AI Chatbots Failed to Urge Emergency Care in About 41% of Vietnamese Scenarios Versus 15% in English

4,000 Brigham and Women’s Nurses Postpone Wednesday Strike to Vote on Hospital’s Best and Final Contract OfferBrigham and Women’s nurses postponed the Oct. 14 strike to vote on a final offer. See the pay, insurance, and staffing terms, and what patients face. Parents Should Check Amazon Bath Seats and Dressers, Giant Bike Lock Keys, and Avengers Medals in New RecallsNew CPSC recalls cover Amazon baby bath seats and dressers, Giant LED bike lock keys, and Avengers medals. See what you own and how to get refunds. Chicago Marathon’s Record Field of More Than 55,000 Faces a Yellow-Flag Day with Highs Near 71 DegreesThe Chicago Marathon starts under a yellow flag with a forecast high near 71, above the 65-degree normal. Heat illness signs and when to stop. Records Detail Immediate Jeopardy Citation After Charlotte Nursing Home Resident in Wheelchair Was Fatally Struck in 2023Records show inspectors cited a Charlotte nursing home at the immediate jeopardy level after a resident in a wheelchair was fatally struck on a busy road. Mental Health The Brain’s So-Called Christmas Spirit Network Lit Up in 10 Holiday Lovers, but Red Pictures Might Explain ItAn fMRI study found five brain regions responded more to Christmas images in people who celebrate, but the authors admitted red may be to blame. Frequent Mental Distress in Adults 18 to 29 Climbed From 12.7% to 21.3% Since 2012 Despite Recent DeclinesCDC data show 1 in 5 adults 18 to 29 report frequent mental distress, up from 2012 but down two years straight. Where to find help today. A Strong Smell Could Set Off His Sneezing Fits, but His Muscles Told a Stranger StoryStrong smells set off a teacher’s sneeze-like fits for nine years. Muscle recordings showed rapid jerks unlike a textbook sneeze reflex. Louisiana Finds New Synthetic Opioid Cychlorphine in at Least Two Deaths, Warns Extra Naloxone Doses May Be NeededLouisiana found cychlorphine in at least two overdose deaths since December 2025. See where the opioid has spread and why naloxone may need repeating. Innovation AI Chatbots Failed to Urge Emergency Care in About 41% of Vietnamese Scenarios Versus 15% in EnglishA simulated test presented at ACG 2026 found that AI chatbots missed urgent GI warnings far more often in Vietnamese. What families should know. AI Tool Helped Neurology Residents Tell Epilepsy From Functional Seizures Correctly Up to 89% of the TimeA University of Minnesota AI tool reads clinical notes to help doctors tell epilepsy from functional seizures. What it found and why it isn’t ready. A Medical Paper Called a 37-Year-Old Study Recent, and Nearly 1 in 5 Recent Citations Were at Least a Decade OldA BMJ analysis of 1,000 papers found 17.7% of references labeled recent were at least 10 years old, and some dated back as far as 37 years. No Measles, Mumps or Chickenpox Cases Followed Accidental Live Vaccines in 220 IBD Patients on Immune-Modifying DrugsNo measles, mumps, or chickenpox cases followed accidental live vaccine administration in 220 IBD patients on immunosuppressive drugs. Why live shots are still avoided. Healthy Living Isaias Outages Left Gulf Coast Fridges Warm for Days: Which Groceries to Toss and Which to KeepAfter Isaias, Gulf Coast families face costly fridge decisions. See which foods USDA charts say to toss, which can stay, and where help may exist. Only 40% of Americans Have Strong Confidence in Hospitals and Clinics While 52% Trust Science a LotNew Wellcome data show Americans trust science more than most nations but doubt hospitals. What patients say drives the gap and what can help. San Francisco Expands Doxy-PEP to Women and Others as Officials Report 24% Drop in Syphilis CasesSan Francisco now offers the after-sex antibiotic doxy-PEP to more people as syphilis and chlamydia rates fall. What it is, who qualifies, and the risks. Queens Pharmacy Owner Arrested Trying to Board One-Way Toronto Flight in Alleged $31 Million HIV Medicaid FraudA Jackson Heights pharmacy owner faces fraud charges over HIV drug kickbacks and buybacks. What patients and taxpayers need to know about the case. Covid – 19 FDA Proposes Requiring Companies to Report Food Additives That Now Reach Grocery Shelves Without Federal ReviewFDA has released its proposed GRAS rule requiring companies to notify the agency about food additives. Here is what the proposal does and does not do. Federal Regulators Warn Some Hair Dryer Brushes Lack the Part That Stops ElectrocutionCPSC warns hair dryer brushes sold on Walmart.com and TikTok Shop lack immersion protection. The seller has not agreed to a recall or refund. A Recalled Teething Toy Broke the Pattern Parents Were Told to Look ForA newly recalled pull-string teething toy has a different model number and design from earlier recalls, and one choking incident has been reported. Trying to Lower Blood Sugar, Cholesterol or Blood Pressure? Different Workouts Suited Different MarkersA pooled analysis of 53 trials found that different exercise types helped different metabolic markers. Most of the evidence was low certainty. PHILIPPE LOPEZ | AFP via Getty Images Share on Twitter Share on Facebook Share on Pocket Artificial intelligence chatbots were far less likely to tell someone to seek emergency care when the same medical emergency was described in Vietnamese instead of English, according to research presented at the American College of Gastroenterology’s 2026 meeting in Nashville. The tools failed to recommend emergency care about 41% of the time in Vietnamese, compared with 15% in English.

