Uhlík vs Křemík · Czech podcast about artificial intelligence

Neurologist Ondřej Volný: bring your AI diagnosis to the doctor, but use common sense

Episode 19 · 30 September 2026 · 49:16 · Guest: Ondřej Volný

More and more people talk their symptoms through with a language model first and only then go to the doctor. Neurologist Ondřej Volný of the University Hospital Ostrava explains in the Uhlík vs Křemík podcast where this kind of pre-diagnosis helps, where it does harm and why doctors should not brush off informed patients. The conversation was recorded at the Černá AI festival in Ostrava.

Dr. ChatGPT has replaced Dr. Google

Ondřej Volný chairs the Czech Society for Artificial Intelligence and Innovative Digital Technologies in Medicine and advises the health minister on AI. He confirmed to hosts David Salomon and Adam Hanka that patients consult language models more and more often. In his view, there is no way to forbid it. At the same time, he points out that these are not tools designed for differential diagnosis.

He also warns of the so-called symptom spiral, even though it does not affect everyone. More anxious people ask the model the same questions over and over, and the answers can paradoxically frighten them even more. The model works with probabilities, and a list of ten or twenty possible diagnoses may include something that sounds incurable, even though in reality it may not be.

To show how common this is, Volný cites a study according to which roughly every fourth or fifth question put to language models concerns health, including the health of family members. According to data from the big providers, most of these questions come at times when medical care is least available, that is after working hours and after midnight.

The diagnosis was right, the doctor took offence

Adam Hanka described his own experience. While running from the tram to a talk, he tore a muscle. A friend asked GPT about his symptoms and the answer pointed to a muscle rupture. He then went to the university hospital with this diagnosis and told the doctor he had it from artificial intelligence. "At that moment the doctor stopped talking to me." A nurse finished the treatment, but the medical report ended up with exactly the same diagnosis.

Volný likes the way Adam went about it. He did not upload any personal data to the chat, he only described what happened and where it hurt. In Volný's view, doctors should not condemn a patient who asked a language model in good faith. "I think that we doctors and health professionals, too, have to face the truth." Language models have many weaknesses, but they handle pre-diagnosis relatively accurately.

Together with colleagues and in cooperation with patient organisations, the Czech Association for Artificial Intelligence and the Ministry of Health, Volný wrote a handbook for patients, a ten-point guide on how to approach language models. He also advises patient organisations that people should anonymise their information and not upload their personal identification number, home address or complete medical records to the chat.

What AI in medicine can do today and tomorrow

Volný distinguishes the language models patients use from software certified as a medical device. Such software evaluates scans, ECGs, skin findings or the back of the eye, but it is always trained for one specific diagnosis. A tool that evaluates stroke on CT will not recognise meningitis, and a tool for multiple sclerosis will not assess epilepsy.

Asked whether AI will ever be a better doctor than a doctor, he said he does not think so. When it comes to diagnosis from a description of symptoms, however, an AI Dr. House is in a way already here. As an example, he mentioned Microsoft's diagnostic orchestra, in which five AI agents trained on the hardest case reports reason about a patient. It is not yet a certified medical device.

Looking ahead, Volný imagines tools built right into the information system where doctors write discharge reports: with one click, the doctor has AI rewrite a finding into language the patient understands. His dream is a clinic without a computer, where he talks with the patient for ten minutes and the record is created from a transcript of the conversation.

When not to ask the model at all

The hosts also discussed how a patient can tell when to see a doctor. Adam Hanka advises asking AI for a thorough explanation of its reasoning, so that people have a basis for their own decision, not just an answer to go or not to go. Volný would welcome a patient's cookbook on how to ask the right way.

For some symptoms, however, the discussion is over. "Of course, if someone has a problem such as chest tightness and shortness of breath, I would definitely not ask a language model whether to call the ambulance, I would call it." The same applies when the corner of someone's mouth droops, an arm weakens or they have trouble speaking.

The full conversation with Ondřej Volný is available in the Uhlík vs Křemík podcast.

The episode is in Czech. Czech page with chapters, quotes and the full transcript