Probably. Usually quietly. This page walks through where artificial intelligence already shows up in your care, what it actually does, and the one procedure where the evidence is strongest: the colonoscopy.
Only 16% of U.S. adults know AI has been used in their own care,[9] yet the FDA has authorized well over a thousand AI-enabled medical devices,[11] and most large hospital systems run prediction models on every admission. It hides in three places.
Language models read clinical notes to pull out lab values, symptoms, and trial eligibility. At Mass General Brigham, a tool called RECTIFIER screens charts for clinical trials, catches urgent findings buried in notes with up to 98% accuracy, and nearly doubled enrollment in a randomized trial of about 4,500 patients.[12]
About three in four FDA-authorized AI devices are radiology tools.[11] Convolutional neural networks flag nodules, bleeds, and fractures. Pathology AI reads biopsy slides, detecting colorectal cancer with an AUC of 0.988 against 0.970 for expert pathologists.
Real-time detection draws boxes around polyps during a colonoscopy, and computer vision tracks instruments and anatomy in robot-assisted surgery. This is the newest and least invisible kind of medical AI, and the focus of this page.
Pick a specialty. Each one says how likely you are to have met AI there, what it does, one real example, and where to read more.
Tick what applies. The answer updates as you go. Nothing is stored or sent anywhere.
Nothing checked yet
Even with nothing ticked, the back office that scheduled your last visit and coded the bill almost certainly used some machine learning.
Colonoscopy is the best test case for medical AI. It is live video, it depends entirely on one person's eyes, and the thing being looked for, a precancerous polyp, is small, pale, and easy to hide behind a fold.
Colorectal cancer is one of the leading causes of cancer death worldwide, and it is largely preventable: find the polyp, remove it, and the cancer never forms. The score every endoscopist is judged by is the adenoma detection rate, or ADR: the share of screening colonoscopies in which at least one precancerous adenoma is found.
In a landmark 2014 study in the New England Journal of Medicine, each 1-point rise in a doctor's ADR cut their patients' risk of an interval cancer, one that appears between screenings, by about 3%.[1] The professional benchmark is 25%. Elite endoscopists reach 35 to 50% by washing the lining, flattening folds, and withdrawing the scope slowly for six to nine minutes.
Yet in tandem studies, where a second colonoscopy immediately follows the first, 25 to 35% of adenomas turn out to have been missed, usually because they were flat, tiny, hidden behind a fold, or seen late on a long day.[4]
Rough illustration: it compounds the roughly 3%-per-point hazard ratio from Corley et al.[1] It is not a prediction for any one patient.
Computer-aided detection, or CADe, is a convolutional neural network that watches the colonoscope's video feed frame by frame and draws a box over anything that looks like a polyp, in under 40 milliseconds, using object-detection designs from the YOLO family (“You Only Look Once”).
The first system to clear the U.S. FDA was GI Genius, in April 2021, through the De Novo pathway (DEN200055).[2] The FDA's conditions matter: it is an adjunct. It cannot diagnose, it cannot replace a biopsy, and the doctor still decides what to remove.
Flip the switch to see what changes. The round polyp is easy. The flat, pale patch at the upper left is a sessile serrated lesion, the kind human eyes miss most, and the kind AI helps most with. In one trial, right-sided serrated lesion detection doubled, from 12.3% to 24.8%, when AI was combined with a fold-flattening cuff.
Randomized trials have now compared the same doctors with and without the box, including a meta-analysis of 44 trials and 36,201 patients.[3] The pattern is consistent: more adenomas found, especially the small and flat ones, with no spike in needless biopsies.
A Danish multicenter trial of 795 patients pushed the screening-population ADR from 58.1% to 74.4%.[5] The gains come disproportionately from 1 to 5 mm adenomas and flat lesions, exactly the ones that tandem studies show humans miss.
Research from Dr. Tyler Berzin's group at Beth Israel Deaconess and Harvard showed that even small public video datasets, stretched with augmentation, can train a network to 0.73 sensitivity and 0.96 specificity, catching serrated adenomas in 93% of positive images.[23]
| Measure | Without AI | With AI |
|---|---|---|
| Adenoma miss rateCut by more than half when AI is on [3][4] | 35.3% | 16.1% |
| Adenomas found per colonoscopyPatients with several lesions get fully cleared in one visit | 1.11 | 1.35 |
| Harmless tissue removed by mistakeNo significant change (p = 0.54). The extra finds are real adenomas, not junk | 17.1% | 15.1% |
In the United States, procedures are billed with CPT codes, and there is no code for AI-assisted detection. A hospital eats roughly $10 per procedure in licensing, while the savings, cancers that never need treating, land with insurers years later.
