AI and robots in medicine: what's real and what's hype
Nov. 17, 2026 · The Science Bits
A computer program may be helping with your care before you ever see the doctor. It might double-check your mammogram, type up your visit notes or help Medicare review a procedure. Some of this is a real help. Some of it is oversold. And a lot of it happens without patients knowing. I went through the studies, government records and news reports to sort it out in plain English.
A quick word on terms. The AI in this story is mostly what experts call machine learning: software that learns to spot patterns by studying lots of examples, like thousands of past X-rays, instead of following a fixed list of rules. An algorithm is the set of steps the software follows.
1. Where AI is already in the clinic
Reading scans. The U.S. Food and Drug Administration (FDA), the agency that decides which medical devices can be sold, says it has allowed more than 1,600 AI-enabled medical devices on the market.1 About three out of four are for radiology, the field that reads X-rays, CT scans, MRIs and mammograms.2 (A radiologist is the doctor who reads those images.) The FDA notes that its list isn't a complete count, and being on the list doesn't mean a hospital near you uses the product.
The strongest evidence comes from breast cancer screening. In Sweden, about 106,000 women were split by chance into two groups. Researchers do this so the comparison is fair. One group got the usual setup, where two radiologists read every mammogram. In the other, AI flagged spots worth a closer look and helped decide which mammograms needed one reader and which needed two.3
The AI group's radiologists caught more cancers at screening (about 81% of cancers, compared with about 74%), with the same low false-alarm rate in both groups. (A false alarm, or false positive, is when a test says something may be wrong and it isn't.) They also had 44% fewer mammograms to read.34 The researchers also counted cancers that turned up between scheduled screenings, which can be a sign something was missed. The AI group did no worse there: about 1.6 per 1,000 women, compared with about 1.8. That small gap could be chance, so the study doesn't prove AI cut those cases.35 Limits: it was one country, one brand of mammogram machine and one AI product.
AI is also used in stroke care. Some hospitals run software that checks brain scans for a blocked blood vessel and alerts the stroke team right away. In a test at four hospitals, it shortened the time from a patient's arrival to the start of the clot-removal procedure by about 11 minutes.6 With a stroke, every minute counts. The study didn't show that patients recovered better 90 days later, though.
Typing up your visit. One fast-growing use is paperwork. An AI scribe is a note-taking app that listens during your visit and writes a draft of the doctor's notes. The doctor is supposed to read and fix the draft before it goes into your chart.
Doctors are using AI a lot more than they used to. In an American Medical Association survey early this year, 72% of physicians said they used at least one AI tool on the job, up from 38% in 2023. (You may see 81% in headlines. That bigger number also counts doctors who weren't sure which AI tools their practice uses.) The most common use was summarizing medical research, followed by paperwork.7 In one study where doctors were randomly given a scribe or not, the scribe modestly reduced how worn out they felt.8
The drafts aren't always right. A hallucination is when AI adds something nobody said. Ontario, Canada, tested 20 approved scribe products using two pretend doctor visits. The province's auditor general (its official government watchdog) reported in May that every product had at least one problem, and nine made up details.9 In a review of 356 real AI-drafted notes, about 18% left something out and about 11.5% included something made up, though most had no serious error.10 A study at UCLA recorded one mild safety problem: advice the doctor gave was missing from the patient's plan.11 A good sign at your own visit: the doctor tells you the app is on and says they'll check the notes before they go into your chart.
If you live in Texas
Two state laws now cover this. Since September 2025, health care practitioners who use AI to help with a diagnosis, including AI suggestions about diagnosis or treatment based on your record, must tell you and must review anything the AI adds to your record.1214 Since Jan. 1, 2026, providers must also tell you when you're interacting with AI as part of a health service or treatment.1314 These laws don't promise you'll hear about every behind-the-scenes use.
A warning from the past. Hospital AI doesn't always work as well as advertised. In 2021, University of Michigan researchers tested a widely used program that warns staff when a patient might be developing sepsis, a life-threatening reaction to an infection. In their test, it missed about two out of three cases and set off a lot of alerts.15 The maker, Epic, questioned how the study was done.16 A 2026 test of a newer version at four health systems found it was better at telling sick patients from others, but results varied a lot from hospital to hospital and many alerts were false alarms.17 The lesson: these tools need checking by people other than the ones selling them.
