When I first moved to the United States in 2014, I developed a skin rash during the winter. I had grown up in Brazil, where my skin had never really dealt with that kind of cold, and I had no idea what was happening.
Before going to urgent care, I Googled it. Not because I thought Google was more qualified than a physician, but because healthcare in the U.S. felt incredibly expensive to me at the time and I wanted to know whether this was something I actually needed a doctor for or something I could manage myself.
The appointment felt unfamiliar almost immediately. In my experience with healthcare in Brazil up to that point, the computer had rarely been a noticeable part of the interaction. Plenty of physicians I had seen were still using paper charts. The conversation happened between you and the doctor, and whatever documentation was happening stayed mostly in the background.
At this urgent care, the screen was suddenly very much in the foreground. I remember the doctor working through a series of questions on the computer, reading them to me and entering my answers as we went. At the end, he gave me a diagnosis, recommended Cetaphil, and sent me home.
I cannot tell you what was happening clinically on his side of that screen. Maybe he knew what the rash was as soon as he saw it. Maybe the software was documenting what he already knew, helping him rule out something more serious, or allowing him to get through a busy urgent-care workflow efficiently. I certainly was not qualified to know.
I can only tell you how the encounter looked from my side. I answered questions. He entered the answers. The computer seemed to guide the process. And when I left, I remember thinking, Couldn't I have basically done that at home?
Looking back, I do not think that reaction tells us very much about whether the doctor practiced medicine well that day. It tells us something about what expertise looks like to a patient.
More than a decade later, that memory came back to me while reading a recent study published in the Journal of Medical Internet Research. The technology is obviously very different now, but the researchers were examining a surprisingly similar question: when technology becomes part of the medical decision, what does the patient believe the physician is contributing?
Across two preregistered experiments, researchers showed participants different versions of AI-assisted medical consultations. In some scenarios, AI provided descriptive support, such as helping analyze medical information. In others, it went further and generated a preliminary diagnosis.
The results were more nuanced than the familiar "patients trust doctors more than AI" narrative. Participants were not broadly opposed to physicians using technology. In fact, descriptive AI support was received quite positively. What changed their reactions was the perceived role of the technology in the medical decision.
When diagnostic AI appeared to influence the physician while the physician was forming an assessment, participants reported less trust than they did in some of the other scenarios. But when the physician was described as reviewing the case first and then consulting the AI output, that trust disadvantage disappeared.
That does not mean physicians who use AI earlier in their workflow are not thinking for themselves. A vignette study cannot tell us that, and real clinical decision-making is far more complicated than the scenarios researchers can put in an experiment. It does suggest something important about what patients perceive.
Patients may be trying to understand not simply whether AI was involved, but what the physician contributed that the technology could not. That is a very different question.
Modern medicine asks an extraordinary amount of physicians. Clinical information is expanding, documentation requirements consume time, administrative work competes with patient care, and visits often have to accomplish far more than the schedule realistically allows. Expecting physicians to reject useful technology simply to prove that they are working hard enough would be absurd.
The promise of good clinical AI is precisely that physicians should not have to spend their limited attention on every task a machine can safely assist with. AI can organize information, recognize patterns, surface possibilities, reduce cognitive load, challenge an assumption, or provide another analytical layer. Used well, those tools may allow physicians to direct more of their attention toward the parts of medicine that require context, judgment and human understanding.
The trust problem begins when the patient cannot tell that any of that is happening. There is an enormous difference between the absence of clinical judgment and the absence of visible clinical judgment. Healthcare has reason to care about the second one even when the first is not a problem at all.
Patients do not see the thousands of hours of training behind a physician's interpretation. They do not see which possibilities were discarded, which finding made one diagnosis more likely than another, or whether the clinician recognized that an AI recommendation did not quite fit the person sitting in front of them. They only experience what becomes visible during the encounter.
That makes communication around AI more important than simply disclosing that a tool was used. Imagine hearing, "AI analyzed your scan and suggested X, and I agree." Now compare that with, "I've reviewed your scan, and these findings make me think X is the most likely explanation. We also use an AI tool as an additional check, and it identified the same pattern."
Both statements can describe responsible AI-supported care. Neither tells us whether one physician actually reasoned more carefully than the other. But they give the patient very different information about the physician's role.
That distinction matters because patients are also becoming more capable of using AI themselves. If I had walked into that urgent care in 2014 after having a conversation with today's generative AI rather than running a few Google searches, I might already have had possible diagnoses, explanations, suggested treatments and questions to ask.
If the clinical encounter then appeared to reproduce essentially the same process on a different screen, my question would have been even harder to dismiss: What am I getting here that I couldn't get at home?
The answer, of course, is potentially enormous. A physician has clinical training, examination findings, access to testing, knowledge of the patient's history, responsibility for the decision, and the ability to recognize when the obvious answer is wrong. AI does not erase any of that. But patients cannot value what they cannot perceive.
That may be one of the more overlooked challenges for healthcare organizations implementing AI. Much of the current conversation focuses appropriately on accuracy, safety, bias, workflow integration, disclosure and governance. Those are foundational issues. There is another design question worth adding: after this technology is integrated into the encounter, will the patient still understand what the physician is contributing?
That is not an argument for physicians to perform their reasoning theatrically or narrate every clinical decision. It is not another demand to pile onto clinicians who are already working inside constrained systems. Sometimes a sentence is enough: "This gives me another data point, but here's what I'm seeing in your case." Or, "The system flagged this possibility. I'm less concerned about it because..." Those moments make the clinician's role legible without turning the visit into a demonstration.
There is a lesson here for healthtech marketing too. "AI-powered diagnosis" sounds impressive when the audience is an investor, buyer or technology leader. But healthcare products operate inside relationships, and the same positioning can carry a different implication when it reaches a patient: the machine figured this out.
Perhaps the more valuable story is that the technology gives clinicians another set of eyes, another source of information or another way to check their reasoning, while the physician remains responsible for interpreting what it means for the person in front of them. That does not diminish the AI. It places its value in the right context.
I do not look back at my urgent-care doctor and think he failed to practice medicine. I genuinely do not know what was happening on his side of that computer. I was not in his head. I was on the other side of the encounter, trying to understand where his expertise fit into what I was watching.
What I know is what the encounter communicated to me. The software was visible. His expertise was not.
As AI becomes more capable and more deeply embedded in healthcare, that distinction is going to matter. The goal should not be to keep technology out of the exam room. Physicians need good tools, and patients benefit when those tools make care better. The goal is to make sure that when technology becomes more visible, the physician does not inadvertently become less so.
SOURCES & EVIDENCE
Primary evidence
Journal of Medical Internet Research: Effects of AI Assistance Type and Timing on Patient Trust in Medical Decision-Making (2026)
Evidence boundary
The preregistered vignette experiments support claims about how participants perceived different types and timing of AI assistance. They do not show how physicians actually reason in real clinical encounters or whether one workflow is clinically superior.
When a piece rests on my own data, I say so. When it rests on someone else’s, I say whose, and whether they funded it.