# People trust human nurses more than AI, but the intuitiveness of the advice matters most > Source: > Published: 2026-08-13 14:00:52+00:00 A recent study published in the * Journal of Medical Internet Research* suggests that people still find health advice from human nurses more credible than advice from artificial intelligence. However, the nature of the advice itself plays a substantial role in determining trust, often outweighing the source. The findings indicate that individuals apply similar critical thinking heuristics to both human and machine sources, particularly evaluating whether the recommendations match their own intuitive expectations. Generative artificial intelligence (AI) systems are becoming common tools for everyday information seeking. Millions of users ask platforms like ChatGPT for health and wellness guidance on a regular basis. Because these systems are neither licensed medical professionals nor anonymous web pages, they occupy a unique position in how medical knowledge reaches the public. The idea for the study came from a personal experience with one of these chatbots. Asheley R. Landrum, an associate professor and director of the News Co/Lab at Arizona State University, was snacking at home when she had a minor scare. “This is sort of embarrassing, but most people who know me well won’t be surprised,” Landrum said. “So, ASU has gone all in on AI, and we’ve been exploring it and figuring out how best to use it. One day, I was working from home and eating Doritos, and one kind of scraped my esophagus as it went down (or at least that’s what it felt like). I had ChatGPT open, so I asked,’ I just swallowed a Dorito and it scraped me on the way down. Should I be worried? What should I do?'” The program provided a detailed response, advising her to sip water, eat soft foods, and seek urgent medical attention if she experienced severe symptoms like difficulty breathing or swallowing. “I wondered how many people might use ChatGPT (and similar tools) for health advice,” Landrum said. “So, I ran a small national survey and presented the results at the American Association for the Advancement of Science (AAAS) in 2025. A month or so later, we got in contact with a company developing an AI Nurse and incorporated this type of tool in a pre-registered experiment, which is now published. We are still doing work in this area.” Epistemic trust, which is the willingness to accept information from a source as reliable and act upon it, relies heavily on how people perceive competence and benevolence. Licensed medical professionals usually serve as the benchmark for this type of credibility due to their extensive training. At the same time, patients are increasingly aware of human fallibility, cognitive biases, and structural issues within healthcare systems. To explore these dynamics, the researchers recruited 1,502 adults in the United States, utilizing a panel matched to national demographic census data. Participants engaged in a survey experiment where they read fictional conversations featuring a person seeking health advice. Each participant was randomly assigned to one of three advice sources: a licensed human registered nurse named Nurse Dobson, a specialized AI nurse trained on medical data, or the general-purpose chatbot ChatGPT. The assigned source remained the same for a given participant throughout all experimental scenarios. [Add PsyPost to your preferred sources](https://www.google.com/preferences/source?q=psypost.org) In the first set of analyses, participants read two scenarios categorized by risk level. The low-risk scenario involved dietary advice regarding eating eggs when trying to reduce cholesterol. The high-risk scenario involved whether to seek emergency care for chest discomfort. For these scenarios, the provided advice was randomly manipulated to be either intuitive or counterintuitive. Intuitive advice matched common public expectations, and counterintuitive advice departed from expectations but remained medically accurate. After reading the text, participants rated the credibility of the advice on a scale from 1 to 5, where 3 indicated moderate credibility. Advice from the human nurse received an average credibility rating of 3.53, making it slightly more trusted than advice from the AI nurse at 3.13 and ChatGPT at 3.21. Participants also reported a higher willingness to follow the human nurse’s recommendations. The difference in credibility ratings between the two AI sources was not statistically significant. The intuitiveness of the message exerted a strong influence on trust. Across all sources, intuitive advice received an average credibility rating of 3.46, compared to 3.12 for counterintuitive advice. This gap widened in the high-risk scenario, meaning that counterintuitive advice was penalized more heavily when the medical stakes were higher. “I think that there are two stories here, one for the average person and one for developers,” Landrum told PsyPost. “For the average person, one of the biggest takeaways is that we may not judge AI health advice as differently from human advice as we think we do. People in our study trusted the human nurse more, but they didn’t dismiss the AI sources as untrustworthy.” “In fact, what the advice actually said and whether it seemed intuitively right mattered as much as or more than whether it came from a human or AI,” Landrum continued. “This is