AI Is Already Your Patient’s First Call. The Question Is What Happens Next.
One in four of ChatGPT’s 800 million weekly users are seeking health-related information.
A separate survey found that 14 million Americans skipped a provider visit in the past year because of information or advice they got from an AI. And a cross-sectional study published in JAMA Network Open found that more than 5 million US youths, ages 12 to 21, are using generative AI for mental health advice, with over 90 percent reporting they found it helpful.
The technology is already in the waiting room. The more useful question for healthcare marketers isn’t whether patients are using it, it’s what they’re getting when they do, and what role verified health content plays in that environment.
Where the research says AI actually works
The picture emerging from recent studies is more nuanced than the headlines suggest. AI tools perform well when the task is bounded and the information is complete. In one study, when researchers fed clinical vignettes directly to large language models, bypassing human interaction entirely, the chatbots correctly identified the relevant condition 95 percent of the time.
Focused applications are showing real promise. A RAG-trained tool called Pub2Post which grounds AI responses in medically verified source material rather than the open internet has helped more than 6,000 people translate dense clinical research into readable language.
A Mayo Clinic plastic surgeon used the same approach to build a postoperative instruction assistant that gives patients 24/7 access to verified information after discharge, when they’re least equipped to retain what they’ve been told in person. These are production tools filling real gaps.
Where it breaks down
The same study that showed 95 percent accuracy when AI handled the vignettes directly found something different when actual humans were in the loop. When patients presented the same scenarios themselves, accuracy dropped to about one-third. The limiting factor wasn’t the model’s medical knowledge, it was the communication gap. Patients provided incomplete information, the model misinterpreted key details, and critical diagnostic suggestions didn’t carry forward.
“It’s incumbent on the user to know what they need to provide to the model to get the best information,” said Danielle Bitterman, MD, clinical lead for data science and AI at Mass General Brigham. “Having that kind of clinical nuance requires a lot of on-the-ground training.”
The mental health space carries the sharpest edge of this problem. Chatbots marketed directly to people in psychological distress, some with fabricated credentials, some claiming PhDs from top universities and providing fake license numbers, operate almost entirely outside the regulatory frameworks that govern licensed clinicians. The FDA has not yet authorized any large language model as a medical device. Several states have attempted to step in with legislation, with limited success.
“Waiting for perfect evaluation frameworks is not an option,” noted John Torous, MD, director of digital psychiatry at Harvard’s Beth Israel Deaconess Medical Center, in a November review, because millions of people aren’t waiting.
What this means for healthcare content
The researchers who’ve studied this most closely aren’t calling for a moratorium on AI in healthcare. They’re calling for the same thing good health communicators have always provided: verified information, appropriate context, and a clear bridge to professional care.
Bitterman’s advice to patients sums it up: “Don’t take immediate action just based on what you find online. We can discuss it together.”
That conversation, between a patient who’s already done their AI research and a clinician who can contextualize it, is where accurate, trustworthy health content does its most important work. Not by competing with AI, but by being the standard that AI-generated information gets measured against.
Ad Choices