For five days, Gleb Tsipursky had what he figured was a muscle cramp in his calf. His chiropractor agreed. Then he asked an AI, which immediately flagged deep vein thrombosis and told him to go to the emergency room. The ultrasound found four clots in his left leg. His wife's grandfather died of a pulmonary embolism. So did the mother of her close friend. The cramp was a cliff edge.
The System Said 'Schedule an Appointment'
Writing in The Guardian, Tsipursky describes what happened after his AI tool raised the alarm. He called his primary care office. They told him to schedule an appointment or try urgent care. Urgent care couldn't do the ultrasound. Neither could his doctor's office. The slow path, in other words, meant losing days before someone sent him to the ER anyway.
So he skipped straight to the ER himself, knowing he'd spend hours waiting instead of the few minutes he'd burn at urgent care. That decision, guided by his AI, is the reason he's writing the op-ed instead of appearing in someone else's obituary.
This is how the American healthcare system works for the average patient: it requires you to advocate loudly and correctly for yourself at the exact moment you are least equipped to do so. You are sick, possibly scared, and trying to decode which level of care a frightening symptom actually warrants. Get it wrong and you wait too long. Get it right and you spend half a day in an emergency room waiting room. Heads they win, tails you lose.
What DVT Actually Is, Because Your Doctor Has Six Minutes for You
Deep vein thrombosis is a blood clot that forms in a deep vein, usually in the leg. The CDC describes DVT and pulmonary embolism together as serious conditions that are chronically underdiagnosed. The symptoms, as Tsipursky lays them out, include pain, swelling, warmth and skin color changes, particularly when affecting one leg.
The reason it matters urgently is that part of a clot can break off and travel to the lungs, causing a pulmonary embolism. The National Heart, Lung, and Blood Institute warns that a pulmonary embolism can be fatal when a clot is large or multiple clots are present. Tsipursky had four.
The chiropractor treated it as a muscle issue. That is not a gotcha against chiropractors specifically. The point is that the symptom presentation looked ordinary and was treated ordinarily, right up until it wasn't.
The Science Is Actually There to Back This Up
This is not purely an anecdote. Tsipursky points to a study published in Science, led by researchers affiliated with Harvard Medical School and Beth Israel Deaconess Medical Center, that tested a large language model on clinical reasoning tasks using real emergency department cases. Science News reported that the model was more likely than physicians to include the correct diagnosis among the possible answers it generated.
That is a careful, limited finding and Tsipursky is careful about how he frames it. He is not arguing that patients should take medical advice from whatever chatbot shows up first in a Google search. He is arguing that doctors and AI used together may be safer than either operating alone. A system that surfaces possibilities and pattern-matches across records could reduce the chance that something dangerous gets waved off as ordinary.
The gap between those two things, chatbot roulette versus a supervised AI tool trained on your own medical history, is enormous. Conflating them is how people get hurt.
The Real Risk Nobody Wants to Talk About Straight
To his credit, Tsipursky does not bury the downside. The Guardian itself has reported that one in seven people in the UK are now using AI chatbots for medical guidance instead of seeing a GP. That number should alarm anyone paying attention. A chatbot cannot examine a leg, assess breathlessness, notice distress, or carry any legal or ethical responsibility for what happens to you.
The call Tsipursky is making is for regulation, testing, transparency, and clinical supervision. AI health tools that are accurate, accountable, and deployed under medical oversight. That is a reasonable and relatively modest ask. It is also roughly a decade away from being standard practice under anything resembling current regulatory momentum.
In the meantime, patients are using whatever they have. Some of them will be helped. Some of them will be steered wrong. The healthcare system has decided, largely by inaction, that this is an acceptable arrangement.
What He Actually Built and Why Most People Can't
Tsipursky is the CEO of a consultancy called Disaster Avoidance Experts and writes about AI adoption at work. He built his own AI health tool using his medical records, medications, lab work, and visit notes, drawing on his professional background in AI implementation. He had the expertise, the time, and the access to build something useful.
Most people have none of those things. The median American patient does not have their lab work and visit notes neatly compiled anywhere. The fragmented record-keeping across providers, insurers, and hospital systems makes assembling a coherent medical history an act of genuine determination. The thing that helped Tsipursky was, in part, that he had already done a tremendous amount of work most patients cannot easily replicate.
That is not a reason to dismiss what happened to him. It is a reason to take seriously what equitable access to this kind of tool would actually require.
The Dingo Take
Here is the thing about this story that keeps nagging. A man with a dangerous, potentially fatal condition spent five days treating it as a minor inconvenience, saw a professional who confirmed that incorrect assumption, and was then told by his actual medical system to wait in line. The machine that finally got it right was one he had to build himself. The system built by professionals and funded by decades of insurance premiums told him to schedule an appointment.
The case for AI as a healthcare tool does not rest on the machines being smarter than doctors. It rests on the machines being available at 11pm when your calf has been swollen for five days and your doctor's office is closed and urgent care can't do the scan anyway. It rests on a tool that will not feel dismissive of your concerns because it is running behind. It rests on something that will not see a 40-year-old man with a sore leg and immediately pattern-match to 'not urgent.'
Tsipursky's conclusion is measured and responsible. Regulation. Oversight. AI as a second opinion, not a replacement. All of that is correct. But let's be honest about why any of this is resonating: the existing system left him standing at the edge of a cliff without a guardrail, and a piece of software he built in his spare time is the reason he walked away from it. That is not an advertisement for AI. That is an indictment of everything else.
