AI Misdiagnosis: When a Chatbot Gets Your Symptoms Wrong or Tells You Not to Go to the ER

An AI misdiagnosis can go wrong in two directions, and they are not equally dangerous.

The asymmetry

Over-calling — you use ChatGPT for medical diagnosis and the chatbot suggests something frightening that turns out to be nothing. The result is anxiety, maybe an unnecessary test, maybe an urgent care copay. Unpleasant. Rarely harmful.

Under-calling — the chatbot reassures you that something serious is ordinary. The result is that you stop looking. You don’t go in. You don’t mention it at your next visit because it was already explained.

False reassurance does something a wrong alarm never does: it ends the search. And for conditions where the treatment window is measured in hours, ending the search is the injury.

This asymmetry explains why the clearest AI misdiagnosis case on file in California is not about an exotic wrong guesses, but rather a confident, calm, wrong reassurance.

What the research says about ChatGPT diagnosis accuracy

A February 2026 study in Nature Medicine tested ChatGPT Health on 60 clinician-written case vignettes, each run under 16 variations (960 responses in all), and found that it under-triaged 52% of emergency presentations, rating them less urgent than they were. (Ramaswamy et al., ChatGPT Health Performance in a Structured Test of Triage Recommendations (2026) 32 Nature Med. 1671.) That last qualifier is the one that matters — emergency presentations are exactly where the error is least recoverable.

The public claims have run well ahead of that. OpenAI itself has said its models outperform physician-written answers on a benchmark OpenAI designed, which is the “Doctor ChatGPT” pitch in the company’s own words. (OpenAI, Introducing HealthBench (May 12, 2025).)

Both things can be true at once, and mostly are, which is why “how accurate is ChatGPT in medical diagnosis?” has no single answer. An AI doctor diagnosis can perform impressively on curated exam questions and benchmarks the developer designs, where the presentation is clean, complete, and written by someone who already knows the answer. But real, clinical presentations arrive incomplete, coming from a patient, potentially missing the detail that would have made it obvious. That gap is where the harm lives.

Why chatbot reassurance is unusually persuasive

A search engine returns ten links of varying alarm and leaves the user to weigh them. The disagreement itself is information — it tells one that the question is unsettled.

A chatbot returns one answer, in the second person, in composed prose, with no visible uncertainty. That is why people use ChatGPT for medical diagnosis, and why an AI misdiagnosis lands harder than a wrong search result. It does not hedge the way a clinician hedges, because it has no liability instinct and no memory of being wrong.

Two further features intensify it. These systems are trained in ways that reward agreeable, satisfying responses — and the satisfying answer to “should I be worried?” is usually no. And in long conversations the model has accumulated context that reads like familiarity: it knows your history, your medications, your prior episodes. That feels like a clinician who knows you. Not so.

Patterns that recur in ChatGPT medical diagnosis records

• A serious condition reframed as a familiar one. Cardiac symptoms explained away as anxiety or reflux. A clot explained as a strain.
• Symptoms attributed to a condition one already has. The single most common pattern in these records. A person with a chronic diagnosis reports something new; the chatbot folds it into the existing story. One public complaint quotes an answer describing new pain as “very likely another minor piece of the long story.”
• A threshold for when to worry. For example, if it happens several more times. A number converts a worry into a rule, and rules get followed.
• A condition ruled out by name. For example, this isn’t a clot. Legally this is significant: California’s statutory definition of diagnosis reaches any undertaking “to ascertain or establish whether a person is suffering from any physical or mental disorder,” and a no answers that question as surely as a yes. (Bus. & Prof. Code, § 2038.) Ruling out is diagnosing.
• Lab or imaging values interpreted. Increasingly common now that products connect directly to patient portals. Telling someone their result is normal is a clinical judgment.
• Advice to manage it at home, where the home management is itself a risk — rest and immobility for someone with clot risk factors being the clearest example.
• A new symptom after a medication change, explained as adjustment. Withdrawal and side effects get described as the body settling in, the one reading that keeps a person on the new course.
• Coaching on what to tell, or not tell, one’s doctor. A conversation that scripts the appointment, or suggests leaving something out, has put itself between the patient and the clinician.