a clear example of confirmation bias. So, one practical takeaway is that when health advice ‘just sounds right’ or ‘sounds wrong,’ we should ask ourselves (as non-medical experts) whether we are responding to the evidence or whether it simply matches what we already believe.” Confirmation bias is the psychological tendency to favor information that aligns with a person’s preexisting beliefs. The researchers then presented participants with a morally sensitive scenario about freezing eggs to preserve fertility for later in life. In this scenario, the ideological tone of the advice was manipulated between subjects. Participants received advice that was framed neutrally, conservatively framed to emphasize natural conception, or liberally framed to emphasize reproductive autonomy. Across all sources, participants rated the conservatively biased advice as less credible. They were also less willing to recommend taking the conservatively framed advice. The researchers anticipated that the AI sources might be perceived as less biased than the human nurse, but the data did not support this hypothesis. Adding a specialized “AI nurse” label also failed to elevate trust above the general-purpose chatbot. “For developers, there are takeaways for designing AI-health tools,” Landrum explained. “There’s a lot of emphasis right now on whether people trust AI, but our findings suggest that’s probably too simple a question. Trust depends on the person, the situation, and how the information is communicated. Simply labeling something as a specialized health AI tool didn’t make people trust it more than ChatGPT, for example.” “Where design may really matter is when good medical advice is surprising or goes against what people expect,” Landrum added. “In those cases, especially when the stakes are high, an AI system probably shouldn’t just give someone an answer, it may need to explain why the recommendation makes sense, where it comes from, and when someone should seek additional medical care. (Although, we can ask if the justification is persuasive or simply the presence of a justification is persuasive, see Langer’s copymachine study!).” This references a classic 1978 psychology experiment by Ellen Langer, which found that people were more likely to comply with a request to skip a line for a copy machine if the person provided any reason at all, even a meaningless one like “because I have to make copies.” Finally, the researchers evaluated overall perceptions of source competence and benevolence. They collected measures of the participants’ medical skepticism, political voting behavior, religiosity, and prior experience with generative AI. Older adults and individuals with prior AI experience tended to give higher competence and benevolence ratings across all three sources. A notable relationship emerged regarding medical skepticism, which is a mindset where individuals doubt the necessity of professional medical intervention and feel confident managing their own health. For participants scoring higher in medical skepticism, perceived competence of the human nurse tended to drop. At the same time, higher medical skepticism was associated with higher perceived competence of the specialized AI nurse. “We found a really cool interaction where people who are more skeptical of traditional medical authority saw the AI Nurse and the Human Nurse as equally competent, but people who are lower on such skepticism saw AI Nurses as much less competent than human nurses,” Landrum said. “This finding was exploratory and we are seeing if it holds in another study. If it does, it might mean that AI Nurse tools could be a way to reach those who are typically less likely to go to medical professionals for advice.” Certain limitations contextualize these findings. The scenarios used in the experiment were hypothetical vignettes, which do not perfectly mirror the emotional weight of real health decisions. “The biggest caveat is that we studied how people evaluate health advice, not whether they can tell when AI medical advice is actually right or wrong,” Landrum cautioned. “These were also hypothetical situations, so we don’t know whether someone facing a real health scare would respond the same way. As I said above, I’d also be a little cautious with our finding about people who are more skeptical of traditional medical professionals. It’s potentially really important, but that part of the study was very exploratory and needs to be replicated.” Additionally, participants only read text from the chatbots rather than generating their own prompts. Real-world trust in a specialized medical AI might depend heavily on its user interface, institutional branding, or active dialogue. “This work was done with set advice and we simply manipulated who it was purportedly from,” Landrum said. “That is, we said whether it was from ChatGPT, or an AI Nurse, or a human nurse. The advice itself did not vary. People did not actually interact with any AI tools. In our current and ongoing research, we are having participants actually interact with AI tools (in addition to survey questions) to better understand how they may use AI for health advice, when they may use AI for health advice, why, etc.” The study, “[Trusting Generative AI for Health Advice: Preregistered Survey Experiment](https://doi.org/10.2196/97882),” was authored by Asheley R. Landrum, Nitin Verma, and Amanda Kehrberg.