Doctors cite context, privacy and false reassurance risks

More Americans are reviewing medical information online before talking with a doctor. A legal provision that took effect in 2021 expanded near-real-time access to lab reports, imaging results and clinical notes, and some patients now upload those records to AI chatbots.

Dr. Jason Goldman, an internist in Coral Springs, Florida, said doctors are increasingly responding to conclusions patients have reached through online searches and AI tools.

“We’re trying to refute what the patients have convinced themselves that they have based on whatever they’ve searched and reviewed,” Goldman said. “You can get everything from, ‘You have a cold,’ to ‘You’re dying of cancer.’”

Goldman said some patients have used chatbot conversations to justify stopping lifesaving cholesterol medication. Others, he said, have questioned recommended vaccines after AI tools validated their hesitancy.

“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.

A 2026 poll from health policy research nonprofit KFF found that roughly a third of adults use AI chatbots for health advice and 19% use them specifically to interpret lab results or medical tests. OpenAI said more than 300 million people ask health-related questions on ChatGPT each week.

Technology companies are expanding their health offerings. Anthropic’s Claude can connect to some users’ medical records, while Epic Systems has an AI assistant built into its electronic medical-record software. OpenAI rolled out widespread access in July to a health feature that can link users’ medical records and wearable-device data and identify trends in their results over time.

Ashley Alexander, OpenAI’s vice president of health products, said the model can “reason better than clinicians.” OpenAI has also said the model is not a replacement for a doctor.

Doctors say the same result can require different treatment depending on information a chatbot may not have or may misread. Dr. Mary K. Mulcahey, chief of sports medicine at Loyola University Medical Center, said AI could struggle to distinguish among types of meniscus tears. Treatment recommendations can differ based on the tear, a patient’s activity level and the amount of arthritis, she said.

More of Mulcahey’s patients are also submitting imaging results to AI tools.

“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,” Mulcahey said.

Mulcahey and other doctors said AI can still help explain complex medical terminology in language patients can understand.

A 2024 study examined how well AI chatbots simplified pathology reports. The study found that ChatGPT lowered the average reading level of the reports to that of a seventh-grader and was medically correct nearly 98% of the time. But ChatGPT falsely reported that lymph nodes were cancer-free when no lymph nodes had been tested. The study authors said that error, which another chatbot also made, was the most common mistake they encountered.

Megg Gawat, a 32-year-old software engineer, used ChatGPT after receiving blood-work results last year. He said he wanted an explanation of medical jargon and acronyms without waiting for a clinician to call and valued being able to ask questions at any time without embarrassment.

The chatbot initially told him, “You’re in great shape,” Gawat said. He caught several errors, including a misreading of his low vitamin D level. After Gawat corrected the chatbot, it urged him to take a vitamin D supplement “immediately,” with a red siren emoji. A nurse later recommended a supplement as well, Gawat recalled, but with much less urgency.

A separate case study published in the American Journal of Geriatric Psychiatry this spring described a 65-year-old veteran who entered results from two previous memory tests into ChatGPT. The chatbot told the patient that mild cognitive impairment diagnosed by his doctor had progressed to dementia. The case study said the patient had not declined.

An OpenAI spokesperson said single case studies are not good representations of the company’s broader user experience. The spokesperson said newer models have better capabilities and safeguards, including additional encryption for users’ integrated health data.

Privacy concerns also extend beyond accuracy. Brad Malin, a Vanderbilt University professor of biomedical informatics who studies health-data privacy, said patients who upload health data to a chatbot waive the federal protections granted under the Health Insurance Portability and Accountability Act. Companies largely set their own privacy policies and can change those terms, he said.

“In this situation, it is not a patient, it’s a consumer,” Malin said. “For the most part, it’s buyer beware.”