When AI tells one not to go to the ER (delayed emergency care)

Most medicine tolerates a delay. A rash can wait until Monday. A sore knee can wait a week. A narrow set of conditions cannot wait at all, and those are the ones where a reassuring wrong answer does the most damage: not a wrong pill, not a bad procedure, just a confident, calm answer that reframes something urgent as something ordinary, and the hours that pass afterward.

What delayed care actually looks like

The now-public California complaint filed by a Florida pastor sets out the shape clearly. In 2025 he asked ChatGPT about recurring dizziness and unstable blood pressure. The chatbot, he alleges, treated both as relatively minor, suggested he stay home and remain “recliner-bound,” and told him his condition would need something on the order of eight to ten more episodes before it counted as serious. He followed that advice. The intake version is “ChatGPT told me not to go to the ER.”

Weeks later he suffered a massive pulmonary embolism — clots in both lungs. His physicians, according to the complaint, linked the embolism to the prolonged immobility. Among its claims is a negligence count that uses California’s medical-licensing statutes as the standard of care. The New York Times reported on it.

Read purely as a sequence of events, two things stand out. The advice was not exotic; it was ordinary-sounding and calm. And the recommended course of action — rest, stay put, wait for a threshold — was itself a risk factor for the very thing that happened.

The conditions where hours matter

Not every delay is a case. Below are representative diagnoses where a few hours change the outcome, and where a reassuring answer does real work:

• Pulmonary embolism — chest pain, breathlessness, a fast heart rate, calf pain or swelling, recent surgery, immobility, a long flight, hormone therapy.
• Stroke — face droop, arm weakness, speech trouble; the treatment window is measured in hours from the last time you were well.
• Heart attack — chest pressure, jaw or arm pain, nausea, sweating. Presentations in women and in people with diabetes are often described as atypical and are more easily dismissed.
• Sepsis — fever or low temperature, confusion, fast breathing, low blood pressure, an infection that keeps worsening.
• Appendicitis — pain migrating to the lower right; perforation risk climbs with delay.
• Diabetic ketoacidosis — thirst, frequent urination, vomiting, deep breathing, fruity breath.
• Testicular or ovarian torsion — sudden severe pain; salvage is measured in hours.
• Meningitis — headache with fever, stiff neck, light sensitivity, rash.
• Ectopic pregnancy — one-sided pelvic pain with bleeding; rupture is a surgical emergency.

Two features of these records that matter most

Numeric thresholds. Wait for eight to ten more episodes. Go in only if the pain passes a 7. A number converts a vague worry into a rule, and a rule is something people follow. It also manufactures a permission structure for waiting — every episode short of the threshold now reads as evidence that waiting was correct.

Thinning disclaimers. In at least two public complaints, early answers carry “it is important to consult a healthcare provider,” and the caveats appear less and less as the conversation lengthens, until they stop. OpenAI has publicly acknowledged that its safeguards “can sometimes be less reliable in long interactions” because parts of the model’s safety training may degrade. That degradation is most consequential in exactly the conversations that run longest — the ones where someone is frightened and keeps checking back.

Time is a treatment decision

Telling someone their symptoms do not require care is not only an opinion about symptoms, but rather a decision about treatment. California’s Business and Professions Code section 2052 reaches an unlicensed person who diagnoses or treats a condition, and “you do not need to go in” is treatment direction. California’s Supreme Court has read “treatment” to reach even an admonition to rest in bed and drink fluids. (People v. Burroughs (1984) 35 Cal.3d 824.)

