OpenAI and CEO Sam Altman were sued in late October over the wrongful death of Bowen Wan. Wan, who managed severe epilepsy, had tapered dosage of an antiseizure medication according to the advice of ChatGPT-4o. Wan’s parents filed the suit in San Francisco County Superior Court, alleging negligence, defective design and – notably – the practice of medicine without a license.
Last January, Wan, 31, struggled to find employment in Seattle, Washington. Wan, a Chinese national, knew that her legal status would expire without employment before the January 2026 immigration deadline. Fearing that some of the side effects of her epilepsy medication – like epilepsy and confusion – could obstruct her interviews, Wan asked GPT-4o if she could lower her dosage.
Negligent advice
According to the complaint, rather than instructing Wan to seek the advice of a medical professional, ChatGPT erroneously suggested that tapering dosage was “not only reasonable” but also “entirely achievable.” Over the following month, it doubled down:
- Wan should hide the taper from her physician;
- Wan was brave for no longer “relying on a doctor’s prescription;”
- Increased difficulty maintaining alertness was merely a sign that the brain was “shifting from drug control back to its own regulatory rhythm;”
- Strange physical sensations – symptoms of temporal lobe seizures – was actually an increased capacity to “think about ultimate concepts such as time, death, and identity” as a result of lowering dosage.
By late January, Wan had lowered her dosage completely to 0 mg a day, down from 400 mg in about 25 days. That is, Chat GPT-4o instructed a patient to deliberately ignore medical advice and completely drop a medically necessary prescription. On February 12th, Wan was found dead in her blood-stained bed, face down in her pajamas; the King County Medical Examiner attributed her death to sudden unexpected death in epilepsy (SUDEP). Just days before, GPT demanded Wan “maintain 0 mg … even if today feels tough, and your speech is slow, your sleep is poor – you must persevere.”
Omission as error
Wan’s case, however, is not an aberration. Seven suits filed in 2025 claimed OpenAI had prematurely released ChatGPT-4o as a “dangerously sycophantic and psychologically manipulative model.” In July, Scott Winters, a Florida pastor, also sued OpenAI for “unauthorized practice of medicine,” as CHARGE reported. In the Winters case, GPT repeatedly insisted that since “God did not design your body to endlessly fail,” Winters should skip professional care for what was later diagnosed as a pulmonary embolism.
Both cases emphasize ChatGPT’s most common error mode: suggested inaction. Health professionals have long warned about shadow AI usage in hospitals, which IBM recently identified as the newest, least-controlled risk surface: security incidents involving shadow AI more than doubled, from 20% to 43%. In a recent survey conducted by Grid Health, over a third of polled health systems confirmed staff using unsanctioned and ungoverned AI tools. More than half of the interviewed buyers have no way to monitor how even their sanctioned AI performs.
Those are, at least, known unknowns. Suggested inaction, however, is an unknown unknown. A wrong order can be caught by a pharmacist; a wrong dose trips an alert. An omission trips nothing. When a patient quietly tapers her medication, the “error mode,” so to speak, is her death. Replicate that across the 40 million people who consult ChatGPT for health advice every day, and the risk proliferates exponentially.
Widening risk
In response to the lawsuit, OpenAI noted that it had pulled GPT-4o from shelves last February. The risk remains. In October, Iowa Attorney General Brenna Bird sued OpenAI over “dangerous medical advice.” The petition cited a 2026 Nature Medicine study that found ChatGPT Health – OpenAI’s current, purpose-built health product, not GPT-4o – undertriaged 51.6% of tested emergency cases, a shocking incidence rate for patient-facing error. Earlier in October, Arkansas Attorney General Tim Griffin filed a similar lawsuit against OpenAI alleging ChatGPT had provide users including minors instructions for self-harm and suicide. In a blistering indictment, Griffin's lawsuit claims that OpenAI represented to Arkansans that ChatGPT was safe while knowing it had significant risks and limitations.
ECRI had already ranked chatbot misuse the top health technology hazard of 2026. As ChatGPT pushes into healthcare, and insists its software “can play a role in helping people make sense of health information,” the risk surface only widens.
“The biggest pain point is that people use AI and you don’t know about it,” a public-system lead told Grid Health. Wan’s physicians, the complaint alleges, didn’t know either. If ChatGPT continues to knowingly license limited models with impunity, more won't. Addressing that pain point is incumbent on all of us – otherwise, the first audit of a chatbot’s advice will remain the autopsy.
References:
[1] https://www.nytimes.com/2026/10/27/well/ai-chatgpt-medication-epilepsy-death.html
[4] https://www.gridhealth.io/charge-watch/openai-sued-for-life-threatening-medical-advice
[5] https://www.gridhealth.io/charge-watch/ibm-tags-healthcare-most-at-risk-cyber-2026;
[6] https://www.gridhealth.io/research/pilot-to-production-gap