2. Robots in the operating room
When people hear "robot surgery," many picture a machine operating on its own. That's not what happens in hospitals today.
The most common system, the da Vinci robot, was used in about 3.15 million operations worldwide in 2025, according to its maker, and about 11,100 were installed by the end of that year.18 The surgeon controls it from a console, looking into a 3D viewer and working hand controls. The robot turns those movements into small, steady motions of thin tools inside the patient. Think of it as a very precise set of remote-controlled hands. It doesn't make any decisions on its own.
Is it better? It depends on the operation. Both robotic and regular laparoscopic surgery are done through small cuts. The difference is whether the surgeon holds the long, thin tools directly or controls them from the console. For gallbladder removal, a review of the research found most results were about the same, but the robot took longer and cost more.19 A 2026 review of weight-loss surgery also found higher costs and no clear improvement in results.20 For some complicated operations, surgeons say the robot gives them a better view and more control, but that isn't the same as proof that patients do better.21
The "robot surgeon" headline. In July 2025, Johns Hopkins researchers reported that a robot called SRT-H handled a long stretch of a gallbladder operation by itself: finding, clipping and cutting the duct and artery, a series of 17 steps. It did this eight times on pig gallbladders outside the body, after training on videos of surgeons, and a person still reloaded clips and changed tools.2223 It's a real milestone, but it hasn't been tested on living patients.
3. Medicine at home: chatbots and wearables
Health chatbots. A chatbot is a program you type or talk to. Tools like ChatGPT run on AI trained on huge amounts of text, which is why they sound so natural. Sounding confident isn't the same as being right, though. OpenAI says more than 300 million people a week ask ChatGPT health questions.24 It started ChatGPT Health as a limited test in January 2026 and opened it to U.S. adults in July. The feature can connect to your medical records and Apple Health.2425 Anthropic (the company that makes Claude, which helped research this story) and Google offer health tools too.2627
How good are they? Two big studies published this year give a mixed picture:
- Oxford University (Nature Medicine, February 2026). About 1,300 people in Britain worked through made-up medical situations, from a cold to bleeding around the brain. Tested alone, the chatbots named a fitting condition about 95% of the time, but picked the right next step only about 56% of the time. When real people used the chatbots, results dropped: they named the right condition only about a third of the time and chose the right next step less than half the time. People who used their usual methods, mostly internet searches, did better at naming the condition.2829 The trouble went both ways: people left out key details, and the chatbots sometimes misread information or gave mixed advice.30
- Mount Sinai in New York (Nature Medicine, February 2026). Researchers ran 60 doctor-written cases through an early version of ChatGPT Health, 16 ways each. It handled obvious emergencies like a stroke or a severe allergic reaction well. But in about half of the true emergency tests (33 of 64), it said the person could see a doctor in a day or two instead of going to the ER. It also went too far the other way: for about 65% of minor problems that could be handled at home, it recommended more care than needed, usually an office visit.31 OpenAI said the test didn't reflect how people normally use the tool and that newer versions do better.32 That claim hasn't been checked independently yet.
Doctors are split. Dr. Robert Wachter of UC San Francisco encourages patients to use these tools. Dr. Adam Rodman, a Harvard Medical School researcher, says they can help you get more out of a real doctor's visit, but he warns against using them to decide whether something is an emergency.32
Your privacy. Here's something many people don't realize. HIPAA, the federal law that protects your medical records, covers most doctors, hospitals and health plans and the companies working for them. An app you choose to use on your own generally isn't covered.33 Your original records stay protected, but the copy you upload is covered mainly by the company's policies and some federal and state consumer rules.34 OpenAI says connected health records aren't used to train its AI or target ads,24 but that's a company promise, not HIPAA.
