Medieval scribes spent their days copying texts. AI scribes spend their days doing something scribes never did — interpreting them. That distinction matters, and most people building medical AI get it wrong in exactly the same way.
Here's what a Benedictine scriptorium in the year 1230 can tell you about the scribe we shipped to a psychiatrist last month.
The easy assumption (and why it's wrong)
The word scribe invites a lazy picture: a quiet person in robes, hand cramped around a quill, copying someone else's words verbatim. Modern AI scribes get pitched the same way — "speech goes in, note comes out, done." Speech-to-text dressed up in a lab coat.
That picture is wrong in both eras.
A medieval scribe was not a photocopier. They were a first-pass editor, a verification layer, a craftsman whose hand shaped every sentence. The manuscripts we inherited from the scriptoria of Reims and Wearmouth-Jarrow are filled with marginal corrections, scribal abbreviations expanded, textual variants compared, and the occasional wry note — "Now I've written the whole thing, for Christ's sake give me a drink." They knew what they were copying. They noticed when the exemplar was wrong.
Strip that judgment out and you don't have a scribe. You have a machine that fogs glass.
What they actually shared
Look at a scriptorium operating at steady state and you see four structural features that matter just as much in modern clinical AI:
Scribes were first-pass editors, not copiers. The scribe normalised spellings, expanded abbreviations, corrected obvious exemplar errors, and silently imposed house style. They were not creating the text, but they were not passing it through either. In a modern clinical scribe, the transcript step is maybe 5% of the work. The other 95% is mapping what was said into a structured note the clinician's specialty actually expects. Psychiatry notes don't look like pediatrics notes. Abbreviations get expanded. Medications get reconciled against the chart. That's scribal work, not stenography.
Verification loops were non-negotiable. A scriptorium used a corrector — a second pair of eyes who cross-checked the finished copy against the exemplar and signed off. Without that loop, errors compounded across generations of manuscripts. In an LLM scribe, the corrector is an eval harness: a test suite that runs on every prompt change, catching drift before it reaches a doctor. Skip the evals and your scribe silently gets worse over months. Nobody sees it until a bad note lands in a chart.
The tools shaped the output. Vellum versus parchment, oak-gall ink versus lamp-black, reed pen versus quill — each changed what could be written and how. The scribe's output was not separable from the materials. In modern clinical AI, the equivalent choices are: which LLM, which ASR, which retrieval system, what guardrails, what context you allow. Swap the model and the note changes flavor. That's not a bug. It's physics. Pick the tools deliberately.
Scribes and scholars are different roles. A scriptorium didn't expect its scribes to be theologians. There were monks for that, in separate rooms, on a different schedule. When product teams ask an AI scribe to also diagnose, they're asking a scribe to be a scholar. That's how you get AI "hallucinating" a condition that wasn't discussed. The scribe writes down what happened. The scholar — the doctor — decides what it means. Good systems preserve that division.
What this looks like in code
When we build a clinical scribe, the architecture mirrors the scriptorium layout:
- The scribe — a specialty-tuned prompt that shapes transcripts into notes the clinician's EHR expects. It never diagnoses. It never adds facts.
- The corrector — an eval harness that runs on every deploy, catching drift in tone, format, and factual grounding. It blocks releases that regress.
- The colophon — every generated note carries a footer stating what model produced it, what guardrails ran, what retrieval context was used. If a doctor questions a line, we can retrace it.
- The scholar — the doctor, who has the final say, always. The scribe drafts; the doctor signs.
That's it. Four roles, nine centuries apart, describing the same operating model.
Why this matters now
Every month we meet a healthcare team that has bought or built an AI scribe and is disappointed with it. Almost always, the failure mode is the same: they treated the scribe as a transcription layer instead of an editorial one. They skipped the corrector. They let the scribe drift into the scholar's chair.
The fix isn't a bigger model. It's older than the printing press.
If you're building a scribe — for medicine, legal, voice notes, meetings — borrow the scriptorium's shape. You'll ship better software. And your doctors will get their evening back.
We build AI-enabled software and help businesses put AI to work. If you're introducing a scribe (or any AI layer) into real clinical workflows, we'd love to hear about it. Get in touch.