Quick Answer: A family has sued OpenAI after an epileptic woman died following her reliance on ChatGPT for medical advice, according to ABC10. Chatbots sound authoritative even when they're wrong, and epilepsy leaves little room for error. Use ChatGPT to understand terms and prepare questions, never to make medication or seizure decisions.
A family's lawsuit against OpenAI puts a hard question in front of anyone who has typed a symptom into a chatbot late at night: when does ChatGPT medical advice become dangerous? According to ABC10's coverage, the case centers on an epileptic woman who died after turning to the chatbot for guidance. The case is still unresolved, so this isn't a verdict. But the failure pattern it points to is one clinicians have worried about for a while.
What the Lawsuit Alleges, and What It Doesn't Prove Yet
Start with the distinction that matters most. A lawsuit is a set of allegations, not a finding. A court still has to decide whether the chatbot's output was a substantial factor in the death, and that's a hard causal question.
Epilepsy complicates causation. People with the condition can die for several reasons, including seizure-related injury, drowning, and SUDEP (sudden unexpected death in epilepsy). Any competent defense will argue that the death had multiple contributing factors, and any plaintiff will have to show which one the chatbot influenced.
Here's the practical takeaway even before the court rules. If a chatbot's advice changed what someone did with their medication, their seizure first aid, or their decision to seek emergency care, that's the part that matters. Read the original report yourself, and treat headlines as the start of the story, not the end.
Why Epilepsy Leaves Almost No Margin for Error
Most people think of epilepsy as a condition that's either controlled or not. In practice, it's a moving target. Seizure control depends on adherence to antiseizure medication, sleep, triggers, and timing, and small changes can tip the balance.
Published epilepsy research commonly cites SUDEP risk at roughly 1 in 1,000 adults with epilepsy per year, with substantially higher rates in people whose seizures are uncontrolled. That's a low absolute number, but it's the reason clinicians push so hard on seizure control and emergency planning.
Here's a cause-and-effect chain worth knowing. When someone stops or reduces an antiseizure medication abruptly, seizure risk can rise sharply, and a single missed or doubled dose can cause a breakthrough event. A chatbot that gives a generic answer about "adjusting" a dose, without knowing the drug, the level, or the patient's history, can push a reader into exactly that risk.
Why Chatbots Fail on Medical Questions in Ways You Won't Notice
The popular advice is "just double-check what the AI says." That advice is incomplete. Checking an answer requires the very knowledge you were asking for in the first place. If you didn't know that a symptom was a red flag, you won't know to verify it.
The more useful framing is calibration. Large language models generate fluent text by predicting likely next words, and fluency is not the same as accuracy. A hallucination, meaning a confident statement that isn't supported by reliable sources, looks exactly like a correct answer on the screen.
Here's the counter-intuitive part. ChatGPT is often right about general information, and that's precisely what makes it risky. Consistent accuracy on textbook questions trains users to trust the chatbot on the one question where a mistake is fatal. Most guides skip this, but the danger isn't frequent error. It's rare error delivered with the same tone as frequent accuracy.
This next part matters more than it looks. A chatbot answers the question you typed, not the one you didn't know to ask. It may explain what a medication does without mentioning the interaction that makes your situation different, or describe seizure first aid without flagging when to call emergency services.
Comparing Your Options: Chatbot, Clinician, or Epilepsy Organization
Not every source of health information carries the same risks or the same accountability. The table below compares the three options most people actually use.
| Source | Strongest use | Main limitation | Accountability |
|---|---|---|---|
| ChatGPT (OpenAI, general-purpose LLM) | Explaining terms, drafting appointment questions | Can state errors fluently; no access to your chart; cannot examine you | Terms of use direct users not to treat output as a sole source of truth; no clinical duty of care |
| Neurologist or epileptologist | Diagnosis, medication changes, personalized seizure action plans | Limited appointment time; wait times for specialists can run for months in many areas | Licensed; subject to professional standards and malpractice liability |
| Epilepsy Foundation (US nonprofit) | Plain-language seizure first aid, helpline access, local resources | Educational, not individualized | Nonprofit; educational content is clinically reviewed |
The trade-off is real. A chatbot is available at 2 a.m. and costs nothing, while a specialist may be months away. That access gap is why people turn to chatbots in the first place, and it's why the design of these tools matters so much. General-purpose chatbots typically aren't marketed as medical devices, which leaves them in a regulatory gray zone compared with clinical decision support software regulated by the FDA.
