The patient's AI got there first
A chatbot's differential or a smartwatch alert belongs in the history, as a record of what worried the patient; physicians should treat it neither as a diagnosis nor as a threat.
Listen to this special topic
read by an AI voiceOn Oct. 5 a randomized trial from Guangzhou, China, reported that a language model agent took better histories than ophthalmology residents. Before their consultations at a tertiary eye hospital, 172 patients with non-emergency appointments were interviewed either by the agent, which was guided by question trees built around symptoms, or by a resident. The agent's histories scored an average of 17.7 points higher, after adjustment, on a 100-point quality scale, and its interviews ran a median of 11.2 minutes against the residents' 3.1. The authors propose a division of labor in which software gathers the history and clinicians concentrate on the examination and on putting the findings together.
The AI that patients bring to the clinic is different. It is a general chatbot that the patient questioned alone, or a watch worn day and night, and it arrives as a list of possible diagnoses or a screen of numbers. About a third of American adults turned to an AI chatbot for health information in the past year, and a quarter say they use one to work out what is causing their symptoms. The physician has a short visit in which to respond, and the right response is to take what the patient brings as history: a record of what the patient noticed, asked and fears. It is not a diagnosis to be confirmed or refuted, and it is not a challenge to be fended off.
It is not a diagnosis, because a chatbot's answer depends on what the patient told it. In a preregistered Oxford study of 1,298 people, language models working alone through ten written medical scenarios named a relevant condition in 94.9% of cases. Participants randomly assigned to use the same models named one in fewer than 34.5% of cases, often because they gave the models too little information, and those left to use whatever they usually would, mostly an internet search, did better. Urgency is harder still. When researchers at Mount Sinai tested ChatGPT Health with clinical vignettes in January, it sent textbook emergencies such as stroke to the emergency department, but in 33 of 64 responses about diabetic ketoacidosis and an asthma attack with impending respiratory failure it advised being seen within 24 to 48 hours instead. Its advice in borderline cases also shifted, mostly toward less urgent care, when a relative or friend in the vignette played down the symptoms. OpenAI said the test did not reflect typical use; at launch it had said the product is not intended for diagnosis or treatment.
Nor is it a threat. Patients brought printouts from the internet long before chatbots. In a national survey of 1,050 American physicians conducted in 2000 and 2001, 85% had seen a patient bring internet information to a visit; in their judgment it helped care when it was accurate and relevant and harmed it when it was not. Most physicians accept the newer version: in an AMA survey this year, 70% called patients' use of general-purpose AI for health information positive or neutral when used appropriately. Patients' reasons are ordinary: in KFF's poll two-thirds of users wanted quick information, and about one in five cited the cost of seeing a clinician or the difficulty of getting an appointment. About four in ten who used AI for physical health did not follow up with a doctor or other provider. The patient who brings a chatbot's list to the visit is one who did.
Treating it as history puts it where physicians already know what to do with it. A patient's account of symptoms is incomplete, colored by worry and sometimes wrong, and physicians still take it seriously, test it against the examination and write it down. A chatbot's differential is the same kind of evidence. It shows what the patient described, which possibilities are now on the patient's mind and which of them the patient fears. Many patients will not offer it unprompted: 29% of physicians in the AMA survey said they had not had a single patient disclose using AI, although about a third of adults use it for health.
Time is the practical objection. In 21 million primary care visits in 2017, the physician's time with the patient averaged 18 minutes, and in an older study of videotaped visits by elderly patients, the longest topic got about five minutes and each of the others about a minute. Taking the patient's research as history should cost less time than rebutting it item by item. Asking what the patient read and what worries them is also a way of drawing out the patient's agenda, which clinicians did in 36% of recorded encounters in one study; in two-thirds of those, they interrupted the patient after a median of 11 seconds.
Wearable data need one adjustment. They are measurements rather than a story, and they should be read like the result of a screening test. In 2024, 41% of American adults used a wearable device. Sleep scores, heart rate variability and similar outputs can be sold as wellness features, which the FDA, under guidance revised in January, does not intend to examine as medical devices when they are low risk and meant only for wellness. The few features that are cleared have modest numbers. Apple's hypertension notification, cleared in September 2025, detected 41.2% of people whose home blood pressure averaged 130/80 or higher in the company's validation study and stayed silent for 92.3% of those below it; with a negative likelihood ratio of 0.64, a silent watch only modestly lowers the chance of hypertension. Abnormal readings also generate work. Of 264 patients whom Mayo Clinic evaluated for an abnormal pulse detected by an Apple Watch, 30, or 11.4%, received a clinically actionable cardiovascular diagnosis.