Wearables. A wearable is a gadget you wear that tracks your body, like a smartwatch. In September 2025, the FDA cleared an Apple Watch feature that warns of possible hypertension (high blood pressure).35 It doesn't measure blood pressure the way a cuff does. It looks for patterns in its light sensor's readings over 30 days.36 In Apple's own testing data, it alerted about 41% of people who had undiagnosed high blood pressure and missed about 59%. Among people who didn't have high blood pressure, about 8% got a false alert.37 Think of it like a smoke detector, not a firefighter. A warning means check your blood pressure with a good upper-arm cuff or see your doctor. No warning doesn't mean you're fine.
Also worth knowing: in January 2026, the FDA issued new guidance easing oversight of some low-risk wellness features and some decision-support software used by doctors.3839 It wasn't a free pass for every gadget, but it means some products that track or interpret health data won't go through FDA review.
4. The frontier
These are things you'll see in headlines that are still mostly in labs and small trials.
- Brain implants. Implants that let paralyzed people move a computer cursor or type with their thoughts really work, but only for a small number of people in research studies. Neuralink, for one, said in January 2026 that 21 people were in its trials.4041 I covered this in depth in Cyborg brains: how close are brain implants, really?
- Estimating your future health risks. European researchers built an AI called Delphi-2M that estimates a person's chances of developing more than 1,000 diseases. It learned from about 400,000 people in the U.K., including their health records and lifestyle information, and was checked against records for about 1.9 million people in Denmark. The researchers say it carries the biases of that data and isn't ready for medical decisions.4243
- Robots that operate on their own. See the Johns Hopkins work above.
- AI and new drugs. Companies are testing AI to find possible new medicines. Whether it makes the whole process faster is still an open question. That's a big enough topic for its own Science Bits piece.
5. Hype vs. reality
Here are claims you'll hear and how they hold up.
- "AI is better than doctors at reading scans." Not quite. The best evidence, from breast cancer screening, shows radiologists using AI caught more cancers than radiologists without it. It's teamwork, not a replacement.3
- "Robots are doing surgery now." Robots help with millions of operations, but in hospitals the surgeon controls them.18 A robot doing part of an operation by itself has only been shown in the lab.22
- "AI is fixing doctor burnout." It's helping a little. Scribes modestly eased doctors' exhaustion in a trial,8 but their notes can include mistakes someone has to catch.910
- "You can ask a chatbot instead of seeing a doctor." Not proven, and risky in an emergency. In testing, people using chatbots did no better than people using their usual methods,28 and one health chatbot told people they could wait in about half of the emergency tests.31
- "Your smartwatch will catch your health problems." Sometimes. The Apple Watch blood pressure alert misses more people than it catches.37
- "FDA-cleared means proven to help patients." Not exactly. About 96% of AI devices on the FDA's list came through what's called the 510(k) route.44 The FDA checks that the device is about as safe and effective as one already on the market. That can require testing, but it isn't proof that patients end up healthier than with usual care.45
One more to watch: AI in Medicare approvals. Since January 2026, traditional Medicare has been running a six-year pilot program called WISeR in six states, including Texas. Private companies use AI, with licensed clinicians making the final call, to review certain services before Medicare pays. These include some spinal steroid shots, nerve stimulator implants (devices that send small electrical pulses to nerves) and knee scopes for arthritis. Doctors can ask for approval ahead of time, or do the procedure and have the claim reviewed before payment.46 Researchers who looked at the first six months found fewer of those services were being done in the pilot states than elsewhere: about 15% fewer of the targeted back injections and almost 20% fewer of the other services. They counted services performed. They didn't measure denials, whether the care was needed, or how much the AI itself had to do with it.4748 Medicare advocates and some lawmakers want the program stopped, and as of mid-November it's still running.4649 If you're on traditional Medicare in Texas and a procedure is held up, ask the office whether WISeR is involved. If a claim is denied, you can appeal.50
The "use it or lose it" worry. A 2025 study at four clinics in Poland looked at colonoscopies done without AI before and after the clinics started using an AI helper. Afterward, doctors found at least one adenoma (a type of growth in the colon that can sometimes turn into cancer) in about 22% of those exams, down from about 28%.51 The study can't prove AI caused the drop, and the clinics were doing about 74% more colonoscopies in the later period, which may also explain it.52 Other studies show AI helps doctors find more of these growths while it's switched on,53 but experts still don't know whether that prevents more cancers.54 It's still a fair question to ask: what happens when the computer is down?
