Μake us preferred on Google If patients turning to the internet for medical advice wasn’t enough, doctors now have a new, more powerful version of “Dr. Google” to contend with.
The majority of American adults today view their test results online before talking to a doctor, research shows. That’s thanks in part to a legal provision that took effect in 2021 to help patients get near-real-time access to health information , such as lab reports, imaging results and clinical notes.
A growing share of those patients are uploading their test results straight into AI chatbots , leaving some medical professionals concerned about hallucinations, misinterpretations and privacy issues.
“We’re trying to refute what the patients have convinced themselves that they have based on whatever they’ve searched and reviewed,” said Dr. Jason Goldman, an internist in Coral Springs, Fla. “You can get everything from, ‘You have a cold,’ to ‘You’re dying of cancer.’”
Patients have used conversations with chatbots to justify stopping lifesaving cholesterol medication, Goldman said. Others have questioned recommended vaccines because AI validated their hesitancy, he added.
“I’ll give them a recommendation and they’ll say, ‘Let me go ask AI,’ and I’m like, ‘Then why are you here?’” Goldman said.
Roughly a third of adults turn to AI chatbots for health advice, and 19% use AI specifically to interpret lab results or medical tests, according to a 2026 poll from health policy research nonprofit KFF.
OpenAI said more than 300 million people ask health-related questions on ChatGPT each week.
Companies are eager to cash in on the interest. Anthropic’s Claude can connect to some users’ medical records, and Epic Systems, the electronic medical-records behemoth, has an AI assistant built into its software.
In July, OpenAI rolled out widespread access to a health feature that can link with users’ medical records and wearables data, and among other things tell users about trends in their health results over time. The model can “reason better than clinicians,” the company’s VP of health products, Ashley Alexander, said. The company has said the model isn’t a replacement for a doctor.
Clinicians are skeptical.
Dr. Mary K. Mulcahey, chief of sports medicine at Loyola University Medical Center, who specializes in shoulder and knee injuries, said AI would likely have a hard time distinguishing between different types of meniscus tears. Her treatment recommendations differ based on the type of tear, how active a patient is or how much arthritis they have, she said. AI-driven answers culled from sifting information in large language models might lack important context or fail to interpret it correctly.
More of Mulcahey’s patients are feeding imaging results into AI, too.
“That does become a little bit dangerous if the patients come in already thinking they know what the treatment should be without taking the expertise of the physician or surgeon into account,” she said.
Mulcahey and other doctors said AI can be helpful in some cases, including to help break down complex medical jargon into patient-friendly terminology.
Research from 2024 examined how effective AI chatbots were at simplifying pathology reports. It found that ChatGPT dropped the average reading level to that of a seventh-grader and was medically correct nearly 98% of the time. However, it falsely reported that lymph nodes were cancer-free when no lymph nodes had been tested. The study authors noted that this mistake, which another chatbot also made, was the most common one they encountered.
When Megg Gawat, a 32-year-old software engineer, got the results of his blood work last year, he was confused by some of the medical jargon and acronyms. Rather than wait for a clinician to call, he asked ChatGPT for its interpretation. He was drawn to the freedom of asking the chatbot a question about his results any time of day, and without judgment or embarrassment.
The AI was complimentary at first: “You’re in great shape,” it wrote. But Gawat caught several errors in real time, including a misreading of his vitamin D level, which was low. When Gawat corrected the AI, its tone shifted to urgency, telling him to start taking a vitamin D supplement “immediately,” with a red siren emoji.
When a nurse called to relay his results, she also suggested that he add a vitamin D supplement, though with much less urgency, Gawat recalled.
Another patient, a 65-year-old veteran whose experience was chronicled in a case study published in the American Journal of Geriatric Psychiatry this spring, fed ChatGPT the results of two previous memory tests. The bot told the patient, whose doctor had previously diagnosed him with mild cognitive impairment, that his condition had progressed to dementia. In reality, according to the case study, the patient hadn’t declined at all.
A spokesperson for OpenAI said single case studies aren’t good representations of its broader user experience, and that its newer models have better capabilities and safeguards, including extra levels of encryption on users’ integrated health data.
Patients should also understand that when they upload health data into a chatbot, they’re waiving the federal protections granted under HIPAA, the Health Insurance Portability and Accountability Act, said Brad Malin, a professor of biomedical informatics at Vanderbilt University who studies health-data privacy. Companies largely set their own privacy policies and can change those terms, he added.
“In this situation, it is not a patient, it’s a consumer,” said Malin. “For the most part, it’s buyer beware.”
Write to Alex Janin at alex.janin@wsj.com
