Can AI transcribe a parish register? I tested three top models

Transcribing parish registers of the Russian Empire (the books of births, marriages and deaths kept by Orthodox churches) is one of the most tedious parts of genealogy. Hours at the screen, letter by letter, record by record. And when the handwriting is messy, you just want to close the laptop.
Iâd long wanted to check: what if I give an AI a folder of scans and say âfigure it out yourselfâ? This week I finally ran a fair test.
The setup
- 20 pages of a parish register from 1917
- Handwriting of average legibility, scans of average quality
- The same prompt for every model
- The task: produce an Excel spreadsheet with the transcribed records
My main goal was minimal involvement: hand over the folder, go have some tea, come back to the result.
Results
Claude Opus 4.6 (via Claude Code) Score: 3/10
It worked for about an hour. Most of the records in the final table were empty: the model transcribed bits and pieces, but there was no complete picture. I donât recommend it for this task.
GPT-5.3-Codex (via Codex CLI) Score: 0/10
An hour and a half of work and zero results. It simply couldnât handle it, and I had to stop it manually.
Gemini 3 Pro (via Gemini CLI) Score: 5/10
The best of the three, with caveats. It worked for about 40 minutes and needed 3 restarts: the model canât manage its context and crashes once thereâs too much data.
It got many of the names, dates and places right. But there are quite a few errors and gaps.
The main thing: you can already take the table Gemini produced as a starting point and fill it in by hand. That really speeds things up, since youâre not starting from scratch.
Honest conclusion
As of today, you canât fully hand over the transcription of a parish register to an AI. None of the three top models I tested could do it out of the box.
But Gemini gives you something you can work with: a draft that you complete and correct. That saves hours.
In the next post, I turn this into an app you can try on your own registers, no terminal required.