The strongest objection is that calling this history undersells it, because the software is sometimes the better diagnostician. In a secondary analysis of a randomized trial published in 2024, GPT-4 working alone scored 16 percentage points higher on diagnostic reasoning than physicians using conventional resources, and in the trial itself physicians given the model did no better than those without it. In Guangzhou, in an analysis the authors call exploratory, the agent's own diagnoses from the history alone beat the residents'. But those models had complete written cases or asked their own structured questions. A patient's chatbot has only what the patient typed, which is one reason the Oxford participants did poorly with models that did well alone; the other was that they often failed to carry the models' correct suggestions into their own answers. And in Guangzhou, adding the eye findings helped the residents more than the agent, significantly narrowing the gap. Taking the output as history does not mean discounting it. A diagnosis on a patient's printout deserves the same consideration as any other possibility the history raises, and sometimes it will be the right one.
None of this takes long. Early in the visit the physician should ask every patient whether they have looked the problem up, asked a chatbot or brought readings from a device, since many will not say so otherwise. What the patient typed shows what the software was told, and the possibility that worries the patient most is what the visit has to answer. Red flags come before the list, because a printout that reassures is no reason to stop looking. Device data should be read like a screening result, with two questions: which feature produced the reading, and whether it is cleared for that use. Anything that matters should be confirmed with a validated measurement such as a home blood pressure cuff, an ECG or a sleep study. A wellness score diagnoses nothing, and a silent watch rules nothing out.
Then the list deserves an answer, however brief: why each serious possibility does or does not fit, and why a test will not be ordered. Copic, a liability insurer, suggests telling patients that AI tools "do not account for downstream harms of unnecessary testing." The reasoning belongs in the chart in neutral terms, such as "patient inquired about." And the patient should leave able to say the plan back, the teach-back method, knowing what would warrant coming back sooner and invited to bring the whole chatbot conversation next time rather than just its conclusion.
In Guangzhou the authors would have software ask the questions and leave the examination and the synthesis to clinicians. A patient's chatbot attempts the synthesis with nothing but what the patient typed. Physicians should take its work as part of the history, examine the patient and do the synthesis themselves.
- Luo and colleagues (npj Digital Medicine, published Oct. 5, 2026, as an accelerated article preview "subject to further edits") randomized 172 of 175 approached patients with non-emergency appointments at a tertiary ophthalmic hospital in Guangzhou, China, between May 10 and June 10, 2025, to pre-consultation history-taking by a large language model agent built "using a dual-agent architecture with skills-engineered symptom-oriented logic trees" or by ophthalmology residents. The agent scored higher on the MedHistory score (range 0 to 100; adjusted mean difference 17.7 points, 95% CI 14.1 to 21.3), received higher patience and empathy ratings (5 vs 4 and 5 vs 3 points) and had longer interactions (median 11.2 vs 3.1 minutes). "Test recommendation performance did not differ significantly across stages." In exploratory analyses the agent's diagnostic accuracy from history alone was higher (F1 score 0.85 vs 0.68), "whereas ocular signs produced greater improvement among residents (interaction effect -0.19; P = 0.003), significantly narrowing the gap." The authors conclude that the findings support "a hybrid workflow where clinicians prioritize their expertise for physical examination and diagnostic synthesis." Neither the abstract nor the supplementary information says who gave the patience and empathy ratings; the protocol lists patient satisfaction among the secondary outcomes. The "ocular signs" were written descriptions of slit-lamp findings, visual acuity and intraocular pressure given to both groups after the history stage. The trial was open-label, at a single hospital, and registered as NCT06824389 and funded by the National Natural Science Foundation of China and other public and university grants; the authors declare no competing interests.
- KFF's poll of 1,343 adults (Feb. 24 to March 2, 2026) found that 32% had turned to AI chatbots for health information in the past year, 29% for physical health and 16% for mental health; 65% of users cited quick or immediate information as a major reason, 19% the cost of seeing a health professional and 18% not having a regular doctor or being unable to get an appointment. Of those who used AI for physical health advice, 58% said they later followed up with a doctor or health care provider; 41% of users said they had uploaded personal medical information, and 77% of adults were concerned about the privacy of medical information given to AI tools. Pew Research Center (survey of June 22 to 28, 2026) found that 34% of Americans have ever used AI chatbots for at least one of eight health reasons and that "A quarter of Americans say they use AI chatbots to diagnose symptoms." OpenAI says more than 300 million people a week ask ChatGPT health questions (July 2026, a company figure) and that Health "is not intended for diagnosis or treatment" (January 2026). In the AMA's survey of 1,692 physicians (Jan. 15 to Feb. 2, 2026), "29% have not had a single patient disclose it," and 70% view patients' use of general-purpose AI for health information as positive or neutral when used appropriately. In a survey of 355 clinicians conducted by Ipsos for Wolters Kluwer, which sells clinical reference software, 60% of doctors and nurses said patients frequently or very frequently bring in AI-generated information.
