Artificial Intelligence in healthcare gets a lot of praise — and often for good reason. More and more patients, and even doctors, are placing their trust in AI tools. What many of us fail to notice is that the same technology can also cause real harm, sometimes with grave consequences. That cuts against the very first principle of medicine: first, do no harm.
Most such cases never make the news. The five cases below — each showing a different kind of danger, and all reported in the media or documented by researchers — reveal how AI advice or AI-driven tools malfunctioned and left patients seriously harmed. They are almost certainly a small fraction of the true number. Together, they point to a single lesson: in healthcare, AI must be used with the utmost caution.
Case 01It can tell patients to stop their prescribed medicine
This is the most direct danger of all. A patient asks an AI tool about their treatment, takes its words as medical permission, and stops a drug the doctor told them to keep taking.
A young transplant patient lost her new kidney
A 30-year-old woman had received a kidney transplant. After an AI tool told her that her “normal creatinine” meant she no longer needed her medicines, she reportedly stopped taking her antibiotics. Within weeks, her transplanted kidney began to fail, her creatinine shot up, and she was back on dialysis.
Senior kidney doctors at NIMS pointed to a worrying pattern: even well-educated patients are now acting on AI answers without checking with their own care team.
As reported in Indian news media (2025).
Case 02It gives one-size-fits-all advice that ignores your condition
An AI tool does not examine you. It does not know your other illnesses, your medicines, or your test results. So its “general health tips” can be exactly wrong for the person reading them.
A diabetic man's sodium dropped dangerously after cutting out salt
A 62-year-old man with diabetes followed a plan from an AI tool that told him to cut out salt completely. He lost weight rapidly and his blood sodium fell to a dangerous level. As one government kidney specialist put it: “General tips ignore the patient in front of you.”
As reported in Indian news media (2025).
A man poisoned himself swapping salt for a chemical
Wanting to cut salt from his diet, a 60-year-old man asked an AI tool what to use instead and understood it to suggest a chemical called sodium bromide. He bought some, used it for three months, and ended up in hospital for three weeks with poisoning, paranoia, and hallucinations. The AI tool never warned him the chemical was unsafe to swallow.
Reported as a medical case in the Annals of Internal Medicine: Clinical Cases (2025).
Case 03It can hand out dangerous do-it-yourself advice
Some AI tools do not stop at bad diet tips — they tell people to carry out risky procedures on themselves.
A man was injured trying to treat his own piles
A 35-year-old man reportedly followed an AI tool's instructions to place rubber bands around his haemorrhoids (piles) himself. The result was an injury serious enough to need emergency medical treatment.
As reported in news media.
Case 04It can make people put off seeing a real doctor
Perhaps the quietest harm of all is delay. A confident answer makes people feel reassured, so they treat themselves at home and reach a doctor only once things have turned serious.
Self-treatment delayed care until it became an emergency
A 42-year-old office worker had ongoing tiredness and mild stomach discomfort. Instead of seeing a doctor, he turned to an AI tool and began treating himself based on its suggestions. Weeks later his condition worsened and he needed emergency care.
Doctors say this is no longer a rare event. Some clinics now report 40 to 50 patients each with problems caused by AI-guided self-diagnosis — wrong drug doses, stopping prescribed medicines, or starting new treatments without asking a qualified doctor.
As reported in Indian news media (2025).
A confident answer is not the same as a correct one. An AI tool can sound calm and certain while being completely wrong about the person reading it.
Case 05Even the machines in the operating room can be wrong
The danger is not limited to the AI tools patients use at home. AI is now built into surgical equipment too — and when one of those tools is wrong, the surgeon may not realise it until the damage is done.
An AI surgery tool linked to strokes and skull injuries
After a popular surgical navigation system (the TruDi system, used in sinus operations) added AI in 2021, safety reports to the US regulator jumped from about 8 to more than 100 — roughly a 14-fold rise by late 2025. At least 10 patients were reportedly injured. The tool allegedly misled surgeons about where their instruments were inside patients' heads, leading to leaks of spinal fluid, a punctured skull base, and at least two strokes after a major artery was damaged.
The makers deny that the technology directly caused the injuries, and the lawsuits are still ongoing — but the pattern alarmed safety experts.
Reuters investigation (February 2026).
When an AI tool tells a surgeon where to cut and is wrong, there is often no way to know until the harm is done.
The Bigger PictureQuieter dangers behind the scenes
Beyond these individual patient stories, AI carries deeper risks inside hospitals and health systems:
It can be unfair. A widely used US tool that decided who got extra care rated many Black patients as healthier than they actually were, because it used past spending as a stand-in for how sick someone was. It affected around 200 million people. (Science, 2019)
It can be used to deny care. US insurers have been sued over AI tools used to cut short care for elderly patients; in one case, about 9 out of 10 denials were overturned on appeal — yet very few patients ever appealed. (STAT & CBS News; ongoing lawsuits)
It can fail quietly and cry wolf. A widely used hospital tool meant to warn staff about sepsis (a dangerous infection) missed about two-thirds of cases when tested independently, while raising so many false alarms that staff stopped trusting it. (JAMA Internal Medicine, 2021)
It can make skilled people rusty — the “lazy doctor” phenomenon. After a few months of leaning on AI, experienced doctors became worse at spotting growths on their own — their unaided detection fell from about 28% to 22%. (The Lancet Gastroenterology & Hepatology, 2025)
What To DoHow to use AI without getting hurt by it
The goal is not fear — it is care. If you are a healthcare provider, a few simple habits can protect both your patients and your practice:
Ask about AI use when taking a history
Make a patient's use of AI for medical advice a routine, mandatory question during history-taking and initial assessment.
Tell patients plainly: an AI tool is not a doctor
Warn them never to stop, start, or change a prescribed medicine — or treat themselves — on the strength of an AI answer. Give them an easy way to check with your team first.
Keep a qualified person in charge of every clinical decision
Treat every AI output — whether an AI tool's answer or a surgical system's guidance — as a suggestion that a trained professional must check, never the final word.
Verify AI output before it reaches the patient
A qualified person should review any advice or information an AI produces before it is passed on to a patient.
Be open, and protect patient data
Tell patients when AI is used in their care, obtain consent where needed, and guard their information carefully.
AI is a tool, not a colleague
None of these stories mean you should avoid AI. The same technology that caused these harms is also catching cancers earlier and giving doctors their time back. The difference between help and harm is almost never the tool itself. It is whether a careful, well-trained team — and a well-informed patient — is watching over it.
So use AI — but keep your eyes open, keep a human in the loop, and never let a confident screen replace good judgement.
