AI in Healthcare: Real Uses, the Evidence and the Risks

In brief

AI in healthcare is software that reads scans, drafts clinic notes, flags risk and helps design drugs, and the FDA's list held 1,614 AI enabled device authorizations at its 4 September 2026 update, 76% of them in radiology. Trials show modest gains: one 238 physician trial found a scribe cut time in notes by 9.5%.

What is AI in healthcare, and how big is it now?

AI in healthcare is software that learns patterns from medical data and then helps with a task: reading an image, drafting a note, predicting a risk or proposing a molecule. It sits inside tools clinicians already use. The FDA keeps a public list of AI enabled medical devices, and its page was current as of 4 September 2026.

A chrome stethoscope looped around a rainbow glass sphere on a black plinth, an upright terracotta clipboard holding one blank sheet, and a terracotta pencil on a third plinth, in a charcoal studio under white rim light.

I counted the rows on the FDA list myself. Decisions dated 2025 number 335, the most of any year. The first half of 2026 already holds 181. The FDA says the list is "not a comprehensive resource of AI-enabled medical devices," and it counts authorizations, not separate products. Treat 1,614 as a floor.

Radiology holds 76 of every 100 AI enabled devices on the FDA list
Radiology1,230 entries76%Cardiovascular154 entries9.5%Neurology73 entries4.5%All other panels157 entries9.7%each square is about 1% of the 1,614 entries on the list

The same page says the FDA will explore ways to tag devices built on foundation models, from large language models to multimodal systems. It gives no count of those yet. That matters, because the chatbots most people mean by AI aren't what the list mostly holds. It mostly holds image tools.

What are examples of AI in healthcare?

The best examples are the ones with a named product and a published result. This table lists eight, each with its source. Where a number comes from a news report of a study rather than the paper, the cell says so.

UseNamed exampleWhat was reportedSource
Breast cancer screeningMASAI trial, SwedenAlmost 106,000 women. 29% more cancers found with AI support (338 against 262 cases). Radiologist workload down 44%Lancet Digital Health, February 2025, as reported by EurekAlert
Eye screeningIDx-DR87% sensitivity and 90% specificity for more than mild diabetic retinopathyBajwa et al., Future Healthcare Journal, 2021
Radiotherapy planningInnerEyePreparation time cut by up to 90% for some cancersBajwa et al., 2021
Clinic notesNabla and Microsoft DAX Copilot238 physicians. Nabla users cut time in notes 9.5% more than usual care. DAX showed no clear changeLukac et al., NEJM AI, 26 November 2025
Drug discoveryRentosertib for lung fibrosisPhase 2a, 71 patients, 12 weeks. Lung function rose 98.4 mL on the 60 mg dose and fell 20.3 mL on placebo, a secondary endpoint in a trial not powered for efficacyNature Medicine, 2025, as reported by trade press
Early warning alertsEpic Sepsis ModelAUC of 0.63 at one academic center, 38,455 hospitalizations. It missed 67% of sepsis cases and scored high on 18% of hospitalizationsWong et al., JAMA Internal Medicine, 2021
Clinician chatChatGPT for CliniciansLaunched 22 April 2026 for verified US cliniciansOpenAI help center release notes
Device oversightFDA AI enabled device list1,614 entries as of 4 September 2026FDA

The mix tells the story. The strongest evidence sits with imaging and note taking, and the weakest with predictive alerts. One row is a failure, and it belongs here on purpose.

The sibling guide to ChatGPT for Clinicians covers the chat row in depth. AI for nurses covers the bedside side of the job.

Is artificial intelligence in healthcare the same as AI in medicine?

Mostly yes. People use artificial intelligence in healthcare for the whole system, including billing, scheduling and records, and AI in medicine for the clinical part, such as diagnosis and treatment. The examples above sit on the clinical side, apart from the scribes, which sit on both. Search results for either phrase show the same tools.

Two more areas show up in every review but have thinner evidence than imaging. Bajwa and colleagues say language tools could automate administrative tasks such as documenting patient visits in electronic health records, and they name remote monitoring through wearables and sensors as a use. Treat both as promising, with the scribe trial below as the one controlled test here. A useful split is by who touches the result. A scan reader changes what a doctor sees. A scribe changes what a doctor types. A drug model changes what a lab tests, years before any patient does. Each carries a different risk, so each needs a different check.

What do the trials actually show?

They show real but modest gains, and they show the gains depend on the setting. Three cases give the range, from a clear win to a clear miss.

Case 1, breast screening. In the Swedish MASAI trial, almost 106,000 women were split between AI supported screening and standard double reading. The Lancet Digital Health reported in February 2025 that AI support found 29% more cancers, 338 against 262, with 44% less radiologist workload and only seven more false alarms. The Lancet followed in January 2026 with interval cancers, those found between screening rounds: 1.55 per 1,000 women with AI against 1.76 without, 12% fewer. Both reports come from one Swedish screening program, so they don't describe every clinic.

Case 2, clinic notes. Lukac and colleagues randomized 238 UCLA outpatient physicians to Microsoft DAX Copilot, Nabla or usual care between November 2024 and January 2025. Nabla cut time in notes by 41 seconds, from 4:30 to 3:49, against 18 seconds in the control group. DAX showed no clear change on that measure. The authors call the burnout findings secondary and in need of confirmation. About 15% of physicians never used their scribe, and the authors say time spent editing drafts isn't captured, so the saving may be overstated.