- Bean and colleagues at the University of Oxford (Nature Medicine, Feb. 9, 2026) gave 1,298 participants living in the UK ten medical scenarios: tested alone, the models (GPT-4o, Llama 3 and Command R+) identified conditions in 94.9% of cases and the right course of action in 56.3%, but participants using them "identified relevant conditions in fewer than 34.5% of cases and disposition in fewer than 44.2%, both no better than the control group"; participants in the control group, who mostly used a search engine or trusted websites, had 1.76 times higher odds of identifying a relevant condition. Ramaswamy and colleagues at Mount Sinai (Nature Medicine, Feb. 23, 2026) tested ChatGPT Health with 60 clinician-written vignettes under 16 conditions (960 responses) between Jan. 9 and 11, 2026; it under-triaged 51.6% (33 of 64) of responses to gold-standard emergencies and over-triaged 64.8% (83 of 128) of nonurgent ones, and when family or friends minimized symptoms its recommendations shifted in edge cases (odds ratio 11.7), mostly toward less urgent care. The emergency responses came from two scenarios, diabetic ketoacidosis and an asthma exacerbation; in a supplementary analysis of four textbook emergencies (stroke, anaphylaxis, meningitis and aortic dissection; 128 responses) none was under-triaged. An OpenAI spokesperson told NBC News that the study did not reflect how ChatGPT Health is typically used or how it is designed to function. Mount Sinai advises that for worsening or concerning symptoms, such as chest pain or shortness of breath, people seek care directly rather than rely solely on a chatbot. Goh and colleagues (JAMA Network Open, Oct. 28, 2024) randomized 50 physicians: access to GPT-4 did not significantly improve their diagnostic reasoning, and in a secondary exploratory analysis the model alone scored 16 percentage points higher than physicians using conventional resources. Murray and colleagues (Journal of Medical Internet Research, 2003) surveyed 1,050 U.S. physicians in 2000 and 2001: 85% had had a patient bring internet information to a visit, and 38% said it harmed their time efficiency.
- An analysis of the National Cancer Institute's HINTS survey (European Heart Journal: Digital Health, January 2026) estimated that the share of U.S. adults using wearable devices rose from 30.2% in 2020 to 41.1% in 2024; most respondents were willing to share their data with clinicians, but 19.2% had done so in 2024. The FDA's general wellness guidance, revised Jan. 6, 2026, describes its compliance policy for "low risk products that promote a healthy lifestyle"; Covington & Burling notes that it now covers non-invasive products that estimate physiologic parameters, including blood pressure, for wellness uses. Cleveland Clinic notes that "HRV that's normal in one person may not be normal for someone else" and that wrist devices "usually aren't sensitive enough to detect HRV accurately." A study funded by Oura (Sensors, 2024) of 35 healthy adults over one night found sensitivity of at least 95% for sleep versus wake but 50% to 86% for sleep stages, against polysomnography.
- The FDA cleared Apple's Hypertension Notification Feature (510(k) K250507) on Sept. 11, 2025, for people 22 or older not previously diagnosed with hypertension. In Apple's validation study its sensitivity was 41.2% and its specificity 92.3%; a family physician writing on the American Family Physician blog gives likelihood ratios of 5.35 and 0.64 and notes that the watch "is not a substitute for a home BP monitor." Apple's sleep apnea notification (K240929, cleared Sept. 13, 2024) had a sensitivity of 66.3% and specificity of 98.5% for moderate to severe sleep apnea in Apple's validation, and Apple says "the absence of a notification isn't intended to indicate the absence of sleep apnea." In the Apple Heart Study (New England Journal of Medicine, 2019, funded by Apple), 2,161 of 419,297 participants were notified of an irregular pulse, and atrial fibrillation was present on a follow-up ECG patch in 34% of the 450 who returned a patch with analyzable data. Mayo Clinic researchers (JAMIA, 2020) reviewed 264 patients evaluated for an abnormal pulse detected by an Apple Watch, only 41 of whom had a documented alert; 30 (11.4%) received a clinically actionable cardiovascular diagnosis.
- Neprash and colleagues (Medical Care, 2021) analyzed 21 million primary care visits from 2017: "the average primary care exam was 18.0 minutes long." Tai-Seale and colleagues (Health Services Research, 2007), studying videotaped visits by elderly patients from 1998 to 2000, found a median visit of 15.7 minutes covering six topics, with about five minutes on the longest topic and 1.1 minutes for each of the rest. Singh Ospina and colleagues (Journal of General Internal Medicine, 2018) found that clinicians elicited the patient's agenda in 40 of 112 encounters (36%) and, when they did, interrupted after a median of 11 seconds in 27 of 40. AHRQ describes teach-back as "a way of checking understanding by asking patients to state in their own words what they need to know or do about their health."
- Copic, a medical liability insurer, published guidance on June 2, 2026, for visits in which patients bring AI-generated information. It suggests acknowledging the patient's work ("Let's walk through what you found and how it applies to your situation"), explaining that AI tools "do not account for downstream harms of unnecessary testing" and charting in neutral terms such as "patient inquired about" or "patient requested discussion of," with the clinical reasoning for accepting or declining suggestions. The AMA's advice to patients (May 20, 2026) is to use chatbots to explore possibilities and prepare for visits, and to protect their privacy.Sources: Copic, American Medical Association