Artificial intelligence in medicine is no longer experimental. The US Food and Drug Administration maintains a public list of AI-enabled medical devices it has authorised, and by 30 March 2026 that list had grown to 1,524 devices. Headlines about AI in healthcare often focus on chatbots, but most of these devices do something quite different. Here is what they actually do, where they are used and what we know about how well they work.
What counts as an AI medical device
An AI-enabled medical device is software, sometimes combined with hardware, that uses artificial intelligence or machine learning to help diagnose, monitor or treat patients. Examples include software that analyses a chest X-ray for signs of disease, a tool that flags an irregular heart rhythm in an ECG, or a system that estimates the risk of a patient's condition getting worse.
In the United States, such devices need FDA authorisation before they can be sold for medical use. Most are cleared through a pathway that compares them with similar existing devices, rather than through large new clinical trials.
Radiology dominates
The most striking pattern in the FDA list is how heavily it leans toward medical imaging. According to an analysis of the list published in July 2026, about 76 percent of the devices, around 1,163, are in radiology. When imaging tools in related fields such as cardiology and neurology are included, imaging accounts for close to 80 percent.
Cardiovascular devices make up roughly 10 percent, and neurology about 3 percent. Other specialties, such as pathology, ophthalmology and gastroenterology, account for smaller shares.
Why imaging leads
Radiology suits AI for several reasons. Medical images are digital and stored in standard formats. Hospitals have huge archives of images with diagnoses attached, which can be used to train AI. And image analysis is a pattern recognition task, which is exactly what modern AI does well.
What these devices actually do
Most AI medical devices are narrow tools designed for specific tasks. Common examples include:
Triage and prioritisation. Software that scans incoming CT scans for signs of conditions such as stroke or bleeding in the brain and moves urgent cases to the top of a radiologist's list.
Detection support. Tools that highlight possible abnormalities, such as lung nodules on chest scans or suspicious areas on mammograms, for a doctor to review.
Measurement. Software that automatically measures structures, such as heart chamber size or tumour dimensions, saving time and improving consistency.
Image quality. AI that improves image clarity or reduces scan time, for example by reconstructing MRI images from less data.
Heart rhythm analysis. Algorithms that detect irregular rhythms such as atrial fibrillation in ECG recordings, including from wearable devices.
Eye disease screening. Systems that analyse retinal images to detect diabetic retinopathy.
In almost all cases, these tools assist clinicians rather than replace them. A doctor reviews the AI's output and makes the final decision.
What about generative AI and chatbots?
Despite the excitement around large language models like ChatGPT and Claude, the FDA had not authorised any generative AI-enabled device for marketing as of mid-2026, according to the same analysis. In March 2026, the FDA granted breakthrough device designation, a programme meant to speed up review, to a patient-facing generative AI application. The agency also plans to tag devices based on foundation models separately.
This means that general-purpose AI chatbots used for health questions are not FDA-authorised medical devices. They can be useful for understanding information, but they are not approved for diagnosis or treatment.
The evidence question
A large number of authorised devices does not automatically mean strong evidence that they improve patient care. Research has raised important questions.
- A study published in JAMA Network Open examining hundreds of devices found clinical performance studies reported for only about half.
- In a review of 691 devices, many summaries did not report study design, training data size or demographic representation.
- Very few devices cited randomised controlled trial data, and fewer than 1 percent reported patient health outcomes.
- A 2025 study linked a number of AI devices to recalls, with many occurring within a year of clearance.
None of this means the devices do not work. Many perform well in their specific tasks. But it does mean that hospitals, doctors and regulators should look closely at the evidence for each tool, especially how it performs across different patient groups.
Why this matters for India
Many AI medical tools used in India are developed by Indian companies or adapted from international products. India's medical device regulator, the Central Drugs Standard Control Organisation, regulates software as a medical device under India's Medical Devices Rules.
The FDA list matters because it is one of the most detailed public records of AI medical devices worldwide, and many companies seek FDA clearance as a mark of credibility. But performance in one population does not always transfer to another. An AI trained mostly on data from American or European patients may perform differently on Indian patients, so local validation is important.
What patients should know
- AI is already part of your care in many hospitals, often behind the scenes, helping radiologists read scans faster or flagging urgent cases.
- A doctor should still review AI results and make decisions.
- It is reasonable to ask your doctor whether AI was used in your diagnosis and how its results were checked.
- General AI chatbots are not substitutes for medical advice.
The bottom line
More than 1,500 AI medical devices have FDA authorisation, and most of them help doctors analyse medical images. They triage urgent cases, detect possible problems, measure structures and improve image quality, almost always as assistants to clinicians. The evidence behind many devices is thinner than the numbers suggest, and no generative AI device had been authorised as of mid-2026. AI is already improving parts of medicine, but careful evaluation and human oversight remain essential.
