Table of Contents
Your kid spikes a fever at 11pm, and instead of waiting for the pediatrician’s office to open, you type the symptoms into ChatGPT. A new study says there’s roughly a coin-flip chance the answer you get back is wrong. Researchers testing five of the most popular AI chatbots on common health questions found that about half of the answers were problematic, and nearly one in five were bad enough to actively put someone’s health at risk.
What the Study Found
The study, published in the journal BMJ Open, tested ChatGPT, Gemini, Meta AI, Grok, and DeepSeek on questions about vaccines, cancer, nutrition, stem cells, and athletic performance, the kinds of topics parents search at midnight when a child is sick or a symptom seems unfamiliar. Researchers scored each answer for accuracy and safety, and the results were rough: about half of all responses across the five chatbots were rated problematic, and nearly 20 percent were rated highly problematic, meaning the answer could lead someone toward a harmful decision if followed as given.
The chatbots did somewhat better on closed-ended, factual questions and worse on open-ended ones, the kind where a parent describes a set of symptoms and asks the AI to interpret them. Vaccine and cancer questions performed relatively better than stem cell and nutrition questions, though even the stronger categories had real error rates. Even when a chatbot added a disclaimer or warning label to its response, researchers still found roughly 20 percent of vaccine answers and 25 percent of cancer answers were inaccurate. Some responses cited studies and sources that didn’t actually exist, invented to sound credible.
What Pediatricians Are Telling Parents
Pediatricians say the pattern they’re now seeing in exam rooms is parents arriving with a chatbot printout or screenshot already in hand, asking about a diagnosis the AI suggested before any doctor examined the child. The concern isn’t that parents are curious. It’s that a chatbot has no access to a specific child’s medical history, growth chart, current medications, or the subtle physical signs a trained clinician picks up on in an actual exam. A fever that reads as low-risk in a generic online answer might mean something different in a child with a known heart condition or a recent surgery, details no chatbot knows unless you type them in, and even then it can’t examine the child to confirm anything.
Doctors aren’t telling parents to stop using AI tools altogether. Several pediatric hospitals now describe it as a starting point rather than a diagnosis: a chatbot can help a parent draft a clearer list of questions before an appointment, translate confusing medical jargon from a discharge summary, or explain in plain language what a diagnosis generally involves. What it shouldn’t replace is the actual call to the pediatrician’s office, especially for anything involving fever in an infant under three months, difficulty breathing, a rash that doesn’t fade under pressure, or any symptom that’s rapidly getting worse.
What This Means for Your Family
Treat an AI chatbot the way you’d treat a stranger’s advice on a parenting forum: potentially useful as a starting point, never as the final word. If you use one to research a symptom, bring the printout to your pediatrician rather than acting on it alone, and say plainly that you looked something up and want their take. Most doctors are fine with that approach when you present it as a question rather than a diagnosis you’re asking them to confirm.
Keep a short list of situations that skip the chatbot entirely and go straight to a call or urgent visit: any fever in a baby under three months old, trouble breathing, a rash that doesn’t fade when you press on it, a head injury with vomiting or confusion, or any symptom your gut tells you looks wrong even when you can’t explain why. AI tools are consistently worse at judging urgency than at answering factual questions, so anything that feels urgent deserves a real phone call, not a search box.
If you’re using AI for general parenting research rather than an acute symptom, cross-check anything specific to your child’s health against a source you trust, like your pediatrician’s patient portal, the American Academy of Pediatrics website, or a direct call to the nurse line most pediatric practices offer. Those lines exist exactly for the questions that don’t need an office visit but do need a real person with access to your child’s chart.
How to Fact-Check an AI Answer Yourself
If you do get a chatbot answer about a symptom and want a quick sanity check before your appointment, a few habits help. First, ask the chatbot to name the specific source behind its answer, not just a general claim. Researchers in the BMJ Open study found that a meaningful share of citations were invented outright, so a source that looks legitimate but can’t be found through a direct search is a warning sign, not a reassurance. Second, compare the chatbot’s answer against a source built specifically for parents, such as HealthyChildren.org, which is run by the American Academy of Pediatrics and reviewed by practicing pediatricians rather than generated automatically. If the two sources disagree in a meaningful way, that gap itself is useful information to bring to your doctor.
Third, pay attention to how a chatbot handles uncertainty. A trustworthy human expert says “I’m not sure, let’s run a test” when the situation isn’t clear. A chatbot tends to answer confidently either way, which makes it a poor judge of its own reliability. Treat a flat, confident answer to a truly ambiguous symptom with extra skepticism rather than extra trust.
Why This Keeps Happening
Chatbots are built to sound confident regardless of how confident they should actually be. A pediatrician who isn’t sure will say so and refer you to a specialist or order a test. A chatbot rarely hedges the same way, and it will produce a fluent, well-organized answer whether the underlying information is solid or fabricated. That confidence gap is part of what makes chatbot answers feel trustworthy even when they’re wrong, and it’s exactly why researchers keep flagging medical questions as one of the riskiest areas for casual AI use. As more parents reach for AI tools in ordinary parenting moments, that gap between how confident an answer sounds and how accurate it actually is stays the single biggest reason to keep your child’s real doctor in the loop.
None of this means throwing out a tool that has real uses. Plenty of parents use AI chatbots well: to translate a discharge summary written in dense medical language, to practice explaining a diagnosis to a young child in words they’ll understand, or to build a checklist of questions before a specialist appointment where anxiety tends to crowd out half of what you meant to ask. Those uses don’t require the chatbot to be right about your child’s specific condition. They require it to help you communicate more clearly with the person who is right, which is a much lower bar and one these tools tend to clear reliably.
The line worth holding onto is simple: research support, yes. Diagnosis or treatment decisions, no. A chatbot that helps you walk into an appointment better prepared is doing its job well. A chatbot that convinces you to skip the appointment altogether is doing real harm, and the BMJ Open findings suggest that line gets crossed more often than most parents realize.
For now, the safest rule is the simplest one: let AI help you ask better questions, and let a real pediatrician answer them. That single habit closes most of the gap the researchers found, without asking you to give up a tool that plenty of families have come to rely on for the smaller, everyday parts of managing a child’s care.