The findings come from a conference abstract that tested simulated digestive-health scenarios. It did not involve real patients, and it has not been published as a full peer-reviewed study.

For a parent deciding whether a child’s stomach pain can wait until morning, or an adult caring for an older relative, chatbot advice can shape what happens next. A tool that sounds confident but misses a warning sign could delay care. The researchers said the results raise concerns about delays in seeking care, a risk the study measured in test scenarios rather than in real patients.

The study, titled “Does AI Speak Gastroenterology Equitably?” and presented as Oral 62, was among ACG’s newsworthy abstracts for the 2026 meeting. The meeting runs Oct. 9 to 14. Researchers tested four large language models across the seven most common languages spoken in the United States.

According to the authors’ summary, the same clinical scenario could receive different advice depending on the language used. Failure to recommend emergency care rose from 15% in English to about 41% in Vietnamese. The researchers also found errors in English and larger gaps in several other languages. The ACG summary did not name the four models or report results for each language.

“A tool’s ability to respond fluently in a language does not mean its medical advice is reliable,” said Harika Dadigiri, MD, the abstract’s listed author contact. “Language equity is part of patient safety.”

Dadigiri said AI is already part of how patients look for health information. She encouraged clinicians to ask patients what advice a chatbot gave them, and in which language, to help catch misunderstandings.

Vietnamese Americans are the fourth-largest Asian-origin group in the country, according to the Pew Research Center. About 72% of Vietnamese Americans ages 5 and older speak Vietnamese at home, and about 56% speak English proficiently, which means many may prefer health information in Vietnamese.

Those families are concentrated in a few major metro areas. A 2013 Census Bureau language report, based on 2007 to 2011 survey data, found about 233,000 Vietnamese speakers in the Los Angeles area, along with roughly 89,000 in Houston and 61,000 in Dallas-Fort Worth. Those figures are older, but they show where language gaps in AI tools could reach the most households.

The risk is not limited to Vietnamese speakers. The researchers said gaps appeared in several of the seven languages, and even English answers were not error-free.

This is early evidence. The study is a conference abstract based on simulated scenarios, not patient outcomes, and it has not undergone full peer review. It does not show that anyone was harmed or that patients actually delayed care because of chatbot advice.

The results also may not apply to every chatbot or every version of a model, since AI tools are updated often. The study focused on digestive-health emergencies, so it is unclear whether the same gaps appear for heart, stroke, or pediatric scenarios.

What the study does show is a measurable difference in how the tools handled identical emergencies across languages. The question extends to hospitals and health plans adding chatbots to patient portals and phone lines. If those tools are offered in several languages, testing emergency advice in each language would help show whether every patient gets the same safety net.

The researchers advised patients not to rely on a chatbot alone to decide whether concerning symptoms can wait. Tell your doctor or nurse what the chatbot said so they can correct any misunderstanding, and ask for an interpreter when you need one.

Some symptoms need urgent care regardless of what an app says. These include vomiting blood, black or bloody stools, severe or worsening abdominal pain, chest pain, trouble breathing, fainting or sudden confusion. Call 911 or go to an emergency department if these occur. For milder symptoms, a nurse advice line or a clinician’s after-hours service is a safer first step than a chatbot alone.

Language help is also a legal right in many care settings. Under federal civil rights law on language access, programs that receive federal health funding must provide language assistance, including interpreters and translated documents, free of charge. Many hospitals, clinics, and pharmacies receive such funding, so patients can ask for an interpreter in Vietnamese or another language at no cost.

Outside researchers will need to see the full methods and model-by-model results. ACG meeting sessions will be available on demand through March 31, 2027, and MedicalDaily will report if a full peer-reviewed paper is published.

What did the study find? In simulated digestive-health emergencies, four AI chatbots failed to recommend emergency care about 41% of the time in Vietnamese, compared with 15% in English. Gaps also appeared in several other languages.

Was this tested on real patients? No. Researchers used simulated scenarios. The work is a conference abstract presented at ACG 2026 and has not been published as a full peer-reviewed study.

Which chatbots were tested? The ACG’s public summary says four large language models were tested but does not name them.

Who could be most affected? People who prefer health information in a language other than English, including many Vietnamese-speaking families in areas such as Los Angeles, Houston, and Dallas-Fort Worth.

Which symptoms need emergency care no matter what a chatbot says? Vomiting blood, black or bloody stools, severe abdominal pain, chest pain, trouble breathing, fainting or sudden confusion all need urgent medical attention.

Can I get a free interpreter at a hospital or pharmacy? Programs that receive federal health funding must provide free language help. Ask the front desk or pharmacist for an interpreter.