There is a useful contrast in the case law. A physician who refused to prescribe over the telephone and sent the patient to a doctor was held not negligent. (Barton v. Owen (1977) 71 Cal.App.3d 484.) On the other hand, a physician who prescribed based on nothing but a web questionnaire, never speaking to the patient, could be prosecuted in California for practicing medicine without a license. (Hageseth v. Superior Court (2007) 150 Cal.App.4th 1399.)

The difference between those two cases is whether the speaker substituted itself for care or routed the person toward it. A bare “call 911” routes the person toward care. Substantive direction given after or instead of a referral substitutes for it.

What a review of a delayed-care record looks for

Nobody can tell you over the phone whether you have a case. A serious review looks at whether the record shows:

• symptoms disclosed with enough specificity to be recognizable;
• an answer addressed to you, in the second person, rather than general information;
• reassurance, a threshold, or an instruction to wait;
• an absence of urgent referral, or a referral overtaken by contrary direction;
• reliance — you actually waited because of it;
• a diagnosis whose outcome turns on timing (e.g. the “ChatGPT delayed my diagnosis” element);
• a clinical record supporting that earlier care would have mattered.

Not every element must be present. Missing one is not a reason to skip the call. If your version is “AI told me not to go to the hospital” and you turned out to need one, that is enough to have the record read.

“ChatGPT told me not to go to the ER, but I went anyway.” Do I still have a claim?

Usually the answer turns on the delay. If you went hours or days later than you otherwise would have, and the outcome changed in that window, the record is still worth reviewing. The timeline is what makes or breaks a delayed-care claim: preserved timestamps on both sides, the conversation and the chart, are the single most valuable thing you can give a reviewing attorney.

What makes it a legal question

An AI wrong diagnosis alone is not a claim. Physicians are wrong sometimes too; the law does not treat every error as actionable.

What changes the analysis is the combination California law looks for: an unlicensed actor, diagnosing or treating a particular person, who relied, and was harmed.

Business and Professions Code section 2052 reaches an unlicensed person who diagnoses or treats a physical or mental condition. And because California bars corporations from practicing medicine at all (Bus. & Prof. Code, § 2400), the question is never whether software can hold a license, but rather whether the company’s product did what only a licensee may do. The statute’s definition of diagnosis covers any undertaking “by any method, device, or procedure whatsoever, and whether gratuitous or not, to ascertain or establish whether a person is suffering from any physical or mental disorder.” (Bus. & Prof. Code, § 2038.)

The threshold question is not whether AI can diagnose correctly, but rather individualization — whether the output described a disease in general or told you what you had. More on where that line sits →

If it happened to you

Medical first. If symptoms persist, get seen, and say plainly that you delayed and why.

Then preserve:

1. Export your conversation data — ChatGPT: Settings → Data Controls → Export Data. Instructions →
2. Do not delete the thread or the account, and do not go back to continue it.
3. Screenshot the exchange with dates visible.
4. Request medical records for the same window.
5. Write down the timeline while it is fresh.

Then have it read. “ChatGPT misdiagnosed me” is the start of a conversation, not a conclusion; whether a record supports an AI misdiagnosis claim depends on what is in it. Review costs nothing.

AI medical advice claims →

Why this firm

The Schenk Law Firm is a San Diego personal injury and mass tort firm that reviews AI medical advice claims nationwide. Attorney Benjamin S. Schenk was appointed by Hon. Ethan P. Schulman to serve as a member of the Plaintiffs’ Steering Committee, a leadership role in the California coordinated proceeding covering ChatGPT product liability cases. The Schenk Law Firm has recovered more than $25 billion on behalf of clients. Past results do not guarantee future outcomes.

We are experienced with emerging-technology injury claims — underage online gambling, social media addiction, and AI-related mental health litigation among them. We also represent the family of Bowen Wan, the center of The New York Times’ recent reporting on AI medical advice. Cases involving AI medical advice are new, the law governing them is developing, and no lawyer can promise an outcome in any of them. What we can tell you is whether a record shows the pattern a ChatGPT lawsuit requires.

General information, not medical or legal advice. For a possible emergency, call 911.