Few areas generate as much excitement, or as much caution, as AI in medicine. The potential benefits are large: earlier diagnosis, less paperwork for clinicians and faster drug development. The risks are equally serious, because mistakes can harm patients. Here is where things stand, with links to the evidence.
Updated September 2026: we added sources for each claim, the results of a randomized trial of AI scribes, and the revised EU timeline for AI in medical devices.
Where AI is already in use
Medical imaging. This is the most established area. The US Food and Drug Administration keeps a public list of AI-enabled medical devices it has authorized. An analysis of that list counted 1,451 devices by the end of 2025, about three-quarters of them in radiology. The FDA notes that its list is not exhaustive. These tools help flag possible findings on X-rays, CT scans and mammograms, prioritize urgent cases and measure structures more consistently.
Clinical documentation. “Ambient” AI scribes listen to a consultation, with patient consent, and draft clinical notes. The evidence is now more than anecdotal. In a randomized trial published in NEJM AI in late 2025, UCLA Health tested two scribe tools with 238 physicians across 14 specialties. One tool cut note-writing time by about 10%, the other’s reduction was not statistically significant, and both were linked to modest improvements in burnout scores. The researchers also reported that notes occasionally contained clinically significant errors, mostly omissions, and stressed that physicians must review every note.
Drug discovery. AI helps predict protein structures, work recognized by the 2024 Nobel Prize in Chemistry, half of which went to Demis Hassabis and John Jumper of Google DeepMind for AlphaFold’s protein structure prediction. AI is also used to find promising molecules. In 2025, Nature Medicine published a phase 2a trial of rentosertib, a drug for idiopathic pulmonary fibrosis whose target and molecule were found with generative AI. It was a small study of 71 patients, and larger trials remain the real test.
Operations. Scheduling, bed management, billing and coding are less glamorous but significant uses.
Patients and chatbots
Many people now ask general AI assistants health questions. Used carefully, they can help you understand terminology, prepare questions for an appointment or make sense of a diagnosis you have already received. Used carelessly, they can give confident wrong answers, miss warning signs or offer reassurance when urgent care is needed.
The World Health Organization’s guidance on large multimodal models in health warns that such models can produce “false, inaccurate, biased, or incomplete statements” that could harm people making health decisions, and recommends independent audits when they are deployed at scale.
For patients Use AI to understand and prepare, not to diagnose or change treatment. Seek urgent care for serious symptoms, check important information with a clinician, and be cautious about sharing identifying health details. Our privacy guide explains how.
The limits and risks
- Performance can drop in the real world. Tools that do well on test data can perform worse in different hospitals, on different equipment or with different patient populations.
- Bias. If training data underrepresents certain groups, performance can be worse for them, as we explain in our guide to AI bias.
- Automation bias. Clinicians may over-trust a tool’s suggestion, or become less vigilant.
- Errors in generated text. The UCLA trial found omissions and other mistakes in AI-drafted notes, which is why review is not optional.
- Accountability. When an AI-assisted decision goes wrong, responsibility can be unclear.
- Privacy. Health data is among the most sensitive information there is.
How it is regulated
AI that diagnoses or guides treatment is typically regulated as a medical device, which requires evidence of safety and effectiveness. In Europe, AI in medical devices is also classed as high-risk under the EU AI Act. After the 2026 amendments, those AI Act requirements apply from August 2028. General-purpose chatbots answering health questions sit in a grayer zone, which is one reason their makers include disclaimers and direct users toward professional care.
An assistant, not a replacement
AI’s most proven healthcare value today is as an assistant to professionals: catching things, speeding up paperwork and accelerating research, not replacing clinicians. The best evidence so far shows real but modest gains, along with errors that need human review. The promise is real. So is the need for evidence, oversight and humility.
Sources
- Artificial Intelligence-Enabled Medical Devices, US Food and Drug Administration
- FDA AI/ML-Enabled Medical Devices by the Numbers (2026), PharmaDossier
- Ambient AI Scribes in Clinical Practice: A Randomized Trial, NEJM AI
- UCLA study finds AI scribes may reduce documentation time and improve physician well-being, UCLA Health
- The Nobel Prize in Chemistry 2024, NobelPrize.org
- A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trial, Nature Medicine, June 2025
- WHO releases AI ethics and governance guidance for large multi-modal models, World Health Organization, January 2024
