A reliable AI is not a confident AI

A clinician asks an artificial intelligence a clinical question. The answer arrives in seconds, well built, assured. It may well be right. Nothing in the way it is written lets you know.

That is the real problem, and it is not the one usually named. Models are blamed for being wrong. They are wrong sometimes, like any tool. What makes them risky in medicine is that they are wrong in exactly the same tone they use when they are right. Dr Nabil Bouali, cardiologist at Poitiers University Hospital and president of the College of Cardiologists in Training, puts it plainly: the tool can state something false without flinching, and you can be caught out by it.

A confident answer and a reliable answer look alike. That is precisely the difficulty.

Why this weighs more in medicine than elsewhere

In most settings an error is corrected on the next attempt. In a consultation it is not. And responsibility cannot be delegated: whatever happens, it is the clinician who signs the prescription and answers for the decision.

Dr Bouali draws a conclusion from this that settles the adoption debate on its own: you cannot feel responsible for something you do not understand. A tool that does not show its reasoning therefore cannot sit inside a chain of medical responsibility. Not because it is poor, but because its output cannot be owned.

This is why explicability is not a convenience. It is the condition of entry.

What goes in

The reliability of an answer begins before the question, in what you agree to read.

JuisciGPT draws on two things: an indexed corpus of selected scientific literature, reindexed continuously, and a scientific web search that runs on every query. Every source is labelled by provenance, so you can see what a statement rests on instead of assuming it. Nothing is frozen at a training cut-off.

Every journal is checked against the list of recommendable journals published by the French Conference of Deans of Medicine and CNU Santé, and against other independent registers. A journal that none of them vouches for is heavily down-weighted. We do not grade journals on our own judgement. We rely on lists drawn up by those whose job it is.

Knowing where a statement comes from is a modest guarantee, and one of the few that can genuinely be given.

What we show

An answer without its path is worth nothing. Every statement is tied to its source, and every source is presented with what you need to judge it: the rank of the journal, its standing in citations, how closely it matches the question, whether guidelines have taken it up.

The confidence score shown is calculated from the weight of those sources. It is never generated by the model. The distinction is not cosmetic: a figure produced by the same model that wrote the answer is not a measurement, it is one more opinion.

And the engine does not stop at titles. On a search it tries to retrieve the full text of the publications it keeps. When the full text is not available, it works from the abstract or an excerpt.

What we do when we do not know

This is the part that makes the rest credible.

No source, no answer. If the search returns nothing, the agent declines instead of producing a plausible text. On a question that carries a care decision, when a publication type is explicitly requested, the engine applies that filter strictly and reports what is missing rather than widening the net in order to have something to say.

A general-purpose model will always answer you. Refusing to answer is a feature, not a failure.

What we do not do

The tool is not meant for patient data: users are asked to leave out anything that could identify a patient, and it works only from published information. Hosting is in the European Union, and conversations are deleted with the account, except an exchange selected for expert evaluation, which we keep to evaluate the engine.

A notice stays visible at all times, in the eight languages of the application: AI can make mistakes, check the source.

We do not present ourselves as a better judge of science. We do not produce this literature and we are in no position to grade it. We show the path to the source, and critical reading stays with the clinician.

The real question is not the one being asked

Public debate asks whether artificial intelligence should be used in healthcare. The facts have already settled it. The INSIGHT-AI France survey, run among 756 professionals by the College of Cardiologists in Training and the Cardio-AI Circle of the French Society of Cardiology, found that 23% use AI at least weekly and 40% occasionally, while 7.8% have had any formal training in it.

The divide is not between those who use it and those who do without. It is between those who can check what they are told and those who cannot.

A reliable AI is not one that is always right. It is one where you can tell when it is wrong.

---

Source: Bouali N, Domart S, Amara W, et al., on behalf of the INSIGHT-AI-France investigators. Artificial intelligence adoption in French cardiovascular care: a multiprofessional survey of barriers and facilitators. European Heart Journal — Digital Health. 2026;7(4):ztag042. doi:10.1093/ehjdh/ztag042