Americans are embracing artificial intelligence for medical advice faster than the institutions meant to oversee their care. Roughly one-third of U.S. adults have used AI for health information, and one in four did so in the past month, often through chatbots like ChatGPT. The trend is most pronounced among younger people and those who lack reliable access to a doctor, including the uninsured and patients who cannot afford a visit.
That pattern reflects unmet demand for healthcare, but it also carries risk. When a chatbot is wrong, the consequences tend to fall on the people least able to absorb them. Doctors are moving in the same direction, largely on their own. OpenEvidence, a free clinical reference that physicians can use without institutional support, grew from about 3 million consultations a month to roughly 18 million a month by December 2025. Individual experimentation is normal, but system-wide adoption will be necessary to produce significant benefits for society.
The gap between enthusiasm and results is not unique to medicine. Across industries, 88% of AI pilots never make it to production, and a recent working paper from the National Bureau of Economic Research found that only about one in 10 companies report any productivity gains from AI over the last three years. One of the main impediments, according to analysts who have worked on more than 230 AI projects across 30 industries, is a flawed mental model among leaders. They are treating AI like traditional software, when it is not.
Traditional software is written, deterministic, and built linearly. AI is trained, sometimes even grown, and is largely probabilistic and dependent on its data. It is built in loops, and most of the work happens before anything resembling a product exists. Managing AI like software means specifying the answer before checking whether the data can actually produce such an answer. The consequences of that mismatch can be severe. A widely used sepsis model, for example, was found by an outside team to miss two of every three cases.
The complexity of building reliable AI systems is illustrated by a project that trained a model on mammograms from a range of hospitals without sharing patient data. Partway through training, engineers noticed the model could tell which hospital an image came from, even though nothing in the data said so. It had picked up tiny differences left by the scanners and was using them as a shortcut. A model that leans on such shortcuts fails at the first hospital it has never seen. The team caught the problem and engineered the shortcut out. The resulting model became the first FDA-authorized AI to predict a woman’s five-year risk of breast cancer from a mammogram alone.
That example points to a broader lesson. Following directions assumes someone has been there before and that step-by-step instructions will lead to the destination. But nobody knows where AI is going or how to get there without course-correcting along the way. The better approach is wayfinding, which means moving through deep uncertainty when neither the destination nor the route is known. It is both a leadership philosophy and a technical methodology for developing AI systems. In healthcare, where small changes in practice can have disproportionately large effects, adopting that mindset may be as consequential as handwashing was in 1847, when Ignaz Semmelweis asked doctors in a Vienna maternity clinic to wash their hands in a chlorine solution before examining patients. Deaths on the ward fell from about 18% to about 2% within two months. The evidence was clear, but his colleagues rejected it for decades because it did not fit how they understood disease. The lesson for AI is that the evidence of what works may already be available, but institutions must be willing to change how they think before they can act on it.