A Practical Protocol for Using AI on Health Questions
You don't have to abandon chatbots to use them safely. You do have to set rules before you need them, because the moment of crisis is the worst time to decide what to trust.
Follow these steps:
- Use the chatbot for vocabulary and questions, not decisions. Ask it to explain a term, then write down three questions for your clinician. The rationale is simple: a chatbot is better at framing a question than at answering one that depends on your chart.
- Never change antiseizure medication based on a chatbot's output. Dosing and withdrawal decisions need your prescriber, because the wrong adjustment can trigger the very seizures you're trying to prevent.
- Ask for red flags first, then treat the answer as a floor. Prompt it with "What symptoms mean I need emergency care right now?" Then remember that the list is a minimum, not a complete checklist.
- Log seizure details consistently. Record time, duration, triggers, sleep, missed doses, and the post-ictal state (confusion or fatigue after the seizure). These details are what a clinician needs and what a chatbot can't reconstruct.
- Hard-code your escalation triggers. Call emergency services for any seizure lasting longer than five minutes, back-to-back seizures without recovery in between, injury, or trouble breathing afterward. Those thresholds are standard first aid guidance, and they shouldn't depend on what a chatbot says at the time.
A short checklist also helps when you're preparing for an appointment. Bring:
- A current medication list with doses and timing
- Your seizure log for the past 60 days
- Any questions you drafted with a chatbot, marked as questions rather than conclusions
- The name and number of your emergency contact
Here's where it gets interesting. These steps don't stop the technology from being wrong. They stop a wrong answer from being the last step in your decision.
What This Case Could Change, and What It Probably Won't
Legal scholars and courts are still working out how product liability and Section 230 of the Communications Decency Act apply to generative AI output. The law here is unsettled, and experts disagree about whether a chatbot's answers should be treated like publisher content, a product, or a service. That uncertainty is a reason for caution, not a reason to predict a particular outcome.
Some possible changes are plausible. Companies could add stronger warnings, more consistent escalation to emergency services for high-risk symptoms, and clearer disclaimers on medical topics. Regulators could look harder at general-purpose tools that people use for health decisions.
What probably won't change is the basic gap between what a chatbot can do and what a clinician can do. A model can't examine you, watch you during a seizure, or take responsibility for a dosing decision. That's the part practitioners don't advertise, and it's the part your safety plan has to cover.
Most people stop reading at the legal drama. Don't. The lesson for patients is operational, and you can act on it this week.
Frequently Asked Questions
Is it safe to ask ChatGPT about symptoms?
It can be useful for general education, like explaining what a symptom might indicate or what questions to ask a doctor. It isn't safe as a basis for diagnosis, treatment, or emergency decisions, because it can't examine you and can state errors with confidence. Use it to prepare for care, and contact a clinician or emergency services when symptoms are new, severe, or changing.
What is SUDEP?
SUDEP stands for sudden unexpected death in epilepsy. It refers to an unexpected death in someone with epilepsy where no other cause is found, and it's thought to be linked to seizure activity, particularly uncontrolled tonic-clonic seizures. Published estimates commonly cite about 1 in 1,000 adults with epilepsy per year, with higher risk when seizures aren't controlled. Discuss your individual risk with your neurologist.
How do I use ChatGPT for medical advice safely?
Limit it to explaining terms and drafting questions for your clinician. Never use it to change medication doses or stop treatment, and set emergency triggers in advance, such as any seizure lasting over five minutes. Keep a seizure log and bring it to appointments, so the decisions rest on your actual history.
Why does ChatGPT give wrong medical answers?
Language models generate text by predicting plausible continuations, not by checking facts against a verified medical record. When they lack reliable information or misread the context, they can produce a hallucination that sounds identical to a correct answer. The risk is highest when the question depends on details the chatbot doesn't have, such as your drug levels, other medications, or current symptoms.
Use ChatGPT Medical Advice as a Starting Point, Not a Final Answer
The single most actionable takeaway is this: chatbots are good at framing questions and bad at owning answers, so keep the decisions with a clinician. Before your next health question, write down your emergency triggers and the three medication questions you need answered, and share them with your neurologist. Try building that one-page plan this week, then note how much clearer your next appointment feels. Pass this to someone managing a seizure disorder who has been relying on a chatbot for answers.