How is AI used in healthcare?
Most real medical AI is narrow image analysis, not chatbots. The FDA had authorized roughly 1,450 AI-enabled medical devices by the end of 2025, and about three-quarters are radiology tools that flag findings on scans. Other major uses: ambient note-taking, triage, and drug discovery. Nearly all keep a clinician making the final call.
Why — the first-principles explanation
Medicine has a structural mismatch: the number of images and signals that need reviewing grows faster than the number of humans qualified to review them. A CT scan can be hundreds of slices. A single ICU bed generates continuous waveforms. Meanwhile a radiologist's attention is finite and expensive. That gap is exactly where pattern recognition scales and humans don't — which is why the overwhelming majority of deployed medical AI is image and signal analysis, not conversation.
The regulatory picture makes this concrete. The FDA maintains a public list of AI-enabled medical devices authorized for U.S. marketing. As of the end of 2025 it held roughly 1,450 entries, and about 76% are radiology — algorithms that detect a possible bleed, measure a nodule, triage a stroke to the top of the queue. These are cleared as devices, with defined intended uses and defined populations. That's a completely different thing from a chatbot answering health questions, which is generally not FDA-authorized at all.
The second big use is invisible to patients but huge for clinicians: ambient documentation. Speech models listen to the visit and draft the note. This isn't clever; it just attacks the largest time sink in the job. The third is triage and prioritization — reordering worklists so the dangerous case is read first, which saves time even when the algorithm adds no diagnostic accuracy. The fourth is drug discovery, where models predict protein structures and screen candidate molecules, compressing steps that took years.
What's genuinely uncertain is impact on outcomes. Most authorizations rest on showing the device performs comparably to a standard, not on randomized evidence that patients live longer. Performance also drifts: a model trained on one hospital's scanner and population can degrade at another. And there's a well-documented failure mode where an accurate algorithm changes nothing because the workflow ignores it. So the accurate summary of 2026 is: AI is broadly deployed in medical imaging and documentation, its diagnostic accuracy in narrow tasks is often excellent, and its proven effect on patient outcomes is still thin.
This is medical information, not medical advice. No AI tool described here replaces a clinician, and none should be used to self-diagnose.
An example that makes it click
Think of an airport bag scanner. Thousands of bags roll past, and a human screener's eyes glaze over by hour three. So you add software that draws a yellow box around anything shaped like a blade and pushes those bags to the front of the line.
The software doesn't decide anything. It doesn't open bags. It just makes sure the human looks at the scary bag first, while they're fresh. That's what most medical AI actually is — a very good yellow box around the suspicious spot on the scan, with a radiologist still making the call. The value isn't that it's smarter than the doctor. It's that it never gets tired at slice 240 of 300.
Key facts
- The FDA publishes an AI-Enabled Medical Device List identifying AI-enabled devices authorized for marketing in the United States; the FDA notes it is not a comprehensive resource and is updated periodically.
- Reporting on the FDA list through December 2025 counted roughly 1,451 authorized AI-enabled devices since tracking began in 1995.
- Radiology accounts for approximately 76% of all authorized AI-enabled devices (about 1,104 of them), and took roughly 75% of 2025 authorizations.
- The overwhelming majority of FDA-authorized medical AI performs narrow image or signal analysis, not open-ended generative conversation.
- Consumer AI chatbots that answer health questions are generally not FDA-authorized medical devices and are not cleared for diagnosis.
- Most authorizations rest on demonstrating comparable performance to a predicate or standard, not on randomized trials showing improved patient outcomes.
▶ The 60-second explainer (script)
Most AI in healthcare isn't a chatbot. It's a yellow box on a scan. Here's the real number. The FDA keeps a public list of AI-enabled medical devices authorized for sale in the U.S. Through the end of 2025, that list held roughly 1,450 devices — and about seventy-six percent of them are radiology tools. Software that flags a possible bleed, measures a nodule, or pushes a stroke case to the top of the reading queue. Why radiology? Because medicine has a scaling problem. A single CT scan can be hundreds of slices. The number of images explodes; the number of radiologists doesn't. Pattern recognition scales. Humans get tired at slice two hundred forty. Think of an airport bag scanner drawing a box around anything blade-shaped and moving that bag to the front of the line. It doesn't decide. It doesn't open the bag. It makes sure the human looks at the scary one first. Beyond imaging, the big uses are ambient note-taking — AI drafting the visit note so doctors aren't typing all night — plus triage and drug discovery. But here's the honest part. Most of these were approved by showing they perform comparably to an existing standard, not by proving patients live longer. Performance can drift between hospitals. And none of this replaces your doctor.
What authoritative sources say
People also ask
Can AI diagnose me?
No consumer AI chatbot is FDA-authorized to diagnose. FDA-cleared medical AI operates inside clinical workflows with a licensed clinician making the final call, under a defined intended use.
Is AI better than doctors at reading scans?
On narrow, well-defined tasks some algorithms match or exceed average human performance in studies. Generalizing that to 'better than doctors' is unsupported — real patients present outside the tested distribution, and outcome evidence is thin.
Is my doctor using AI in my visit?
Possibly for the note. Ambient documentation tools that transcribe and draft visit notes are widely adopted. Practices should disclose recording; you can ask.
What are the main risks?
Performance drift when a model meets a different scanner or population than it trained on, automation bias where clinicians over-trust flags, and gaps in demographic representation in training data.
Does FDA approval mean it improves outcomes?
Usually not directly. Most authorizations show comparable performance to an existing standard. Evidence that a device makes patients healthier generally requires separate clinical study.