ChatGPT Sued by 8 States Over Kids and Medical Advice — Oct 16, 2026

0:00Eight states just took ChatGPT to court, and medical advice is in the complaint. Welcome to MedAI Times Podcast, your daily update on medical AI. Don't forget to like and subscribe. Eight states sue OpenAI over what ChatGPT tells kids.

0:14The FTC sends 50-page demands to OpenAI and Anthropic. Chatbots miss emergencies twice as often in Chinese and Vietnamese. Penn patients text their brain monitors. Google's AI agent goes looking for a new antibiotic.

0:27A $1.2 billion bet on AI-designed proteins. And open evidence lands inside two more big health systems. On Wednesday, we covered Vanderbilt psychiatrists finding chatbot-linked psychosis in their own patients' charts.

0:40This week, eight attorneys general put that kind of harm in a lawsuit. On October 14, Arizona, Arkansas, Idaho, Iowa, Kansas, Montana, Nebraska, and West Virginia each sued OpenAI in their own state courts.

0:55The legal hook is consumer protection, not medical regulation. The states say ChatGPT was sold as safe for young users, while the company failed to warn about harmful medical guidance, chatbots reinforcing paranoia and delusions, and unhealthy attachment.

1:09The complaints cite cases where ChatGPT allegedly coached children on suicide, and one where wrong medical information allegedly led to an overdose. They want changes to how ChatGPT treats minors, the profits made from young users handed back, and civil penalties.

1:26Arkansas also asks a court to pause testing of advanced models until safeguards exist. The same Wednesday, the Federal Trade Commission sent OpenAI and Anthropic civil investigative demands, each over 50 pages, asking for records on safety incidents, safeguards, marketing, and privacy.

1:44The FTC confirmed sending them. My verdict. Signal for policy, not yet for evidence. These are allegations, not findings. But a general chatbot giving health advice sits outside FDA review entirely, and consumer protection law is now how states are trying to reach it.

1:59At the American College of Gastroenterology meeting in Nashville, Dr. Harika Dadagiri of New York Medical College showed what happens when the same 50 patient questions about gut symptoms are machine translated into six languages.

2:14Four small, cheap models answered each question three times. In English, 91% of answers matched guidelines. In other languages, that fell as low as 64%. The part that should worry clinicians is emergencies.

2:30In English, 15% of answers failed to tell the patient to seek urgent care when they should have. In Chinese, it was 37%. In Vietnamese, 41%. It's a conference abstract with scoring by rubric, and these were not frontier models.

2:46Your patients don't use frontier models either. In Nature Medicine, on October 13th, a Penn team led by neurologist Brian Litt let patients text with the brainwave monitor recording them. Thirteen adults in Penn's epilepsy monitoring unit exchanged 1,307 messages with the system, asking about their recordings, logging mood and sleep, and telling it which seizure alerts

3:11were wrong. That feedback improved the detector's seizure calls, though the press release gives no figure. Thirteen inpatients is a proof of concept. The team wants to try insulin pumps and cardiac devices next.

3:24On October 15th, Google DeepMind published a walkthrough of one researcher and an AI agent working in Google's antigravity platform, hunting for antimicrobial peptides in a bacterium called Streptococcus pseudopneumoniae.

3:38In a single session, the agent filtered protein databases, checked novelty, pulled alpha-fold structures, and drafted a lab plan. The lead is a 97-amino acid bacteriocin with no close match among reviewed proteins, predicted to block bacterial cell division.

3:55Predicted is the key word. No lab result exists yet. It's a promising antibiotic that has not, so far, met a bacterium. Read it as a workflow demo, not a discovery. Danish drugmaker Zeeland Pharma signed with Generate Biomedicines on October 13th to build AI-designed proteins that hit several targets at once for obesity and metabolic disease.

4:16$30 million up front, up to $35 million in early milestones, and nearly $1.2 billion in total milestones. Two programs, with an option for a third. No candidate exists yet.

4:29Only the $30 million is guaranteed. The rest depends on molecules that haven't been designed. Open Evidence kept signing hospitals. On October 12th, Northwestern Medicine said its clinicians can now launch Open Evidence from inside Epic with their hospital login, and nurses and pharmacists get it, too.

4:48A day later, Johns Hopkins Medicine announced a deal covering education and clinical care. Penn and Advent Health signed in the past month. The tool is now inside the chart at several academic centers. Nobody has published outcome data from those deployments.

5:04UnitedHealth used its October 13th earnings call to talk about AI on the insurer's side of the chart. Chief Executive Stephen Hemsley said AI chart summaries now reach more than 40 percent of the company's clinicians, and claimed appeal cycle times got 26 times shorter.

5:22He gave no method. Clinicians will meet this as faster prior authorization decisions and faster denials, too. Which one they get more of, the call didn't say. Meta released MuseSpark Open on October 15th, its first open-weight model since Lama.

5:38816 billion parameters, 42 billion active per token, a 1 million token context, and it runs on two Mac studios. The license requires a built-with label and a separate deal for big commercial API providers.

5:52For hospitals that want frontier-ish models on their own hardware, that's a real option. New from Anthropic, on October 15th, six staff from the evaluation group METR now work inside the company with employee-like access, and METR can publish its risk assessments without an Anthropic veto, apart from narrow redactions.

6:12SecureBio is reviewing the chemistry and biology chapter of its latest risk report. Coming the same week, the FTC asked for safety records. Outside reviewers who can publish freely are worth more than any internal promise.

6:25This week's spotlight is human-in-the-loop label feedback, also called active learning. Instead of labeling thousands of examples up front, the model asks a human only about the cases it is least sure of, and learns from each answer.

6:40The Penn Epilepsy study turns the patient into that human, which is cheap and sensible, and also means the label is only as good as the patient's memory. Every study, filing, and press release from this week is linked in the description, with its evidence level marked.

6:56Has a patient in your clinic acted on CHAT GPT medical advice before calling you this year? Thanks for listening. Find us on YouTube and your favorite podcast app. See you tomorrow.