Case 3, a sepsis alert. Wong and colleagues tested the Epic Sepsis Model on 38,455 hospitalizations at one academic center, Michigan Medicine, and found an AUC of 0.63. An accompanying editorial notes the developer had reported 0.76 to 0.83. The model scored 6 or higher in 18% of hospitalizations and missed 1,709 patients with sepsis, 67% of them. At that threshold its sensitivity was 33%. A model that works at the hospital that built it can fail at yours, which is why outside checks matter.

Drug discovery sits between them. In the 12 week phase 2a trial of rentosertib in lung fibrosis, the sponsor and trade press report a gain of 98.4 mL in forced vital capacity at the 60 mg dose and a fall of 20.3 mL on placebo, in 71 patients, all in China. The authors list small groups and short follow up as limits, and the gain was a secondary endpoint. That's early evidence, not proof of benefit.

What are the pros and cons of AI in healthcare?

The pros are time and detection, and the cons are models that don't travel, errors in drafts and poor data. Each has a source. Bias in training data is a concern the review literature raises widely, and AI ethics covers it.

  1. Pro, time back: the scribe trial found a 9.5% cut in note time for one product, and no clear change for the other.
  2. Pro, more cancers found: MASAI found 29% more cancers with AI support, with a 44% lighter reading load.
  3. Pro, screening where specialists are few: IDx-DR reported 87% sensitivity and 90% specificity for more than mild diabetic retinopathy.
  4. Con, models that don't travel: the Epic sepsis result above is the standard warning.
  5. Con, errors in drafts: physicians in the scribe trial rated clinically significant inaccuracies as occurring occasionally, 2.7 for DAX and 2.8 for Nabla on a five point scale, and one mild adverse event involved omitted patient counseling.
  6. Con, messy data: Bajwa and colleagues write that healthcare data is "highly sensitive, inconsistent, siloed and not optimised for the purposes of machine learning."

The pattern is consistent. AI helps most where the task is narrow and the answer can be checked, as with a scan or a note. It fails most where the task is broad and the stakes hide in the data.

How is AI in healthcare regulated?

In the United States, the FDA reviews AI enabled devices as medical devices, and it publishes the list of those it has authorized. Its AI section also carries guiding principles on good machine learning practice, predetermined change control plans and transparency for machine learning enabled devices.

Most entries use the fastest route. The FDA list shows submission numbers beginning with K, which mark 510(k) clearances, and a few beginning with P, which mark the stricter premarket approval route. The earliest entries date from 1995, so this isn't new. The pace is: 65 decisions dated 2018, 335 dated 2025.

A listing isn't a guarantee. The FDA says the list is not comprehensive, and the Epic example shows that a deployed model can underperform later. Chat tools that answer health questions mostly sit outside this process. Best AI for medical questions compares them, and how to fact check AI answers gives the routine for any claim.

What questions should you ask an AI about this topic?

Use an AI to learn the field, not to decide care. These six prompts work in Claude, ChatGPT, Gemini or Whizi, and each asks for sources you can open. Replace the brackets with your own details.

  1. "Explain how [a use, such as AI mammography screening] works for a reader with no medical training. Then list the two studies a skeptic would cite and a link to each."
  1. "I'm a [nurse, administrator or student]. Which three uses of AI in healthcare matter most for my job, and what does the published evidence say about each? Mark any claim you couldn't source."
  1. "Compare [two tools] for [task]. Use only each vendor's own page and the FDA list. Make a table, and write not stated where you found nothing."
  1. "Here is a news claim about AI in medicine: [paste]. Find the original study, tell me the sample size and setting, and say what the claim leaves out."
  1. "What are the strongest arguments that [a tool] is overhyped? Give each with a source, then say which are well supported."
  1. "Draft ten questions I should ask my hospital about how it uses AI with patient data. Keep each under 20 words."

These prompts help you study and prepare. They don't replace a clinician, and they can't see your records. If you paste patient details into any chatbot, check the tool's privacy terms first.

Where does Whizi fit?

Whizi is one app for asking Claude, GPT and Gemini the same question, which suits research like the prompts above. Web search is on by default in each chat, and Pro adds side by side Compare, so you can read where two models disagree on a study.

The free plan gives 10 messages a day on Whizi AI, Auto and Claude Haiku, with no card. Pro carries all three labs, and its first week is $1.99. Whizi isn't a medical device and doesn't give medical advice.

What can AI in healthcare not do?

It can't diagnose you, replace your clinician or promise a result. The evidence above covers narrow tasks under study conditions. A trial in one Swedish screening program doesn't describe every clinic, and the sepsis result shows how far performance can drop elsewhere.

This page gives no medical advice. For symptoms, medicines or a diagnosis, talk to a licensed clinician. Every figure here comes from the source named beside it, and where a figure comes from a report about a study, the text says so.

Workflow checklist
  • Find the journal, year and sample size behind any claim
  • Check whether the study ran in one hospital or several
  • Ask how big the gain was and against what
  • Search the tool name with external validation
  • Look the product up on the FDA list, and read a listing as authorization only
Common questions

Frequently asked questions

What are examples of AI in healthcare?

AI mammography screening, diabetic eye screening, note taking scribes, radiotherapy planning and drug discovery are examples with published results. The FDA list held 1,614 AI enabled device authorizations on 4 September 2026, and 1,230 were in radiology.

What are the pros and cons of AI in healthcare?

The main pros are time and detection: one trial cut note time by 9.5% and another found 29% more breast cancers with AI support. The main cons are models that fail outside their home hospital, drafts with errors, and messy source data.

Will AI replace doctors?

The sources here don't show that. Every trial above used AI beside a human reader or writer. This page makes no prediction of its own.

Is AI in healthcare regulated?

In the United States, yes: the FDA authorizes AI enabled devices and lists them. Its list says it isn't comprehensive, and general chatbots that answer health questions mostly sit outside that review.