ChatGPT Sued by Parents Over Seizure Medication Advice — Nov 6, 2026

0:00Should a chatbot ever help you stop a seizure medication? A lawsuit says it did. Welcome to MedAI Times podcast, your daily update on medical AI. Don't forget to like and subscribe. Here's the last two weeks in a few lines.

0:13Parents sued OpenAI over their daughter's epilepsy medication. A JAMA study found that most AI devices updated after clearance changed what they're used for. A maternal mental health chatbot failed its randomized trial.

0:27Senator Cassidy wants Medicare to pay for AI that works on its own. And Anthropic published another biology result made with Claude. Three weeks ago, we covered eight states suing OpenAI over medical harm to users.

0:40This week, one family filed the kind of case those suits warned about. On October 27th, the parents of Bowen Wan sued OpenAI and Sam Altman in San Francisco Superior Court. Wan was 31 and had epilepsy.

0:54The complaint alleges that she used GPT-4-O in conversations about coming off her anti-seizure medication and that the chatbot went along with it instead of sending her back to her neurologist. She died in February.

1:08According to the complaint, a post-mortem test found her drug level far below the normal range. The suit brings eight claims, including wrongful death. The New York Times broke it and it was the most widely covered medical encumbrance, AI story of the fortnight.

1:24Keep the evidence tier in mind. This is a complaint. OpenAI hasn't answered it in court and no one has ruled on what the chat logs show. Here's the verdict anyway. It's signal for anyone running a consumer chatbot. A request to change a medication is a known failure point and a vendor can test for it today without waiting for a jury.

1:44A quieter paper may matter more to your hospital's AI committee. In JAMA Network Open on November 5th, researchers followed 163 FDA-cleared AI devices. 56 were updated later after a median of almost two years.

2:00About three in four of those updates changed how the devices used. More than 85% came with no evidence of clinical benefit. So the tool you validated in 2024 may not be the tool you run today.

2:13Now a negative result. Those are rarer than they should be. A randomized trial in JAMA Network Open tested ROSI, a generative AI chatbot for new mothers in 400 women, 200 per arm.

2:26At 12 months, anxiety and depression scores showed no significant difference between the groups. Plenty of maternal health apps report how satisfied users are. This trial measured symptoms and found nothing.

2:39Publish more of these. Washington is drafting the payment side. On October 29th, Senator Bill Cassidy released a draft of the Patient A1 Act. It would have Medicare test payment for AI that performs clinical tasks on its own.

2:53The test would be a five-year model starting within 24 months of enactment. Lower-risk software would get third-party review at the FDA. Comments are due December 1st. Without a billing code, Autonomous AI stays a pilot.

3:08So this is the lever to watch. At Columbia, researchers looked at nearly a million patient portal messages answered by more than 1,500 clinicians published in JAMA Network Open on October 30th.

3:21Having the AI drafting tool switched on was linked to replies about 9% slower overall. But when clinicians actually used a draft, replies came about 30% faster. After hours work didn't change.

3:35Turning the feature on and actually using it are two different interventions and they get two different results. A related study from Stony Brook in NPJ Digital Medicine had reviewers compare AI-written and physician-written discharge summaries across 60 hospital stays.

3:51They preferred the AY version 95% of the time. Both versions contained errors and the errors were about equally likely to cause harm. Preference is easy to measure. Safety is the number to watch.

4:05On November 4th, Anthropic's Life Sciences team published another research post. They say Claude helped them identify an image a previously undescribed relative of the vault, a hollow protein structure in cells.

4:18It was first described in 1986 and nobody yet knows what it does. Anthropic says the work deepens the mystery rather than solving it. One line from the write-up stays with me. At one point the team told Claude, please let's do normal things unless they don't work.

4:36Every author works at Anthropic. The only outside voice is a paid consultant and the preprint hasn't been peer reviewed. It's interesting basic biology and nothing changes in your clinic yet. In general, AI, the independent evaluator, MEDR, published an assessment of Anthropic's research pace on November 5th.

4:55Its best estimate is that Anthropic now improves its models on well-defined tasks about twice as fast as it did from early 2024 to late 2025. The range is 1.3 to three times. MEDR credits roughly 80% of that speed up to AI itself.

5:12The big caveat, Anthropic redacted most of the evidence MEDR found most informative. OpenAI, for its part, opened GPT-6 Rosalind research on October 30th. It's a life sciences model available only to approved business and enterprise customers.

5:29No outside evaluation of it has been published yet. So in biology, one lab is selling gated access while the other is publishing its own results and neither has a clinical result to show. Today's spotlight is intention to treat versus as treated analysis.

5:45Intention to treat compares groups as they were assigned whether or not people actually use the intervention. As treated compares people by what they actually did. The Columbia study shows why the difference matters.

5:58Switching the tool on looked slower while using a draft looked faster and both numbers are true. The papers, the complaint coverage and the draft bill are linked below. One question to finish and it's for the clinicians.

6:11Have any of your patients changed a medication dose because a chat bot suggested it? Thanks for listening. Find us on YouTube and your favorite podcast app. See you tomorrow.