AI Healthcare Tools: Patient Access and Medical Navigation

The Doctor Will Not See You Now

The average American primary care visit lasts between 15 and 20 minutes. Getting one takes about three weeks. The AI is available now, answers in full sentences, and never makes you feel like you’re wasting its time. When you frame it that way, the surprise isn’t that people are using chatbots for medical advice. The surprise is that anyone finds this surprising.

---

The conversation about AI in healthcare keeps organizing itself around the wrong question. Is it accurate enough? Is it safe? Does it cross the line into practicing medicine without a license? These are reasonable concerns, and they deserve serious answers. But they all assume that the interesting thing about AI health tools is what they do. The interesting thing is who benefits from them — and who doesn’t.

Start with why people actually use them. The dominant narrative — that patients are replacing their doctors with chatbots — is not supported by what we know about user behavior. Rock Health’s 2025 Consumer Adoption of Digital Health Survey, which polled 8,000 U.S. adults in December 2025, found that one in three Americans had used an AI chatbot for health information, double the share from the year before.

When researchers looked at what those users did next, 40 percent consulted a healthcare provider. Another 42 percent sought additional sources to verify the AI’s response. Most people are not using AI instead of medicine.

They are using it to navigate medicine — to understand what their lab results mean before the doctor calls back, to walk into an appointment with better questions, to make sense of a clinical note written in language that was never intended for them.

That instinct has a name: agency. Not the desire to diagnose yourself, but the desire to participate in your own care as something other than a passive recipient of decisions made in a register you can’t fully access.

The chatbot, for many users, is first and foremost a translation tool. What it translates is the gap between what medicine knows and what patients are allowed to understand.

That gap is larger than most people realize.

---

The Number That Reframes Everything

Only 12 percent of American adults have proficient health literacy, according to the U.S. Department of Health and Human Services. Twelve percent.

That means 88 percent of the adult population lacks the skills to fully find, understand, and act on health information — including the documents their own doctors send them. Clinical notes are written for clinicians. Lab reports arrive as columns of values and abbreviations with no interpretive context.

Radiology findings are described in a technical register that is, for the overwhelming majority of patients, a foreign language.

This has been true for decades. What changed is that patients now receive these documents directly, through electronic health portals, and are expected to engage with them — often hours or days before speaking to a physician. When Judith Miller, a 76-year-old Milwaukee resident, received flagged lab results and couldn’t get a detailed explanation from her care team, she put the report into a chatbot and asked what the values meant.

NPR reported her case in September 2025 as part of a broader pattern: patients turning to AI not to get a diagnosis but to understand a document their own healthcare system handed them without translation.

Miller’s case sounds like a success story. In some ways it is. But it also assumes something that doesn’t apply to most of the 88 percent: that she knew how to ask the question.

---

The Tool That Rewards the Already Informed

Here is the paradox at the center of the AI health democratization narrative. The chatbot is available to everyone. But it is most useful to the people who are already best equipped to use it — and most dangerous to those who aren’t.

A 2025 paper in the Annals of the New York Academy of Sciences examined what the researchers called GenAI-mediated confirmation bias in health information seeking. The finding was precise: AI is particularly vulnerable to reinforcing what users already believe, because users tend to phrase questions in ways that narrow the field of possible responses.

Someone who has already decided they have a particular condition will ask the AI to evaluate that hypothesis — not to challenge it. The AI, calibrated to be responsive and contextually consistent, will tend to comply. Information that contradicts the user’s existing belief, the research found, is frequently misinterpreted or dismissed. The chatbot becomes a mirror, not a window.

This dynamic is not new. It plays out on Google too, where confirmation bias is just as available and just as tempting. But AI changes the texture of the problem. A Google search returns a list of links that implicitly signals uncertainty — the user sees competing sources, divergent claims, the messiness of the information landscape.

An AI chatbot returns a single, fluent, personalized response in the confident register of someone who has considered the question carefully. There are no competing links. The scaffolding of doubt is gone. Users, the research notes, often interpret that tone as expert validation.

The person most likely to catch this is the person who already knows enough to be skeptical — who can recognize when an answer is too tidy, who thinks to ask a follow-up that complicates rather than confirms.

That is precisely the person who needs the translation service least. The 76-year-old who doesn’t know that framing her question differently would produce a different answer — and might produce a more accurate one — is the person the democratization promise is supposed to reach. She is also the person most exposed to its failure mode.

---

What Better Looks Like, and What It Costs

To be clear about what AI actually does well: it is a genuine improvement over Google as an information environment, for specific reasons. It synthesizes rather than ranks. It contextualizes values instead of listing them.

It can tell a patient that flagged lab results may be clinically insignificant without that patient having to evaluate the credibility of ten separate sources. A proof-of-concept study by researchers at OpenNotes — an academic lab at Beth Israel Deaconess that advocates for healthcare transparency — found that AI models performed well at responding to patients’ questions about clinical notes, with one consistent finding: the quality of the response was a direct function of how the question was framed.

A well-constructed prompt produces a useful answer. A poorly constructed one produces a confident-sounding answer that may point in the wrong direction.

That asymmetry is the real variable. And it is not distributed equally.

Cognitive reflection — the ability to pause and interrogate one’s own assumptions — is the trait most associated with resistance to confirmation bias, according to the same research. It correlates with education, with familiarity with how information works, with the kind of background that lets you recognize when a source, human or AI, is telling you what you want to hear.

It is, in other words, unevenly distributed in ways that track existing social and economic inequality. The people most likely to use AI health tools critically are disproportionately the people who already have better access to healthcare, better health outcomes, and better ability to advocate for themselves in a clinical setting.

This is not a reason to restrict AI health tools or treat the technology as inherently problematic. It is a reason to be honest about what “democratizing access to medical information” actually means when the tool’s value is gated by skills that most people don’t have and no one is systematically building.

---

The Question Nobody Is Asking

ChatGPT Health launched in January 2026. Microsoft Copilot Health followed in March, drawing on records from more than 50,000 U.S. hospitals and provider organizations. Amazon expanded its health assistant the same week.

The regulatory conversation is catching up slowly — several state medical boards are examining whether these products constitute the unauthorized practice of medicine, and the FDA has yet to establish a clear framework for consumer-facing AI health tools.

What is largely absent from that conversation is any serious attention to the education layer. Not health literacy in the traditional sense — the ability to read and understand medical documents — but something more specific: the ability to use AI as a health tool without having the tool use you.

How to frame a question that invites challenge rather than confirmation. How to recognize when a fluent answer is not necessarily a correct one. How to treat the chatbot as a starting point rather than a conclusion.

This is not technically complicated. It does not require a new regulatory framework or a significant investment. It requires someone to decide that it matters — that the difference between an AI health tool that expands meaningful access and one that gives the illusion of access while quietly tracking existing inequalities is worth addressing before the user base doubles again.

In the meantime, the tools will keep improving. The models will get better at flagging uncertainty, at pushing back on leading questions, at distinguishing between what a user wants to hear and what they should know. That trajectory is real.

But it runs on the assumption that the companies building these products are optimizing for the right thing — for accuracy and epistemic humility rather than for engagement and the feeling of a resolved question.

Those are not always the same objective.

---