags:** ai-governance, enterprise-ai, healthcare ## TL;DR The headline story — AI catching the tumor a radiologist missed — is the smaller one. The technology actually changing clinical work today is the ambient scribe that drafts the visit note, because it hands clinicians back an hour. And a cleared device is not a proven one: of 903 FDA-authorized AI devices, only about half reported a clinical performance study and a quarter reported none. ## Key takeaways 1. Most FDA-cleared AI is diagnostic imaging — 1,250+ AI-enabled devices by July 2025, about three-quarters in radiology — but clearance is a marketing authorization, not proof of clinical benefit. 2. Of 903 FDA-approved AI devices, only about 50% reported a clinical performance study and about 25% explicitly reported none. 3. The real adoption win is ambient documentation: scribes that draft the visit note attack the ~2 hours of paperwork per hour of care that drive burnout — an easier sale than a diagnostic model that needs trust, a workflow, and a reimbursement code. 4. The cost is privacy: several scribe vendors reserve broad rights to reuse de-identified consultation data — a recording that outlives the appointment. 5. Watch the clinical-study rate, not the clearance count, and whether ambient tools get held to the same evidence bar as imaging models. Right now they are not. import SignalChart from '~/components/viz/SignalChart.astro'; The story everyone tells about AI in medicine is diagnostic: the algorithm that spots the tumor a tired radiologist missed. That is the smaller story. The bigger one is that the technology actually changing clinical work right now is a microphone in the exam room, transcribing the visit so the doctor can look at the patient instead of the keyboard. The scan-reading models are real, but most of them are still waiting for adoption, and a striking share were never proven to work where it counts. What changed in the room is not the diagnosis. It is who is typing. ## The scoreboard says diagnosis. The floor says documentation. Count the regulatory wins and diagnosis looks dominant. The FDA's public database listed [over 1,250 AI-enabled medical devices as of July 2025, up from about 950 a year earlier](https://bipartisanpolicy.org/issue-brief/fda-oversight-understanding-the-regulation-of-health-ai-tools/), with [295 cleared in 2025 alone](https://innolitics.com/articles/year-in-review-ai-ml-medical-device-k-clearances/). A systematic review found that [roughly three-quarters of authorized devices are in radiology](https://pmc.ncbi.nlm.nih.gov/articles/PMC12595527/), where vendors like Aidoc kept [winning expanded FDA designations through 2025](https://en.wikipedia.org/wiki/Aidoc). Most of them are predictive models that flag a finding, not generative systems that write. By the clearance scoreboard, AI in medicine is an imaging story, and a fast-growing one: the same review counted [221 device approvals in 2023 alone against a handful a decade earlier](https://pmc.ncbi.nlm.nih.gov/articles/PMC12595527/). By the adoption scoreboard, it is a paperwork story. Physicians have long spent [about two hours on documentation and desk work for every hour of direct patient care](https://en.wikipedia.org/wiki/Physician_burnout), the single largest driver of burnout. Ambient documentation tools, which listen to a visit and draft the note, [spread rapidly through 2024](https://en.wikipedia.org/wiki/Automated_medical_scribe) precisely because they attack that hour, not the diagnosis. The diagnostic model needs a radiologist to trust it, a workflow to slot into, and a reimbursement code to justify it. The scribe just needs the doctor to stop typing. One of those is a far easier sale. ## What does not work, stated plainly Here is the part the press releases skip. A cross-sectional study of [903 FDA-approved AI devices found that clinical performance studies were reported for only about half, and a quarter explicitly stated that no such study had been conducted](https://pmc.ncbi.nlm.nih.gov/articles/PMC12044510/). The clearance says the device is safe enough to market. It does not say the device was shown to help patients in the setting where it will be used. A separate taxonomy of [1,016 authorizations noted that the FDA's own list is updated irregularly](https://www.nature.com/articles/s41746-025-01800-1), which tells you how much of this market moves faster than its oversight. That gap is the difference between a cleared algorithm and a useful one. It is the same distance between a pilot that demos well and a deployment that survives contact with the real world — the [gap that strands most AI projects short of production](/agent-roi-pilot-production). An imaging model trained at three academic hospitals can degrade quietly at a rural clinic with a different scanner and a different patient mix, and nothing in a marketing clearance guarantees otherwise. ## The altitude shift Move from the dataset to the person. A device's reported accuracy is an abstraction. The clinician's after-hours charting is not. When an ambient scribe shaves the documentation tail off a workday, the measurable result is a physician who finishes notes before going home instead of at 10 p.m., which is the metric health systems actually buy. The diagnostic model promises a better answer. The scribe gives back an hour. In a profession bleeding clinicians to burnout, the hour is worth more than the answer, at least today. There is a cost on the other side of that ledger, and it is not clinical. Several scribe vendors reserve broad rights to reuse de-identified consultation data, which turns the quiet recording of a private visit into a [data asset the patient never priced](https://en.wikipedia.org/wiki/Automated_medical_scribe). The hour saved is real. So is the recording that outlives the appointment. ## What to actually watch The useful question is not how many AI devices the FDA clears next year. It is what fraction arrive with a clinical study attached, and whether the ambient tools quietly running in exam rooms get held to the same evidence bar as the imaging models that get the headlines. Right now they are not, because documentation is regulated more lightly than diagnosis, even though it is the part patients are actually exposed to. A cleared device that was never studied in the clinic is a permission slip, not a result. Treat the count as marketing and the clinical-study rate as the number that matters, and the medicine gets easier to judge. ## FAQ ### How many AI medical devices has the FDA authorized? The FDA's public database listed over 1,250 AI-enabled medical devices as of July 2025, up from roughly 950 a year earlier. About three-quarters are radiology devices, and most are predictive rather than generative models. ### Does FDA clearance mean an AI device was proven to work? Not necessarily. A study of 903 FDA-approved AI devices found clinical performance studies reported for only about half, with a quarter explicitly reporting none. Clearance reflects a marketing authorization, not a guarantee of clinical benefit in the setting where the device is used. ### What is an ambient AI scribe? It is a tool that listens to a clinical visit and drafts the documentation, which a clinician then reviews and edits. Adoption rose sharply from 2024 because it targets the documentation burden that drives physician burnout, rather than attempting to make a diagnosis. ### What is the main risk with ambient scribes? Beyond transcription errors, several vendors reserve broad rights to reuse de-identified consultation data, raising privacy questions about secondary use. Clinicians and health systems are advised to review data-use terms before deploying these tools.