It aces diagnostic exams but can't sign one: a medical AI CEO's warning about health chatbots

By NewsNation · September 26, 2026.

An OpenAI model got 78% of cases right on a set of complex diagnoses and outperformed some emergency physicians, according to the study cited by NewsNation itself in an interview with Dr. Ome Ogbru, CEO of the medical AI company AI Engines. The figure sounds spectacular, but Ogbru —the very person who cites it— immediately qualifies it: such tests tend to hand the system information that in a real consultation would never be so neatly organized or available up front.

Ogbru, interviewed by Alicia Nieves, has a blunt message for the millions of users who already ask a chatbot about their symptoms: never skip the appointment with a healthcare professional and never treat these platforms as if they were giving medical advice. His practical recommendation is specific: check whether the AI's answer comes from health experts, medical societies or other reputable sources, and then take it to your doctor before acting on it.

It helps to know who is speaking. AI Engines does not sell a consumer chatbot: it builds tools for doctors, scientists and life sciences companies under what Ogbru calls an 'evidence first' approach, with integration of PubMed —the reference biomedical database— and traceable citations so professionals can audit where each answer comes from. It is a business model aimed at professionals that competes, in practice, with the unsupervised use of generic chatbots for health questions; his warning is consistent with his own behavior —he says he never uploads personal data to an AI whose destination he does not know— but it also maps out the exact niche his company wants to occupy: the intermediary between the patient, the chatbot and the licensed physician.

The underlying argument, however, is solid and does not depend on who makes it: an AI model is not board-certified, cannot sign a prescription and cannot assume legal responsibility for a wrong diagnosis. 'They can't own the decision or own the outcome,' Ogbru sums up. That is the line separating assisting from deciding, and it is also the blind spot in much of the media enthusiasm for benchmarks (standardized reference tests): acing an exam of archived clinical cases is a real achievement, but it is not the same as practicing medicine in front of a patient with an ambiguous history, fear and little time.

There is a second warning in the interview that deserves as much attention as the first: privacy. More and more people are dumping complete medical records into generic chatbots without knowing whether that data will feed the model's training or end up in some commercial use. Ogbru recommends checking, before typing anything sensitive, whether the platform offers the option to exclude those conversations from retraining; according to the source, some platforms already include that setting.

Our reading: this piece captures well the stage AI applied to health is at, and it fits what we have been observing in other regulated sectors. In the short term, the problem is not so much that AI diagnoses badly —it often matches or beats standardized tests— but that no one has yet settled who answers when it fails, or what happens to a person's most intimate data once it enters a system that was neither designed nor certified to handle it. That gap in liability and data governance is the real obstacle, not the model's technical capability.

In the long term we remain convinced that the trajectory points upward: an AI able to cross-reference scientific literature with traceable citations, like the one described here, is precisely the kind of tool that can speed up early diagnoses, make a second opinion cheaper to obtain and free up physicians' time for what no machine does well —the human relationship, the reading of a patient's life context. The road to a medicine that helps eradicate diseases and extends years of healthy life runs through traceable, auditable tools like those Ogbru describes. But that future only arrives if, in the meantime, today's patient is protected from the chatbot that answers with total confidence and no license.