Some docs see LLMs as a boon for medical literacy. The common affected person would possibly battle to navigate the huge panorama of on-line medical info—and, specifically, to differentiate high-quality sources from polished however factually doubtful web sites—however LLMs can do this job for them, no less than in concept. Treating sufferers who had searched for his or her signs on Google required “quite a lot of attacking affected person anxiousness [and] lowering misinformation,” says Marc Succi, an affiliate professor at Harvard Medical Faculty and a working towards radiologist. However now, he says, “you see sufferers with a school training, a highschool training, asking questions on the degree of one thing an early med pupil would possibly ask.”
The discharge of ChatGPT Well being, and Anthropic’s subsequent announcement of latest well being integrations for Claude, point out that the AI giants are more and more prepared to acknowledge and encourage health-related makes use of of their fashions. Such makes use of definitely include dangers, given LLMs’ well-documented tendencies to agree with customers and make up info moderately than admit ignorance.
However these dangers additionally should be weighed in opposition to potential advantages. There’s an analogy right here to autonomous automobiles: When policymakers take into account whether or not to permit Waymo of their metropolis, the important thing metric shouldn’t be whether or not its vehicles are ever concerned in accidents however whether or not they trigger much less hurt than the established order of counting on human drivers. If Dr. ChatGPT is an enchancment over Dr. Google—and early proof suggests it might be—it might doubtlessly reduce the big burden of medical misinformation and pointless well being anxiousness that the web has created.
Pinning down the effectiveness of a chatbot corresponding to ChatGPT or Claude for client well being, nevertheless, is difficult. “It’s exceedingly tough to judge an open-ended chatbot,” says Danielle Bitterman, the medical lead for information science and AI on the Mass Normal Brigham health-care system. Giant language fashions score well on medical licensing examinations, however these exams use multiple-choice questions that don’t mirror how folks use chatbots to lookup medical info.
Sirisha Rambhatla, an assistant professor of administration science and engineering on the College of Waterloo, tried to shut that hole by evaluating how GPT-4o responded to licensing examination questions when it didn’t have entry to an inventory of attainable solutions. Medical specialists who evaluated the responses scored solely about half of them as totally appropriate. However multiple-choice examination questions are designed to be difficult sufficient that the reply choices don’t give them totally away, they usually’re nonetheless a reasonably distant approximation for the kind of factor {that a} person would kind into ChatGPT.
A different study, which examined GPT-4o on extra lifelike prompts submitted by human volunteers, discovered that it answered medical questions accurately about 85% of the time. After I spoke with Amulya Yadav, an affiliate professor at Pennsylvania State College who runs the Accountable AI for Social Emancipation Lab and led the examine, he made it clear that he wasn’t personally a fan of patient-facing medical LLMs. However he freely admits that, technically talking, they appear as much as the duty—in any case, he says, human docs misdiagnose sufferers 10% to fifteen% of the time. “If I have a look at it dispassionately, it appears that evidently the world is gonna change, whether or not I prefer it or not,” he says.
For folks searching for medical info on-line, Yadav says, LLMs do appear to be a better option than Google. Succi, the radiologist, additionally concluded that LLMs could be a higher various to internet search when he compared GPT-4’s responses to questions on frequent persistent medical situations with the data offered in Google’s information panel, the data field that generally seems on the suitable facet of the search outcomes.
Since Yadav’s and Succi’s research appeared on-line, within the first half of 2025, OpenAI has launched a number of new variations of GPT, and it’s affordable to anticipate that GPT-5.2 would carry out even higher than its predecessors. However the research do have necessary limitations: They give attention to easy, factual questions, they usually study solely transient interactions between customers and chatbots or internet search instruments. A number of the weaknesses of LLMs—most notably their sycophancy and tendency to hallucinate—could be extra more likely to rear their heads in additional in depth conversations and with people who find themselves coping with extra advanced issues. Reeva Lederman, a professor on the College of Melbourne who research know-how and well being, notes that sufferers who don’t just like the prognosis or therapy suggestions that they obtain from a physician would possibly search out one other opinion from an LLM—and the LLM, if it’s sycophantic, would possibly encourage them to reject their physician’s recommendation.
Some research have discovered that LLMs will hallucinate and exhibit sycophancy in response to health-related prompts. For instance, one study confirmed that GPT-4 and GPT-4o will fortunately settle for and run with incorrect drug info included in a person’s query. In another, GPT-4o continuously concocted definitions for faux syndromes and lab exams talked about within the person’s immediate. Given the abundance of medically doubtful diagnoses and coverings floating across the web, these patterns of LLM habits might contribute to the unfold of medical misinformation, significantly if folks see LLMs as reliable.
OpenAI has reported that the GPT-5 collection of fashions is markedly much less sycophantic and susceptible to hallucination than their predecessors, so the outcomes of those research won’t apply to ChatGPT Well being. The corporate additionally evaluated the mannequin that powers ChatGPT Well being on its responses to health-specific questions, utilizing their publicly obtainable HeathBench benchmark. HealthBench rewards fashions that categorical uncertainty when applicable, suggest that customers search medical consideration when vital, and chorus from inflicting customers pointless stress by telling them their situation is extra severe that it really is. It’s affordable to imagine that the mannequin underlying ChatGPT Well being exhibited these behaviors in testing, although Bitterman notes that among the prompts in HealthBench have been generated by LLMs, not customers, which might restrict how effectively the benchmark interprets into the true world.
