AI in medicine, what it actually does today -

In the United States, 1,614 AI-enabled medical devices are cleared for sale. Three out of four read radiology images. They don’t diagnose in place of the doctor, they flag. The difference between those two words is all you need to understand where we really stand.

The count is kept by the Food and Drug Administration, the US federal agency that authorizes drugs and medical devices, in a public list updated periodically. It’s worth reading the caveat the agency itself puts at the top: the list is not complete, because devices are identified mostly by the terms used in the authorization documents.

AI in medicine, authorized devices, screening and limitations — Most of the approved AI in medicine looks at images. The rest is smaller than it seems.

How many there are and where they sit

The FDA list of AI-enabled medical devices counted 1,614 entries as of its latest update, on September 4, 2026. The breakdown says more than the total: 1,230 belong to the radiology sector (76%), 166 to cardiovascular and 73 to neurology. One hundred and eighty-one received their decision during 2026.

That 76% is not a bureaucratic detail. It means that when people talk about artificial intelligence in medicine, in the vast majority of cases they mean software that looks at an X-ray, a CT scan or an MRI and points the radiologist to where to look. The diagnosis remains an act of the physician, who signs and answers for it.

In the EU the path is different and no comparable count exists. A medical device with AI goes through the medical devices regulation, obtains the CE mark through a notified body and is registered in EUDAMED, the European database anyone can query by device, manufacturer or certificate.

Where the evidence is there

Mammography screening is the case with the best evidence. The MASAI trial, run in Sweden on 105,934 women randomly assigned to two pathways, compared AI-supported reading with traditional double reading. The results, published in The Lancet in January 2026, show a sensitivity of 80.5% against 73.8% for traditional reading, with identical specificity in both groups at 98.5%. The rate of cancers appearing between two screening rounds came out non-inferior, with fewer invasive cases.

The detail that makes that result solid is the unchanged specificity. It means the rise in cancers detected was not paid for with more women recalled needlessly, which is the hidden price of almost every screening tool.

The second documented case concerns the eye. The first system authorized by the FDA to produce a result without a physician reviewing the image analyzes retinal photographs of diabetic patients to detect retinopathy. In the pivotal study on 900 patients enrolled in primary care clinics it reached a sensitivity of 87.2% and a specificity of 90.7%, with 96.1% of images usable. The point is not the number. The exam can be done at the family doctor’s office, where patients already go, instead of at an eye clinic many never reach. Experimental projects such as RetinaMind, which looks for signs of autism and ADHD in the retina, work on the same idea, but they remain research and not authorized tools.

Where it helps, but not through images

On the ward, the most studied tool is early warning for sepsis. The TREWS system, monitored on 590,736 patients in five US hospitals, is associated with an 18.7% relative reduction in in-hospital mortality, but only when the clinician confirms the alert within three hours. The study in Nature Medicine is observational and non-randomized, so it measures the association with staff response, not the effect of the alert itself. Read properly, that research says one precise thing: the value lies in the chain between signal and human reaction, not in the signal.

Then there is the technology no patient notices but which is changing doctors’ days, automatic transcription of the visit. The first randomized trial, published in NEJM AI in December 2025 on 238 outpatient physicians and over 48,000 visits, compared two products with usual care. One cut the time spent on the clinical note by 9.5%, the other produced no significant difference. Burnout and perceived workload improve with both, but those are secondary measures the authors themselves say need confirming, with some clinically relevant inaccuracies reported.

Outside the West the race is faster and less debated. In China, hospitals have started adopting clinical models at scale, as we covered in the piece on CARES 4.0 and adoption in Chinese hospitals. There the problem is not the availability of the technology but the independent verification of results, which is exactly the theme of the next section.

Where it breaks, documented

The most instructive story is that of a proprietary sepsis model deployed in hundreds of US hospitals. An external validation on 27,697 patients at Michigan Medicine, published in JAMA Internal Medicine, found a discriminative ability of 0.63 on a scale where 0.5 equals chance, far worse than the developer had claimed. The model failed to identify 1,709 of 2,552 septic patients and generated alerts on 18% of all hospitalized patients.