6. What it means for you
Questions for your next appointment
- "Are you using AI for my care today? What for?" In Texas, the law requires disclosure in many cases.1213
- "Is this visit being recorded? Who can see the recording, and can I ask you not to use it?"
- "If an app writes my notes, do you check them?" Then read your notes in your patient portal. If something's wrong, you can ask for a correction. The office doesn't have to agree, but it has to answer you.55
- "Did a person review my scan, or just the software?"
- "My watch flagged something. What should we check?" Bring the alert, but let a real test decide.
- "Was this request reviewed by a computer? What are my options?" If an advance request isn't approved, the office may be able to resubmit it. If a claim is denied, you can appeal.50
When you use a health chatbot at home
- Use it to explain words in your lab results or to make a list of questions for your doctor.
- Share only what it needs. Leave out your name, birth date and ID numbers.
- Double-check anything important with your care team before you act on it.
- Think twice before uploading your medical records, since HIPAA protection doesn't follow the copy.33
- Don't use it to decide whether something is an emergency. For chest pain, trouble breathing, signs of a stroke, or anything that feels seriously wrong, call 911.
This is not medical advice. I'm a tech writer, not a doctor. Talk to your own doctor about your health.
My take: After digging through all of this, I'm more hopeful than worried, but not by much. The best AI in medicine right now is the boring kind: software helping radiologists catch more breast cancers, or getting the stroke team moving a few minutes sooner. The flashy stuff, like chatbot doctors and robot surgeons, is either not ready or not what it sounds like. What bugs me most is how much of this happens behind the curtain: notes written by an app, a procedure reviewed by an algorithm, a watch that stays quiet when it shouldn't. The fix isn't to fear AI. It's to ask questions. Your doctor is still the one making decisions about your care, so ask how the computer fits in.
How this was made: AI helped research and draft this story. Claude, made by Anthropic, did the research and first draft, and a second AI, Perplexity, fact-checked it against the sources. I reviewed it myself. The wording is original, and the numbers and findings belong to the researchers, agencies and outlets credited below. Disclosure: Anthropic offers its own health tools, which are mentioned in section 3. This story is not medical advice.
References & credits
- U.S. Food and Drug Administration, "Artificial Intelligence-Enabled Medical Devices" (list and overview).
- PubMed Central, "Analysis of the FDA AI-enabled device list" (Nov. 2026).
- Gommers J. et al., The Lancet, via PubMed, "MASAI trial final results" (Jan. 31, 2026).
- Lång K. et al., The Lancet Oncology, "MASAI interim analysis (screen-reading workload)" (2023).
- Radiology, via PubMed Central, "Independent commentary on the MASAI results" (2026).
- JAMA Neurology, via PubMed Central, "Randomized trial of AI stroke detection and team alerts" (2023).
- American Medical Association, "2026 Physician Survey on Augmented Intelligence" (March 2026).
- Afshar M. et al., NEJM AI, "Randomized trial of ambient AI scribes" (2025).
- Office of the Auditor General of Ontario, "Special report on AI, section 4.3" (May 2026).
- JMIR Medical Informatics, "Real-world evaluation of AI-drafted clinical notes" (2026).
- NEJM AI, via PubMed Central, "UCLA randomized trial of AI scribes" (2025).
- Texas Legislature, "SB 1188 (enrolled)".
- Texas Legislature, "HB 149, Texas Responsible Artificial Intelligence Governance Act (enrolled)".
- Texas Medical Association, "Explainer on the AI disclosure laws".
- Wong A. et al., JAMA Internal Medicine, via PubMed Central, "External validation of the Epic Sepsis Model" (2021).
- Healthcare IT News, "Research suggests Epic Sepsis Model lacking predictive power" (2021), including Epic's response.
- PubMed Central, "Multi-site evaluation of Epic Sepsis Model version 2" (2026).
