A confident answer and a trustworthy answer are not the same thing. Most clinical AI optimizes for the first, and we built Meddies around the second.
The hidden mechanism is the liability
Ask a typical medical chatbot a question and it gives you a fluent, assured reply. It sounds right. It may even be right. But a doctor cannot tell the difference between a correct answer and a confident hallucination, because the model shows nothing but the conclusion. The reasoning, the source, the chain that produced the recommendation: all of it stays hidden.
In most software, a hidden mechanism is an inconvenience. In medicine, it is a liability. A clinician is accountable for the decision, not the tool. Asking a doctor to act on an unverifiable answer is asking them to put their name on a guess. No serious clinician will, and they are right not to.
The source travels with the recommendation
Watch how a careful physician uses a reference. They do not read the guideline cover to cover for every order. They check the specific line that bears on this patient: the dose ceiling, the interaction, the contraindication. The reference is not there to teach them medicine. It is there so they can verify the one thing in front of them, fast, and move on.
A clinical tool should support that exact motion. When Meddies recommends something, the source travels with the recommendation. Not a vague "according to guidelines," but the specific reference, openable in one click. For an interaction, that means Stockley and the BNF. For a national standard, the Ministry of Health protocol number. The doctor checks the line, confirms or overrides, and keeps moving.
Verification is the product
It is tempting to treat citations as a trust badge bolted onto the side of a model. They are not. The source on every answer is the difference, because it changes what the doctor is being asked to do. Without it, they are asked to trust. With it, they are asked to verify. Trust does not scale across a hospital. Verification does.
This also disciplines the system. A model that has to attach a real, citable source to every recommendation cannot wander into confident invention as easily, because there is nothing to cite when it does. The requirement to show the work constrains the work.
The stance has empirical backing. In medical question answering, retrieval-augmented models score higher on faithfulness when each answer must be grounded in a cited source, a pattern reported in studies of citation-enforced prompting for medical RAG.
Summaries get the same treatment. Every line is meant to trace back to the chart, so the clinician can check what is stated. Citations discipline what the model says, though what a summary leaves out needs its own check.
A fabricated citation is worse than none
Showing the work means showing real sources, which means we will not fake them. A source chip on a Meddies answer names a guideline or a reference that actually exists and actually says what we claim. We do not attach invented metrics, implied endorsements, or institutional logos we have no right to. A fabricated citation is worse than no citation, because it manufactures exactly the false confidence we set out to remove.
So the bar for clinical AI is not "sounds like a doctor." It is "lets a doctor check its work in seconds and stay accountable for the decision." A sourced recommendation a clinician can audit beats a confident black box every time, in every specialty, for every patient.
That is why every answer shows its work. It is the only version of clinical AI a careful doctor would let stand beside them.