The Opening
Some weeks the story is what AI does to medicine. This week it was what patients do with AI.
Google published the largest real-world test yet of a conversational symptom checker: nearly 14,000 people describing what actually hurt, in their own words. OpenAI switched on ChatGPT Health for US users and asked for the medical records to go with it. And statehouses kept writing rules for the chatbots that teenagers are already confiding in.
The waiting room has quietly gained a new occupant: the second opinion that arrived before the first one. Your patients are interviewing an algorithm on the drive over, and the algorithm is getting better at interviewing them back.
So this issue lives on the other side of the exam table. What the new evidence really shows, who is building it, where the guardrails are going up, and what it means when the person in front of you has already been diagnosed by something that never went to medical school.
Let's get into it.
- Troy, Ray, and Ibrahim
The chatbot will interview you now
On July 22, Google Research published results from SymptomAI, a conversational agent built on Gemini that interviews you about your symptoms the way a clinician would: asking follow-up questions, probing for detail, then producing a ranked list of possible causes. It ran quietly inside Fitbit Labs as an IRB-approved research study, and 13,917 consenting adults used it to talk through real symptoms in their own words.
The headline finding is about conversation, not computation. Every version of the agent that actively asked questions significantly beat the passive version that waited to be told. And when clinical experts compared assessments head to head, they ranked the AI's differential diagnosis (its list of likely explanations) as the best on offer in 53.3% of cases, edging out clinician reviewers reading the same transcripts.
Why it matters: Symptom checkers used to be multiple-choice quizzes with modest accuracy. This is the first evidence at national scale that a chatbot interviewing real patients, not exam vignettes, can produce a genuinely competitive differential.
The catch: It is a preprint, not a peer-reviewed trial. Ground truth was what participants said their clinician diagnosed two weeks later, and only a fraction returned that answer. The physicians it edged out were reading static transcripts, unable to ask a single question of their own. And you cannot download it: SymptomAI remains a research prototype.
Bottom line: The tools your patients use tonight are not this good, and this one is not available. But the direction is unmistakable: the AI interview is coming, and it will arrive prepared.
A patient walks in with a chatbot-generated list of possible diagnoses. What is your honest first reaction?
We're collecting stories. This week, one question.
This week's issue is about patients arriving with answers, so tell us what that looks like from your side of the exam table.
Tell us about the visit that changed when the patient pulled out a chatbot answer. The printout that caught something real, the confident nonsense you had to gently walk back, or the question you never would have been asked five years ago. Maybe it made the visit better. Maybe it swallowed the first ten minutes. Either way, you saw something the app stores never mention.
We want the real version, not the conference-panel version. Two paragraphs is plenty.
This is open to every corner of medicine: EMS, nursing, pharmacy, techs, registration, environmental services, and administration, all of it. You choose how you're named, whether that's full name, role only, or fully anonymous. We protect patients in every story we run. That is not negotiable and it never will be.
If we run your story, we'll send you a Consult mug as a thank you.
Dr. Jake Sunshine, Google Research and University of Washington School of Medicine
Behind this week's feature is an anesthesiologist who has spent a decade teaching consumer devices to notice when something is medically wrong. Jake Sunshine practices and teaches at the University of Washington, works as a research scientist at Google, and co-led the SymptomAI study alongside Joseph Breda and Daniel McDuff. His career has one consistent pattern: take the gadget already in the patient's home, and make it clinically useful.
Where things stand:
Co-led SymptomAI, the largest real-world evaluation of a conversational symptom checker to date: 13,917 consenting participants inside Fitbit Labs, published this July
Co-founded Sound Life Sciences in 2018, turning smartphones into contactless sonar breathing monitors, validated in more than 200 people who self-inject opioids; Google acquired the company in 2022
His University of Washington research showed smart devices can detect agonal breathing, the distinctive gasping of cardiac arrest, from across the room
Still a practicing, board-certified anesthesiologist and associate professor while running consumer-health research at Google
Illustration: The Consult
Why they matter: Most consumer health AI is built by engineers who have never admitted a patient. Sunshine works the other direction: a clinician embedding the bedside instinct for "something is off" into the phones and watches patients already own. If patient-facing AI is going to be safe, it will be because people who have carried a code pager helped build it.
ChatGPT Health rolls out with your records attached
OpenAI's Health feature is now live for US users 18 and up, on every tier, and it can connect Apple Health data plus medical records from supported systems. OpenAI says the data is not used for training or ads. The world's most popular chatbot now reads labs, meds, and sleep data; your patients' next questions will be personal.
Thirteen states put guardrails on chatbots
The Transparency Coalition counts 14 chatbot safety laws enacted so far in 2026, with Rhode Island the latest to require protocols for suicidal ideation and crisis referrals. The common thread: disclose the bot is a bot, protect minors, and respond to self-harm. Enforcement, as ever, is the open question.
Patients now ask AI before they ask you
A Corporate Insight review of 17 symptom checkers finds 68% of patients consult AI tools first for health questions, ahead of search engines. Most health systems' own tools lag: only 29% accept plain-language input and just 35% tailor their follow-up questions. The front door of medicine is moving, and it is not on the hospital's website.
One in five teens confides in a chatbot
A RAND survey in JAMA Pediatrics finds 19.2% of Americans aged 12 to 21 used AI chatbots for mental health advice last year, and 63.3% told no one. Most found the advice helpful. The unsettling part is the silence: support is happening with no clinician, or parent, in the room.
Twenty years of Dr. Google, and what comes next
Every generation gets a ritual for its fear. For the last twenty years, ours was typing symptoms into a search bar at 2 a.m. and scrolling until the headache became a tumor. "Dr. Google" became a cultural character: a punchline, a warning, the reason "don't google it" turned into standard medical advice.
This was the week the sequel arrived. The search bar now asks follow-up questions. It reads your labs if you let it, remembers your medications, and delivers its differential in the calm, confident voice of something that has never once been sued.
And yet, look closely at what people say they want. In a survey of 6,000 patients released this month, 90% still prefer to reach a human being when they call about their health, and 72% are uncomfortable with AI holding their personal health data. We are handing over our records with one hand and asking for a person with the other.
Maybe that's not a contradiction. Maybe it's a diagnosis. The 2 a.m. search was never really a request for information; it was a request for reassurance, and no ranked list has ever supplied that. What patients seem to be building, with the tools available to them, is a way to arrive at your door already believed, already taken seriously, already partway through the conversation.
The machines are getting very good at the interview. The question this week leaves behind is the one the interview was always for: when the list of possibilities is finally read out, whose voice do we need to hear it in?
Until next Sunday,
For the people keeping medicine human.
