The other day, I was deep in conversation with my co-founder, Aaron Patzer, reflecting on how we each approach health concerns in our personal lives. Despite my background in emergency medicine and his in tech, we had surprisingly similar medical information seeking habits: We both turn to large language models (LLMs), tap into our networks, and—when possible—avoid visiting a doctor.

What struck me wasn’t just our shared behavior, but how universal it likely is. Even people surrounded by health care professionals prefer not to become “patients” unless absolutely necessary. That avoidance instinct runs deeper than cost or logistics. There’s an emotional discomfort—anxiety, fear, vulnerability—that makes many hesitate to seek medical care.

This subconscious aversion is a powerful and often overlooked driver of how the service of health care and health information is consumed. My working hypothesis is that most people will do whatever they can to avoid seeing a doctor, even when they know they probably should. And now, with the emergence of powerful LLMs, avoiding the doctor no longer means avoiding answers.

We’ve entered a new era of health information access. In just the past two years, AI has moved from novelty to infrastructure. Type a medical question into Google today—say, “What does a heart attack feel like?”—and you won’t just get weblinks. You’ll get a composed, AI-generated narrative explanation, complete with differential diagnoses, symptom profiles, and action steps. These responses are conversational, nuanced, and more accessible than any textbook or journal article.

While there’s typically a size 8 font disclaimer to “consult a professional,” the AI-generated content feels like a consultation. It mimics the reassurance—or sometimes the alarm—of a real conversation with a physician. And for millions of people, it’s becoming their first, and sometimes only, point of care.

This changes everything.

Traditionally, doctors have been the gatekeepers of medical information. Even the early internet—think WebMD or Mayo Clinic—relied on physician-written or physician-reviewed content. But the sheer volume and velocity of LLM-generated knowledge now exceed our ability to supervise it. We are no longer the gatekeepers. The algorithm is.

That may sound alarming, but in some ways, it’s a natural evolution of shared decision-making. In clinical settings, we’ve long encouraged patients to weigh options, understand risks, and express preferences. This framework assumes a clinician is nearby to guide the process. But now, AI is playing that guide role at scale—and often without supervision.

The issue isn’t that AI is replacing doctors. It’s that it’s becoming a patient’s first interaction with the health care system. And that interaction isn’t neutral—it shapes expectations, influences decisions, and sometimes delays necessary care.

We saw this vividly during the COVID-19 pandemic. Patients with strokes, heart attacks, and traumatic injuries delayed coming to the ER. Was it fear of infection? Of dying alone? Of overwhelming the system? Likely all of the above. But the behavior was consistent with the core instinct we’re talking about: avoidance. And in an AI-powered world, that avoidance can feel less risky because there’s still a sense of “doing something.”

This means we, as physicians, need to shift our mindset. Our patients have spoken and they want to use this technology to understand their health. We don’t get to decide the if or the how: That ship has sailed. So the question becomes: How do we ensure that the new first step that patients take when they have a clinical question is as safe, accurate, and useful as possible?

We need tools to evaluate AI-generated medical content, not just for factual accuracy but for clinical risk. A prompt like “How do I re-wrap my sprained wrist?” is relatively low stakes. But “Is my chest pain a heart attack?” carries high risk, and even subtle misinformation could lead to catastrophic delays.

Here’s what we need:

• Accessibility standards: Health information should be education-level and linguistically personalized, automatically.
• Reliability metrics: Ways to measure prompt output quality and hallucination risk, especially by clinical domain, should be available to the user and doctors.
• Prompt risk stratification: Frameworks to evaluate how dangerous or urgent a prompt is, so higher-risk queries can be flagged or redirected to less variable sources.
• Content guardrails: Improved oversight—not necessarily by human review, which doesn’t scale—but through training, testing, and alert systems built into the models themselves.

Importantly, we should also recognize that this shift can be a good thing. People want to understand their bodies. They want to feel empowered. LLMs are helping meet that demand in ways that “Dr. Google” never could.

But democratizing access to information carries risk. That’s where the medical community still has a vital role to play—not by policing every piece of content, but by shaping the frameworks, setting the standards, and helping patients understand when AI is enough and when it isn’t.

Patients are changing. The tools they use are changing. And if the algorithm now comes before the doctor, we’d better make sure the algorithm is worthy.

Justin Schrager is an emergency physician.