I will never forget the look on his face. He wasn’t frightened. He wasn’t uncertain. He came into the hospital the way people used to come into courtrooms: with evidence already assembled, a verdict already reached, and very little interest in hearing otherwise. He had consulted an AI program before arriving, which had given him a diagnosis. He had read it, believed it, and arrived not as a patient seeking answers but as a patient defending a conclusion. That look, calm and certain, is the look I have come to fear more than almost anything else in medicine. Not because he was wrong; in fact, he may have been right. He had stopped wondering.

A crisis that predates the algorithm

Diagnostic error is not new and is one of the most persistent and underacknowledged threats in American health care. At least 150,000 patients experience preventable harm from diagnostic errors annually. Diagnostic errors account for more than 8 percent of adverse events in medicine and up to 30 percent of malpractice claims, with cognitive factors, such as premature closure, involved in approximately 75 percent of cases.

Premature closure is the moment the mind stops searching because it believes the search is over. If the history fits, the exam fits, and the working diagnosis feels right, then the physician (or the patient, or the algorithm), stops asking questions. As we have observed, when the diagnosis is made, the thinking stops.

The upstream migration of certainty

What has changed is where premature closure now begins. It no longer originates in the clinical encounter but arrives before the patient even does. A recent survey found that 50 percent of physicians report that their patients mention using AI tools, such as ChatGPT, symptom checkers, and LLM chatbots, at least occasionally before a visit. Among adults under 30 years of age, approximately 25 percent consult AI chatbots monthly for health information and advice.

This is not a story about health literacy. Health literacy, or the capacity to find, understand, and use health information, is a genuine public good. Informed engaged patients ask better questions and can report symptoms more precisely. They participate more meaningfully in shared decision making. However, AI-generated certainty is not health literacy; it is something categorically different. Health literacy opens the diagnostic conversation, but artificial certainty closes it. The patient who arrives having researched their symptoms is a partner in the clinical encounter, but the patient who arrives having received a confident algorithmic conclusion is defending a position. Those are not the same patient, and they require fundamentally different clinical responses, ones that our care delivery systems and our quality metrics are not yet equipped to handle.

What AI does well, and what it cannot do

AI is definitely changing medicine. AI tools are demonstrating value in summarizing clinical literature, flagging patterns in imaging, identifying high-risk patients, and reducing the administrative burden that drives physician burnout. These are very worthy contributions.

However, AI does one thing that the practice of medicine deliberately resists: It produces confident answers immediately. It does not wait, and it does not say, “Let me see what happens tomorrow.” It does not sit with a patient across two visits and notice that something has changed. In medicine, time is one of the most valuable diagnostic tools we have. Diseases evolve, and symptoms declare themselves across days and not seconds. Imaging catches up to pathology on its own schedule. Tests can become positive later.

AI operates in snapshots while medicine is a timeline. This mismatch, confidence delivered instantly into a field that requires deliberate uncertainty, is the structural problem. It is worth naming precisely: This is not a failure of technology, it is a failure of context. AI tools are not designed for medicine’s methodology but instead designed for completion, for answers, and for resolution. Medicine is a discipline that must sometimes resist all three.

The policy discussions we are not yet having

We are having the necessary conversations about AI in medicine: clinician use, liability, bias, and validation standards. These conversations matter, but we are not yet having serious policy conversations about what happens when AI-generated diagnoses migrate upstream into the patient encounter, not through physician tools but through consumer platforms that millions of Americans are already using before they ever enter a clinical setting. These are questions of health systems design and patient safety.

When patients arrive with AI-anchored diagnoses, visit structures designed around 15-minute problem-focused encounters are poorly equipped to address them. The time required to reopen a closed differential, rebuild trust in the diagnostic process, and attend to the actual presenting complaint is not time our reimbursement models currently recognize. We are paying for speed, and artificial certainty speeds the encounter further in the wrong direction. Meanwhile, the downstream costs of diagnostic error, such as repeat visits, delayed treatment, and patient harm, remain among the most expensive and least addressed problems in American health care. Adding a new upstream source of anchoring bias without addressing the structural conditions that make premature closure so dangerous is not a technology problem but a policy failure in waiting.

The treatment is discipline

This is not an argument against AI in health. The clinicians and systems that dismiss AI fully will be left behind and will miss genuine opportunities to improve care. However, a tool that generates confident answers in a field built on disciplined uncertainty must be handled with unusual care. The treatment for artificial certainty is not skepticism of technology but creation and protection of something that clinicians spend careers building: the discipline of remaining uncertain long enough for the truth to appear. That discipline is not instinct but rather the practice of medicine. It is built across years of being wrong in instructive ways and lives in the physician who asks one more question after they think they know the answer. It lives in the patient encounter that begins with genuine curiosity rather than algorithmic confirmation as well as health systems that create space for deliberation before the diagnosis is made.

This case I shared at the beginning did not teach me about AI but instead reminded me about the discipline that medicine has always required. The best clinicians are not the fastest to the diagnosis but are the ones willing to remain uncertain the longest. AI will give us answers faster, but medicine has never been about speed and always been about judgment.

The danger is not AI. The danger is artificial certainty.

Ganesh Asaithambi is a vascular neurologist.