AI in Medicine: What Has Worked and What Hasn’t

Artificial intelligence in medicine is often described in sweeping terms: machines that diagnose better than doctors, or that will replace them. The reality is narrower and more interesting. AI tools are already used in hospitals and clinics every day, mostly in specific, well-defined tasks, and some have strong evidence behind them. Others have performed worse in real hospitals than in the lab. Looking at specific examples, both successes and disappointments, gives a clearer picture than the hype.

Where AI is already in use

The US Food and Drug Administration maintains a public list of AI-enabled medical devices it has authorized, which passed 1,000 entries in 2025. The large majority are in radiology, where software analyzes X-rays, CT scans, MRIs and mammograms. Other common areas include cardiology, such as analyzing heart rhythms and ultrasound images, and neurology, such as flagging possible strokes on brain scans so specialists see them sooner.

Case study: diabetic eye screening

In 2018, the FDA authorized the first autonomous AI diagnostic system, IDx-DR, which detects more than mild diabetic retinopathy from photographs of the back of the eye without a specialist interpreting the image. Diabetic retinopathy is a leading cause of vision loss, and many people with diabetes skip yearly eye exams. Placing a camera and AI system in a primary care office means patients can be screened during a routine visit and referred to an eye specialist only if needed.

Case study: breast cancer screening

In Sweden's MASAI trial, published in The Lancet Oncology in 2023, more than 80,000 women were randomly assigned to standard mammogram reading by two radiologists or to AI-supported screening, in which AI triaged images and highlighted suspicious areas. The AI-supported approach detected about 20 percent more cancers while reducing radiologists' screen-reading workload by roughly 44 percent, without increasing false positives. Longer follow-up is still needed to see whether it reduces cancers found between screenings, but it is one of the strongest pieces of evidence so far for AI in a real screening program.

Case study: protein structures and drug research

DeepMind's AlphaFold system predicts the three-dimensional shape of proteins from their genetic sequence, a problem that had challenged scientists for decades. Its database now covers hundreds of millions of predicted structures, and researchers use it to study diseases and design drugs. Demis Hassabis and John Jumper shared the 2024 Nobel Prize in Chemistry for the work, alongside David Baker for protein design. Predictions still need laboratory confirmation, and faster drug discovery does not yet mean drugs reach patients faster.

Case study: paperwork

Some of the fastest-growing uses are administrative. So-called ambient AI scribes listen, with patient consent, to consultations and draft clinical notes for doctors to review and edit. Early reports suggest they can reduce time spent on documentation, a major contributor to physician burnout, though accuracy, privacy and the need for careful review remain concerns. AI is also used for scheduling, coding and processing insurance paperwork.

Case study: when an algorithm disappointed

Not every tool lives up to its promise. A widely used commercial model designed to predict sepsis, a life-threatening reaction to infection, was evaluated by researchers at the University of Michigan in a study published in JAMA Internal Medicine in 2021. They found it missed many sepsis cases and generated frequent alerts that were not useful, performing considerably worse than its developer had reported. The study became a frequently cited reminder that AI tools need independent testing in the settings where they will be used.

Case study: bias

A 2019 study in Science found that an algorithm used by US health systems to identify patients for extra care management systematically underestimated the needs of Black patients. The algorithm used past health spending as a proxy for health needs, and because less money had historically been spent on Black patients with the same level of illness, it ranked them as healthier than they were. After the researchers worked with the developer, the bias was substantially reduced. The case shows how AI can absorb inequities present in the data it learns from.

What AI does well, and what it doesn't

• Strengths: spotting patterns in large volumes of images and data, working consistently without fatigue, and flagging urgent cases for faster attention.
• Weaknesses: performance can drop when used on different patient populations, scanners or hospitals than those it was trained on, and many systems cannot easily explain their conclusions.
• Accountability: when AI contributes to a mistake, questions of responsibility between doctors, hospitals and developers are still being worked out.

Chatbots and patients

General-purpose AI chatbots are increasingly used by patients to look up symptoms and understand test results. They can explain medical terms clearly, but they can also give confident, wrong answers and do not know a person's full medical history. They are best used to prepare questions for a doctor, not to replace one, and personal health details shared with them may not be protected in the same way as medical records.

What to expect

For most patients, AI will mostly be invisible: a second reader on a scan, an alert in a hospital system or a draft note a doctor edits. The most useful tools are likely to be those that support clinicians on specific tasks and are tested in real-world settings, rather than systems that promise to replace medical judgment.