Why every AI-generated clinical summary should cite its source document

The problem summaries are meant to solve

A patient’s history rarely arrives in one place. Prescriptions from different doctors, laboratory reports from several laboratories, a discharge summary from one hospital and scan reports from another accumulate over years, in different formats and sometimes in different languages. Reading all of it before a short consultation is often impossible, so important details are missed or tests are repeated.

Software that reads those documents and produces a summary can save real time. But a clinical summary is only useful if the clinician can rely on it, and reliance has to be earned by evidence rather than by fluent writing.

Fluent is not the same as correct

Large language models generate text that reads naturally. They can also produce statements that are not supported by the input — a medicine that was never prescribed, a dose copied from the wrong line, a diagnosis inferred from a single ambiguous word. This behaviour is widely described as hallucination, and it is not rare enough to ignore in medicine, where one wrong allergy or anticoagulant dose can cause harm.

Handwriting, poor scans and abbreviations add further uncertainty. A summary that hides where each statement came from gives the clinician no way to tell a well-supported fact from a guess.

What source-backed means

In a source-backed summary, every clinical statement carries a reference to the document it came from — and ideally the page or passage — so the clinician can open the original with one action. A statement such as metformin 500 mg twice daily links to the prescription where it appears. A creatinine trend links to each report that contributed a value.

This changes how a summary is used. Instead of accepting or rejecting the whole summary on trust, the clinician can spot-check the statements that matter most for the decision in front of them.

How a clinician verifies a summary quickly

• Check the high-stakes items first: allergies, anticoagulants and antiplatelets, insulin and other narrow-margin medicines, and any recent critical result.
• Open the source for anything that will change management today.
• Look at the dates: a statement from an old document may no longer be true.
• Treat missing information as unknown, not as absent. A summary that does not mention an allergy has not proven there is none.
• Correct or discard anything that the source does not support.

Drafts, not decisions

An AI-generated summary should remain a draft until a qualified clinician reviews it. The clinician accepts, corrects or discards each part, and responsibility for the clinical decision stays with the clinician, as it always has. Systems should make that review easy and visible rather than presenting generated text as settled fact.

This is not a limitation to be engineered away. It reflects how evidence works in medicine: conclusions are drawn by people who can examine the patient, weigh conflicting information and be accountable for the outcome.

Conflicting documents and uncertainty

Records often disagree. One prescription lists a medicine that a later discharge summary stops; two laboratories report slightly different values; a diagnosis changes after further tests. A good summary does not silently choose one version. It shows both, with their dates and sources, so the clinician can see the conflict and resolve it.

The same applies to uncertain reading. Where handwriting or a scan could not be read reliably, the system should say so rather than guess.

Keeping an audit trail

In clinical work it matters not only what a summary said but who reviewed it and what they changed. An audit trail records when a summary was produced, which documents it drew on, and who accepted, corrected or discarded each part and when. That record supports clinical governance, makes it possible to investigate an error properly, and shows that a human decision stood behind what entered the record.

When you evaluate any system, ask to see the trail for a single patient over a week: who opened the record, which summaries were generated, and which corrections were made. If that cannot be shown, accountability depends on memory.

The limits of any summary

A summary is a way into the record, not a replacement for it. For decisions that are hard to reverse — starting a high-risk medicine, stopping an anticoagulant, planning a procedure — read the key source documents themselves. A good system makes that faster by putting them one click away; it should never make it feel unnecessary.

How Vaidence approaches this

Vaidence AI organises a patient’s records into conditions, medicines, allergies, laboratory values, documents, imaging and referrals, and keeps the original document one click behind every statement. Every output is a draft for a clinician to accept, correct or discard. Vaidence is an early product: it is not clinically validated and does not claim to be, and the public demo at vaidence.com/demo runs entirely on fictional records with precomputed answers.

Questions to ask any clinical AI vendor

• Can every statement in a summary be traced to a specific source document and page?
• How does the system show uncertainty, unreadable text and conflicting records?
• Is output clearly marked as a draft, and how does a clinician correct it?
• Where are records stored and processed, who can access them, and is patient data used to train models?
• What validation has been done, on which data, and what are the known failure modes?
• How are access, changes and reviews logged for audit?

A note on this guide

This guide describes principles for using AI-generated summaries responsibly. It is not clinical guidance for any specific patient, and it does not describe regulatory requirements in any particular country.