How AI Health Advice Is Changing Dental Patients, Treatment Decisions, and Case Acceptance
A 60 year old man arrived at an emergency department convinced that his neighbor was poisoning him. Within hours, paranoia deepened into auditory and visual hallucinations. Laboratory testing was bizarre, and the diagnosis eventually came into focus. He had bromide poisoning.
For about three months, the man had replaced ordinary table salt with sodium bromide purchased online. He told physicians that he had consulted ChatGPT while investigating alternatives to chloride. His blood bromide concentration reached roughly 1,700 mg/L, hundreds of times the reference range. After bromide was stopped and he received supportive care, his laboratory abnormalities and psychiatric symptoms resolved.
The most important detail is what investigators could not determine. They never saw the original ChatGPT conversation. They could not establish exactly what the patient asked, what the system answered, whether warnings appeared, or how he interpreted them. The physicians later tested ChatGPT 3.5 with a similar question and received an answer that included bromide, but that did not reconstruct the original exchange. The careful conclusion is not that ChatGPT told a man to eat sodium bromide. It is that a man developed severe bromism after reportedly using ChatGPT while investigating chloride substitutes, and the information pathway that may have influenced him could not be audited. That distinction matters because dentistry is entering the same world.
A 2026 KFF poll found that 32 percent of U.S. adults had used AI for physical or mental health information or advice during the previous year. Twenty seven percent had used it to research symptoms or medical conditions, 19 percent to interpret tests or diagnoses, another 19 percent to compare treatment options, and 16 percent to help decide whether professional care was necessary. Among people who used AI for physical health information, 42 percent did not later follow up with a healthcare professional.
The reason is simple. AI is immediate, private, inexpensive, patient, and available at two in the morning. Patients may ask a chatbot whether a root canal is safe, whether fluoride is toxic, whether antibiotics are necessary, whether periodontal disease can be reversed naturally, whether an implant is better than a bridge, or whether a recommended crown is really needed. By the time they sit in your chair, they may already have a diagnosis, a treatment preference, a price expectation, and a story about risk.
AI is more persuasive than a conventional web search because it does not simply return links. It synthesizes an answer, explains it conversationally, and adjusts to follow up questions. Sometimes that helps. AI can translate technical language, prepare patients for appointments, and reinforce explanations afterward. In one 2026 randomized study involving 2,400 adults without medical training, a GPT based respiratory chatbot improved simulated diagnostic and triage performance compared with conventional web search. The lesson is not that AI is better than doctors. The relevant comparison is often AI versus whatever the patient would otherwise use.
The danger is that AI can also be wrong in a way that feels right. A 2026 physician led evaluation of 888 chatbot responses to patient style medical questions found that, depending on the model, roughly 22 percent to 43 percent were judged problematic and about 5 percent to 13 percent unsafe. Those numbers are not patient injury rates. The questions were simulated, the outputs were physician rated, and the models have since changed. But the study confirms the practical problem. A fluent answer can be incomplete, outdated, or unsafe while sounding authoritative.
Dentistry is especially vulnerable because so many decisions depend on details a generic prompt cannot capture. A chatbot does not palpate swelling, probe a periodontal pocket, test pulpal vitality, inspect a crack under magnification, evaluate occlusion, judge restorative ferrule, or interpret a CBCT in full clinical context. It may know the literature, but it does not automatically know what matters about the patient in front of you.
That creates a new communication task. Instead of asking only what questions the patient has, consider asking what they have already read, searched, or asked AI about the problem. That reveals the patient’s starting point. It can expose a misconception before it quietly undermines case acceptance and show you which part of the treatment plan needs more explanation.
If a patient says ChatGPT told me I do not need this crown, ridiculing the chatbot may only deepen resistance. A better response is to inspect the reasoning. Ask what the patient was told, identify what was correct, explain what information the AI did not have, and connect your recommendation to findings the patient can see in their own radiographs, photographs, scans, and clinical exam. The strongest rebuttal to generic AI advice is not authority. It is patient specific evidence.
That is also good case acceptance. Patients often reject treatment because of uncertainty, cost, fear, distrust, or recommendations they do not understand. AI may amplify those forces, but it can also reveal them. A screenshot of a chatbot conversation may become a map of the patient’s real objections, whether the issue is fear of overtreatment, concern about toxicity, confusion about prognosis, or the assumption that a cheaper alternative is equivalent.
Dentists should apply the same discipline when using AI themselves. It can be excellent for drafting patient education, simplifying consent language, preparing postoperative instructions, summarizing literature, or brainstorming questions to investigate. But anything that could materially change diagnosis, treatment, medication, or urgency deserves verification. The more consequential the decision, the less acceptable it is to trust an answer simply because it sounds polished.
Privacy matters too. KFF found that 77 percent of Americans were concerned about the privacy of personal medical information supplied to AI, yet 41 percent of people who had used AI for health purposes said they had uploaded personal information such as test results or clinicians’ notes. Dental patients will do the same with radiographs, treatment plans, insurance estimates, and clinical notes. Dentists should also be cautious about placing identifiable patient information into general purpose AI systems unless the workflow meets applicable privacy requirements.
The broader concern raised in Nature Health is surveillance. Healthcare has systems for detecting harm through records, adverse event reporting, malpractice claims, regulators, and research. AI health conversations often occur outside those systems. If bad advice contributes to a harmful decision, the clinician may never know what was said, the technology company may never learn what happened, and researchers may never count the event.
For dentists, the practical response is neither panic nor blind enthusiasm. Assume AI is already part of the patient’s information ecosystem. Ask what the patient has seen. Correct errors without condescension. Show the clinical evidence a generic chatbot cannot see. Use AI where it improves communication and efficiency, but verify anything that could change care. Compared with what?
Join the Conversation!