- Intuitive Surgical, SEC filing, "Fourth quarter and full-year 2025 results" (Jan. 2026).
- INAHTA database, "Robot-assisted vs. laparoscopic surgery for cholecystectomy".
- Current Surgery Reports, "Robotic vs. laparoscopic bariatric surgery review" (2026).
- Al-Saeed A.A. and Alghannam G., Journal of Medicine and Life, "Robotic versus laparoscopic surgery" (Dec. 2025).
- Kim J.W. et al., Science Robotics, "SRT-H autonomous surgery study" (July 2025).
- Johns Hopkins University Hub, "Robot performs first realistic surgery without human help" (July 9, 2025).
- OpenAI, "Launching Health in ChatGPT" (July 23, 2026).
- OpenAI, "Introducing ChatGPT Health" (Jan. 7, 2026).
- Anthropic, "Health care and life sciences announcement" (Jan. 2026).
- Google, "Fitbit and Google Health updates" (Oct. 2026).
- Bean A.M. et al., Nature Medicine, "Oxford study of people using chatbots for medical decisions" (Feb. 2026).
- Reuters via The Express Tribune, "Oxford study shows AI no better than other methods for patients seeking medical advice" (Feb. 10, 2026).
- Bangor University, "New study warns of risks in AI chatbots giving medical advice" (Feb. 9, 2026).
- Ramaswamy A. et al., Nature Medicine, "Mount Sinai safety evaluation of ChatGPT Health" (Feb. 23, 2026).
- NPR via KPBS, "ChatGPT might give you bad medical advice, studies warn" (March 11, 2026).
- U.S. Department of Health and Human Services, "HIPAA FAQ on health apps".
- Federal Trade Commission, "Health Breach Notification Rule".
- FDA, "510(k) decision record K250507 (Apple hypertension notification)".
- Apple Support, "Hypertension notifications on Apple Watch".
- JAMA, via PubMed Central, "Analysis of Apple Watch hypertension notifications" (Feb. 2026).
- FDA, "General Wellness: Policy for Low Risk Devices" (Jan. 2026).
- FDA, "Clinical Decision Support Software guidance" (reissued Jan. 2026).
- Neuralink, "Two Years of Telepathy" (Jan. 2026).
- Reuters via GV Wire, "Elon Musk's Neuralink says it has 21 participants enrolled in trials" (Jan. 28, 2026).
- Shmatko A. et al., Nature, "Delphi-2M" (Sept. 2025).
- Medscape, "Coverage of Delphi-2M" (2025).
- Complizen, "FDA has cleared 1,614 AI medical devices" (2026).
- FDA, "Premarket Notification 510(k)".
- Centers for Medicare & Medicaid Services, "WISeR Model" (status: active) and its FAQ.
- JAMA, "WISeR research letter" (Oct. 26, 2026).
- Brown University School of Public Health, "Free summary of the WISeR study" (Nov. 4, 2026).
- Medicare Rights Center, "WISeR denials report" (Sept. 24, 2026).
- Medicare.gov, "Appeals".
- Budzyń K. et al., The Lancet Gastroenterology & Hepatology, via PubMed, "Colonoscopy study" (Aug. 2025).
- Science Media Centre Spain, "Expert reactions to the colonoscopy study" (Aug. 2025).
- BMC Medical Imaging, "Meta-analysis of randomized trials of AI in colonoscopy" (2026).
- American Gastroenterological Association, "Guideline on computer-aided detection in colonoscopy".
- U.S. Department of Health and Human Services, "Your Medical Records".
Images: graphic made by Just Spilling the Bits. No outside images used.
Watch the 60-second version
More to read
- Cyborg brains: how close are brain implants, really, and what does AI have to do with it?
- Biohacking: what's real, what's risky, and what AI is about to change
- Nanobots that heal you from the inside: how close is it really?
- Robots that do your chores: what's real, what's hype, and when you could own one
- I asked Copilot to rank itself. Then I asked every other AI what it thought.
- Black Friday is three weeks away. Here's how to spot a fake online store before you pay.