The other three ways of breaking are just as well documented. The first is drift over time: kidney injury models validated over nine years keep the ability to rank patients by risk but get progressively worse at estimating absolute risk, because the population beneath them changes. The second is the wrong target: a commercial algorithm used on millions of people predicted healthcare spending instead of illness. Because less was spent on Black patients, at equal scores those patients were sicker. After correcting the bias, the share of Black patients referred to the care program went from 17.7% to 46.5%.

The third is the most recent and the least discussed. In four Polish endoscopy centers, the adenoma detection rate in colonoscopies performed without AI fell from 28.4% to 22.4% after AI had entered the department, according to a study in Lancet Gastroenterology and Hepatology. Operators used to a second pair of eyes looked worse when that second pair was missing.

Finally there is the cost that doesn’t make headlines. In a Korean randomized trial on low-dose CT lung screening, AI doubled follow-up recommendations, from 7.4% to 15.3%, without any cancer emerging in either group during the observed period.

Limitations to keep in mind

The FDA list is not a census. The agency itself warns it identifies devices from the terms used in authorization documents, so the true number differs from 1,614 and nobody knows by how much.
Authorized does not mean tested like a drug. Most devices go through an equivalence procedure against products already on the market, not through a randomized trial like the Swedish mammography one.
Lab results don’t carry over to the ward. The most widespread sepsis model in American hospitals missed nearly two thirds of cases when validated by third parties.
The best studies cited here measure a population, not you. A sensitivity of 80% across a hundred thousand women says nothing about what happens in a single exam. This article does not replace your doctor’s advice in any way.

The chatbot is not a doctor, the numbers say so

Here lies the biggest gap between perception and data. A randomized trial published in Nature Medicine in February 2026 gave 1,298 ordinary people ten clinical scenarios, asking them to work out what they had and what to do, using either a language model or a source of their choice. The models queried alone recognized the condition in 94.9% of cases. The same people using those same models managed it in less than 34.5% of cases, so no better than those who searched elsewhere.

The reason is mundane and serious at once: non-doctors describe symptoms badly, leave out what matters and misread the answer. The authors say it explicitly, standard tests on the models don’t predict this kind of failure, because they measure the machine and not the pairing of person and machine.

Concrete harm already exists in the literature. A 60-year-old man ended up hospitalized with bromism, a rare neuropsychiatric syndrome, after replacing table salt with sodium bromide bought online, having consulted a chatbot for information. The case is documented in a case report journal of the American College of Physicians. It is a reminder that dietary advice taken from a language model can end up in the emergency room. The World Health Organization, for its part, has published guidance on the use of multimodal models in healthcare that warns of the risk of false, inaccurate or incomplete statements.

The rules, with the real dates

There is a lot of confusion here, so it pays to be precise. The EU AI Act classifies as high-risk the systems that are safety components of products already subject to conformity assessment by a third party. Medical devices fall into that category. But the dates have moved: the European regulation known as the Digital Omnibus, approved in July 2026, postponed the obligations for high-risk systems of this type to August 2, 2028. Anyone writing that from August 2026 hospitals must apply the AI Act to medical devices is oversimplifying.

There is a second misunderstanding that touches patients closely. Article 86 of the AI Act introduces the right to receive clear explanations of the role an automated system played in a decision, but it applies only to a category of systems that does not include medical devices. In practice, the right to understand an AI-assisted report today runs through the device rules, the GDPR and above all the responsibility of the physician who signs.

On the data front, the regulation on the European health data space applies from March 26, 2027, with the parts on secondary use of data postponed to 2029 and some deadlines running to 2031. We wrote about it when it was approved, in the piece on the European data space. In the United States, health privacy law covers hospitals, insurers and their providers, but not an app or a chatbot you use on your own: that data sits outside that protection.

