
AI healthcare means software that reads medical scans, flags health risks, drafts clinical notes, answers patient questions, and helps researchers design drugs, with a clinician still making the final call.
Independent tallies of the FDA file count more than 1,600 AI-enabled medical devices authorized for sale in the United States, about three in four of them in radiology. The FDA keeps a public list of AI-enabled medical devices, and the agency warns the list is not comprehensive. This article gives general information, not medical advice for any person.
How AI healthcare reads medical scans
Radiology took to AI first because the work fits the tool: a scan is a grid of pixels, and a model trained on thousands of past scans can flag the odd one out. In Sweden, the MASAI trial tested AI-supported mammography screening in about 106,000 women, with AI scoring each exam and marking suspicious areas.
After two years, AI support found 29 percent more cancers than double reading, cut screen-reading workload by 44 percent, and did not raise false positives. Cancers found between rounds numbered 82 against 93. Lund University reports the final MASAI results, and notes that many regions in Sweden have started to use AI-supported screening.
That trial answers a narrow question well. Scan readers are narrow models, unlike large language models, which predict text rather than read pixels. Trust must be earned deployment by deployment.

Diagnosis beyond the scan
Mayo Clinic built a screening tool that reads a standard ECG, a routine heart trace, and flags a weak heart pump before symptoms show. In a Mayo study it found people at risk of left ventricular dysfunction 93 percent of the time. Mayo holds more than 7 million ECGs, with identifying details removed first, and its AI also runs in Apple Watch.
The same team uses AI on CT scans in emergency rooms, where AI examines scans from people with bleeding in the brain and cuts diagnosis time. AI-guided ECGs also flag atrial fibrillation before symptoms appear.
Nothing here replaces the cardiologist. It moves the first look earlier, when treatment can do more.

How AI helps design new drugs
Drug discovery runs on a brutal failure rate, so drug makers use AI to pick targets and sketch molecules faster. The furthest published test is rentosertib, for idiopathic pulmonary fibrosis, a scarring lung condition. Generative AI picked both the target, a protein called TNIK, and the molecule built to block it.
A randomized, double-blind, placebo-controlled phase 2a trial ran for 12 weeks in 71 patients. Safety was the main goal, and adverse event rates looked similar across all arms. At the highest dose, lung capacity rose by a mean of 98.4 ml while the placebo group fell by 20.3 ml. The trial report in Nature Medicine calls for larger trials of longer duration.
The paper opens bluntly: few AI-discovered drugs have reached human trials. AI can shorten the early hunt, not the years of testing that decide whether a drug works.

Clinical notes and the paperwork load
Ambient scribes listen to a visit and draft the clinical note for the doctor to review and sign. A scribe is one kind of AI agent, software that works in the background while the clinician talks to the patient. Clinicians at Ohio State Wexner Medical Center use Dragon Copilot to draft notes, with the clinician reviewing and finalizing each one.
WHO lists visit documentation among the main health uses of large multimodal models. The pitch is simple: the doctor looks at the patient instead of the keyboard. The catch is the review step, where a smooth draft can still carry invented words.
Nabla, whose tool transcribed about 7 million visits, requires providers to edit and approve notes. The software drafts. The clinician owns the record.

Patient chatbots and hospital operations
WHO lists patient-guided use, such as checking symptoms, among the main health uses of large multimodal models. A chatbot built on a model like GPT-4o can answer at 2 a.m. in plain language. It can also sound sure when it should not: WHO warns these models can produce false, inaccurate, biased, or incomplete statements.
Behind the scenes, Ohio State runs a smart hospital platform at two facilities that connects building, operational, and clinical support systems. Its data project aims to forecast surgical demand and operating room use, so beds and teams match demand.
This work rarely makes headlines. It decides whether clinics run on time and beds are free when needed.

Limits, bias, and safety checks
The strongest warning comes from a study about an older algorithm. Researchers found an algorithm that picked patients for extra care predicted future health costs, not illness. At the same risk score, Black patients were considerably sicker than White patients. Fixing the disparity would have raised the share of Black patients getting extra help from 17.7 percent to 46.5 percent. The study appeared in Science in 2019. The bias lived in the proxy.
An Associated Press investigation found OpenAI Whisper, used in hospital transcription, invents text nobody said. Researchers found 187 hallucinations in over 13,000 clear audio snippets, nearly 40 percent of them harmful or concerning. One invented drug name was “hyperactivated antibiotics”. OpenAI warns against using Whisper in high-risk domains.
Privacy brings its own questions. A California lawmaker refused a form that would have sent her child’s consultation audio to vendors including a cloud provider. Nabla erases the original audio for data safety reasons, so nobody can check a transcript against the recording. Every scribe deployment has to answer where the audio goes and how long it stays.
To judge a claim, check the FDA listing and its exact covered task, then ask for a randomized trial in real clinics, with results checked by age and ethnicity. WHO guidance pushes the same checks in more than 40 recommendations. Ask who reads the output: safe deployments here share one habit, a clinician who reads the draft and signs it. When you compare AI tools, weigh trial evidence above feature lists.

Frequently asked questions about AI in healthcare
What is AI in healthcare?
Software that reads scans, flags risks, drafts notes, answers patient questions, and helps design drugs. A clinician still decides.
Is AI healthcare safe?
It depends on the deployment. Listed devices passed a safety review for a stated use. Scribes and chatbots need a human to check the output.
Can AI replace doctors?
No. The AI does a first pass and a person decides: radiologists read the mammograms and doctors sign the notes. The tools do not carry responsibility for a patient.
Has the FDA approved AI for healthcare?
The FDA authorizes AI-enabled devices, and tallies count more than 1,600 entries, mostly radiology. Check the listing for the exact product and task.
Can AI find new drugs?
AI can help pick targets and design molecules. Rentosertib reached a 71-patient trial with a lung function signal, but it is still in testing, not an approved treatment.
Should I trust a health chatbot with my symptoms?
Use a chatbot for general information, not a personal diagnosis. WHO warns large models can produce false or biased statements. Talk to a clinician about your own health.