Health economists test this with Markov models that follow millions of simulated patients for life and count cost per QALY. The verdict is that AI colonoscopy is a dominant strategy: better outcomes and lower total cost, in up to 73.4% of the scenarios tested.[8][24]
The lesson is bigger than one device. The affordability crisis cannot be fixed by shaving upfront costs. It needs payment that rewards preventing disease rather than performing procedures.
| Line item | Without AI | With AI |
|---|---|---|
| Screening colonoscopy | €101.25 | €111.75 |
| Polyp removal and follow-up | €150.32 | €153.85 |
| Late-stage cancer care | higher | €1,073.80 |
| Net lifetime savings with AI | €4.91 to €14.34 |
Per cohort the model prevented 134 colorectal cancers and 56 deaths: 7.4% fewer cases and 7.6% fewer deaths, plus 0.027 extra quality-adjusted life years per person.[8]
Detection has a cousin: computer-aided diagnosis, or CADx, which tries to tell a precancerous adenoma from a harmless hyperplastic polyp on screen, in real time. If it could reach a 90% negative predictive value for polyps 5 mm and under, the threshold the American Society for Gastrointestinal Endoscopy set in its PIVI statement, doctors could resect and discard tiny polyps without sending them to pathology, saving an estimated $1 billion a year in the U.S.[7]
In multicenter trials CADx accuracy has hovered around 78 to 81%, short of the bar. And one microsimulation found a twist: human-only resect-and-discard cuts the carbon footprint of shipping and processing specimens, but running CADx servers emitted more CO₂ than the pathology it replaced.
So detection is becoming the standard of care, while diagnosis is still an experiment. A tool can be clinically fascinating and still fail on economics, uncertainty, and liability.
Putting an algorithm in the room changes the humans in it. These are the problems researchers have actually measured.
The networks are tuned to miss nothing, so they also flag folds, bubbles, fluid, and the scope's own shadow. Mucosal folds alone caused 39.4% of false alerts in one commercial system and 33.9% in another. Depending on the system, false boxes appear anywhere from under 1 to over 37 times per minute.
Doctors drowning in alerts start ignoring them. Older medication-safety alerts were overridden 49 to 96% of the time, and a box nobody trusts is a distraction, not an aid.
The fix: tune thresholds, tier the alerts, and go quiet while a polyp is already being removed.
In 2025, researchers followed 19 experienced endoscopists in Poland, each with over 2,000 procedures, after months of routine AI use, then measured them with the AI switched off.[6]
Surgeons dropped 8.3 points, and some high-performing centers fell by more than 12. The authors called it the Google Maps effect: hand a task to reliable automation and people stop looking as hard.
The fix: track unassisted performance, train without the overlay, keep the human deciding.
No CPT code means no reimbursement. The hospital pays for the license; the insurer collects the savings. Until payment rewards prevention over volume, adoption stalls even when every model says society wins.[8]
The fix: value-based payment that rewards fewer late-stage cancers, not more procedures.
Hospital IT departments rolled these tools out faster than anyone wrote a consent form. Only 16% of adults know AI was used in their care; 46% have no idea.[9] A generic line buried in a ten-page portal agreement satisfies the letter of transparency and none of its intent.
The fix: say it out loud, in the room, and let people opt out of the parts they can.
Pew Research Center surveyed U.S. adults in June 2026.[9] The gap between how much AI is used and how much people know about it is the whole reason this page exists.
Here is the paradox. The same public that distrusts AI in the exam room uses it at home: 34% of adults ask chatbots health questions,[9] 28% to get information fast, 25% to figure out symptoms, and 18% to raise things they feel too uncomfortable to tell a human. People do not fear the tool. They fear not being told, and not having a say.
None of these are rude. Most doctors will be glad you asked.
Do you use any AI tools when you look at my scans, slides, or scope video?Radiology, pathology, and endoscopy are where it is most common.
Is this visit being recorded or transcribed by an AI scribe? Can I opt out?Ambient scribes are the fastest-spreading AI in primary care. 72% of people want to be told.[9]
If the AI flags something, who makes the final call?Every FDA authorization so far is for an adjunct. The answer should be a person.
What is your adenoma detection rate?For a colonoscopy, this is the expert question. Benchmark is 25%;[1] 35% and up is excellent. Many endoscopists know their number.
Was this tool tested on patients like me?Models trained on one population can quietly underperform on another.
Does the hospital keep checking whether the AI still works?Leading systems run shadow-mode tests and ongoing monitoring, the way they would for a new drug.
Bracketed numbers on this page link here. Primary papers first, then the Wikipedia articles for background.