Cases show the problem is real. The UK data protection authority ruled that a London hospital had not complied with the law by giving an AI company the data of about 1.6 million patients to test a kidney alert system, because those patients could not reasonably have expected that use. More recently the Texas attorney general closed with a settlement an investigation into a generative AI company used in four major hospitals, challenging misleading claims about the product’s accuracy. And a US case confirmed that the data of 3,182 Johns Hopkins patients ended up on an unapproved AI tool.

What you can check, in practice

Check whether the device is authorized. In the United States the FDA list gives for each one the submission number, the company and the decision date, downloadable in full. In the EU the public EUDAMED search lets you look up by device or manufacturer.
Ask your doctor what role the software played. The useful question is not whether they used AI, but whether the report was reviewed by a person and what would change without that support.
Distrust apps that declare nothing. A consumer app that promises clinical outcomes without citing an authorization, a study or a registration number is selling wellness, not medicine.
Don’t bring reports and records into a generic chatbot. Outside the healthcare system that data loses its sector-specific protections, as the cases already before the authorities show.
Use models to prepare for the visit, not to replace it. Writing out the symptom history and the questions to ask works. Asking for a diagnosis, according to the data, doesn’t.

Key Takeaways

Three authorized devices out of four read images. When people talk about AI in medicine they almost always mean radiology, almost never software that decides a treatment.
Mammography screening has the best evidence: more cancers detected without more needless recalls, across over a hundred thousand women. It’s the study model that should be demanded for everything else.
The leap from lab to ward is the breaking point. The most widespread sepsis system in American hospitals missed two thirds of cases when someone outside the company measured it.
The chatbot answers researchers’ questions well and patients’ questions badly: 94.9% accuracy alone, under 34.5% when ordinary people use it.
The AI Act obligations for medical devices kick in on August 2, 2028, not in 2026. In the meantime the device rules and the physician’s signature apply.

Frequently asked questions

1. Can an AI make a diagnosis in place of the doctor?

Short answer. Almost never. Where it happens, the scope stays very narrow. The only known example of a system authorized to give a result without human review concerns diabetic retinopathy, with a standardized photographic exam and a binary question. Everything else flags, measures or sorts by priority. The decision stays with whoever signs.

2. If my report was also read by software, how would I know?

Short answer. Often it isn’t written anywhere. In the EU, the right to an explanation introduced by the AI Act doesn’t cover medical devices. The practical route is to ask the doctor or the facility, who under the device rules must know which tools they use and with what marking.

3. Can I trust an app that analyzes moles from the phone?

Short answer. Only if it declares a verifiable authorization or marking. Even then it is no substitute for a visit. Many such apps remain wellness products, a category that doesn’t require proof of clinical efficacy. The check is done on the public registers cited above.

4. Does AI in hospitals save doctors time?

Short answer. Less than vendors promise. In the first randomized trial on automatic visit transcription, across over 48,000 consultations, one product cut writing time by 9.5% and the other changed nothing measurable. Those are useful results, not revolutionary ones.

5. Can AI make care worse?

Short answer. Yes, in three documented ways. It can err systematically on some patient groups when trained on the wrong target, it can generate needless alerts that lead to extra tests, and it can erode operators’ skill when it’s taken away, as happened to the Polish endoscopists.

6. Will hospitals use teams of artificial doctors?

7. Who is responsible if software gets it wrong?

Short answer. The physician and the facility remain responsible for the clinical decision, while the manufacturer answers for the product under the device rules. The arrival of the AI Act obligations in 2028 will add risk management and surveillance requirements, without moving the signature away from the physician.

The right question to ask

When someone announces AI has beaten doctors at something, one question clears everything up. Did the study measure the software alone or the software in the hands of those who will actually use it?

Almost all the enthusiastic headlines measure the first. The few studies that measure the second, like the one on the 1,298 people with chatbots or the external evaluation of the sepsis model, tell a slower and more useful story, where the technology works when it enters a procedure someone has thought through in full.

Source link