Dr. Google, meet Dr. GPT: Tips on using AI for medical advice for children

It’s 2 a.m., and your 4-year-old just woke up again. This time they’re throwing up, and now they have a high fever. Should you go to the ER? Right away, the parent math begins: a sleepless night, a 20-minute drive, hours in a waiting room, another co-pay, and, of course, your growing worry for your child. Can AI help you figure out what to do? Should it?

Maybe you’re like me and have used AI to compare vacation spots or to figure out how to remove a Sharpie stain from a shirt. But for decisions about my own kids’ health – even as a pediatrics resident who cares for children every day – I still wonder what role these tools should play.

More and more parents are turning to AI for health answers. But how good are those answers? A widely cited 2024 study in JAMA Pediatrics found that a popular chatbot got the diagnosis wrong in about 83% of pediatric cases. Similar studies have called for more expert oversight to safeguard the health information these tools give families. Since this time, however, the capabilities of AI have improved and continue to improve. New studies are needed to benchmark the accuracy of the latest AI models.

I’ve spent time working with these new AI models and helping to train them. Often, to better train AI models, you probe until you find weak spots and work to improve those weaknesses. In doing this work, and as an avid AI user myself, a few things stand out to me. Newer models are incredible: they synthesize huge amounts of information, catch what gets missed, and distill a complex topic to its key point. But they also hallucinate – sounding completely sure while making things up, even inventing sources and links to sites that don’t exist. And they tend to be overly reassuring, which can give you a false sense of confidence.

These drawbacks matter even more in pediatrics. A recent Pediatrics article, from the American Academy of Pediatrics, weighs AI’s pros and cons in children’s care: AI doesn’t grasp human emotion or the give-and-take between family and doctor, and, unlike a clinician, it has no empathy, ethical judgment or feel for a child’s story. Training data may also skew toward adult health, yielding recommendations that are wrong for children. Privacy is a concern too, and the regulations surrounding health-related AI are still an evolving patchwork.

So, when is it okay to use AI for your child’s health? It depends.

Tips to consider

Probably okay to use AI for:

• General, non-urgent parenting questions such as milestones, picky eating, and sleep routines to use for further discussion with your pediatrician.
• Translating health information into a preferred language or into simpler words.
• Organizing thoughts or building a list of questions to ask at the next clinic visit.
• Linking to and summarizing or synthesizing trusted sources like HealthyChildren.org, AAP.org, and MedlinePlus.org.

Probably not okay to use AI for:

• Deciding whether something counts as an emergency, such as a fever in a baby under 3 months, a child who is hard to wake, a seizure, trouble breathing, blue lips, dehydration, or trauma.
• When the medical concern involves a baby or a child with complex medical needs. These are populations where trained professionals should evaluate without delay, as subtle changes in health can have more immediate and serious consequences.
• A health situation involving complex social or ethical nuances (not a strength for AI). To help in these situations, we can rely on strong and capable professionals such as social workers, care coordinators, chaplains, or other medical professionals.
• A mental health or safety crisis. If a child or teen may be in danger, poses a credible danger to others, or is talking about self-harm, reach a trained professional or a crisis line, not a chatbot.
• Anything that needs to be seen. A chatbot can’t examine your child; describing a rash or a limp isn’t the same as a healthcare professional looking at it.

Generative AI is not a replacement for a human caregiver. They can be helpful, even amazing tools in many areas, and I believe it should continue to have a growing role as a supplemental tool in healthcare. However, when it comes to important decisions about your child’s health, it’s important to remember that AI models can hallucinate, rely on false or outdated information, give overconfident answers, or miss important and nuanced ethical or social context. The key, I believe, is to check what they generate with trusted, trained medical professionals and reliable medical knowledge and not lean too heavily on them – especially when a child’s health is in the balance.

So, what should you do at 2 a.m. with a sick 4-year-old? Here’s one thing you can do right now, before you ever need it: find an after-hours nurse line for your clinic or nearby hospital and save the number in your phone. A calm human voice who knows children beats a confident chatbot at 2 a.m. every time.

By Conner Earl, M.D., Ph.D., father of two and first-year pediatrics resident