Reading the Unreadable: How AI Broke a 200-Year Brick Wall

Dynasty House Academy · · 10 min read

Reading the Unreadable: How AI Broke a 200-Year Brick Wall

A field report from a real research session: handing two-century-old church registers to a vision AI, letting it read the handwriting, and writing the findings back into a tree over an open API. The wins, the wrong turns, the false friends, and a brother the records had forgotten.

This is a guest field report — a write-up of one real research session, set down while the work was still warm. The subject is a Lutheran cottager named Johann Zieger, born around 1790 in a village on the old Austro-Hungarian border, who died before 1851 and took the names of his parents with him. Below him the tree was full and sourced. Above him: a blank where four generations should have been. The classic brick wall.

It came down in a single sitting. Not with a paid lookup or a lucky hint, but by handing two-hundred-year-old parish registers to an AI, letting it read the handwriting, reasoning over the whole book at once, and writing the results back into the tree over an open API. It worked — and it also went wrong in useful ways. The wrong turns are the more honest half of the story, so they are here too.

Why the wall was a wall

The records existed the whole time. That was never the problem. The problem was that they were unreadable in every sense a search box cares about.

They were written in Kurrent, the old German cursive that looks, to a modern eye, like a seismograph having a difficult day. They switched language mid-volume — German one year, Hungarian the next — because the border did. The place names had three forms each: Heiligenkreuz to a German clerk, Rába-Szent-Kereszt to a Hungarian one, Újkörtvélyes on a later map. And the family was Lutheran in a mostly Catholic county, which meant every search that assumed the wrong church came back empty. None of this is indexed. You cannot type a name into a box and find it. Someone has to read the books.

The approach: read the whole book, not the one line

The usual way to use a parish register is to hunt for a single entry and squint at it. We did the opposite. We pulled every page of the relevant registers at full resolution — the marriage books from 1784 to 1866, then the baptism book from 1819 to 1839 — and sent each two-page spread to a vision model with one instruction: transcribe every entry, every column, keep the original spelling, and flag anything bearing the family surname.

The output was not a search result. It was the entire book turned into clean, structured text you can reason over — hundreds of marriages and baptisms with bride, groom, both sets of parents, witnesses, dates and villages laid out in rows. Once a forty-year register is text, the questions change. You stop asking "is this the entry?" and start asking "show me every Zieger father across four decades, with the mother's maiden name." That is a different kind of genealogy.

The false friends, the spelling chaos, and a small panic

The transcriptions immediately taught us how wrong our assumptions had been. A few that are worth passing on:

One family, five spellings. The surname appeared as Zieger, Zinger, Zeiger, Ziger and Czieger — sometimes two spellings in a single entry, in one clerk's hand. Spelling was not standardised until civil registration arrived late in the century. The lesson: treat the name as a sound, not a string. A search that demands exact spelling will miss its own subject.

A word that means the opposite of what it looks like. The men were repeatedly called Söldner. In modern German that reads as "mercenary." In these manorial records it means a cottager — a smallholder with a house and a garden and little field land. The clerks proved it themselves: they wrote Soldat, a different word, the forty-one times they actually meant a soldier. Reading old records is half palaeography and half resisting the modern meaning of the word in front of you.

The surname scare. A web search for the name surfaced a Jewish family of the same spelling from the same region, which is the kind of thing that makes you stop and wonder. The records settled it calmly: this family was baptised, married and buried Lutheran for generations, in the church's own books. A shared surname across two communities is a coincidence of German vocabulary, not a shared ancestry — and the documentary trail, not the hunch, is what decides it.

Being confidently wrong, on the record

Midway through, a clever theory took hold. The wife was recorded in one late document as "Elisabeth Peter," yet no Elisabeth Peter ever turned up as a bride in eighty years of marriages. But a surname Petermann did appear, married into the family more than once. Clerks clip names; "Petermann" to "Peter" is an easy slip. It was a tidy explanation, and it was wrong.

The baptism book overturned it without ceremony. There, in 1824, in plain script: a son baptised to Johann Zieger, cottager of Heiligenkreuz, and Elisabeth geborne Peter. Then again in 1830, and again in 1833. The model that produced the transcriptions even distinguished the two surnames on the same pages, writing "Petermann" for a different woman entirely. The maiden name was Peter. The neat theory died on contact with the primary source, which is exactly what neat theories are supposed to do. We note it here on purpose: the value of putting the whole book in front of you is that it can prove you wrong, fast, before a guess hardens into a "fact" in someone's tree.

Where the negatives did the work

The marriage of Johann and Elisabeth was never found — not in either marriage book, across 1784 to 1866. That sounds like failure. It was information. Their first child was baptised in August 1824, so they married around 1823, which lands precisely in the seam between one marriage volume that ends in 1823 and another whose early years are jumbled and thin. The absence had a shape, and the shape pointed somewhere. A flat "no results" tells you nothing; a transcribed book tells you where the gap is and why.

The breakthrough, and a brother nobody remembered

The clincher was a godfather. The baptism of the known son, Johann, in 1824 named one Georg Groller of Heiligenkreuz as godfather. Nine years later, the baptism of his brother Andreas named the same Georg Groller again. Same parents, same village, same sponsor across a decade — the kind of repetition that turns "probably the same family" into "certainly the same family."

And between those two boys sat a third baptism we had not gone looking for: Michael, born 7 March 1830, same father, same mother, same hamlet. A son who had simply fallen out of the family's memory, recovered because we were reading the whole register instead of hunting one name. Two confirmed birth dates to replace soft estimates, a corrected maiden name, and a brother restored — in an afternoon.

The part that actually matters: closing the loop

Findings that live in a notebook are half-finished. The reason this session ended as tree and not as notes is that Dynasty House has an open developer API. Every confirmation went straight from the transcription into the tree — exact birth dates, the corrected name, the newly found brother linked as a child of the couple — each carrying its citation to the register and the date it was read. Nothing was retyped. The agent did the digging; the tree received the results, sourced, in a home the family owns.

This is the whole thesis of an AI-native tree in one motion. The intelligence does the unglamorous, expert labour — reading dead scripts, holding a forty-year book in its head, refusing to confuse a cottager with a soldier — and an open interface lets the results land where they belong instead of in a forgotten document. Most platforms bolt AI on as a chat box beside a tree it cannot touch. Here the tree is something an agent can responsibly read and improve, with changes that arrive as sourced additions rather than silent overwrites.

What we will and will not promise

We will be plain about the limits. The AI misread pages and occasionally returned nothing at all, and the work was rerun until it was complete; a transcription is a strong lead, not gospel, until a human eye confirms the crucial entry against the image. Vision models are good at this now, not perfect. And no amount of cleverness conjures a record that the fire or the flood already took.

What it changes is the economics of the hard part. A brick wall that needed a specialist, a trip to an archive, and a reading knowledge of two dead scripts can now be approached in an afternoon, with the whole book transcribed and reasoned over at once, and the results delivered sourced into a tree you keep.

This is the work we do as a research service at Dynasty House. Point us at the ancestor where your tree goes dark; we find the parish, read the registers, verify the line, and write what we can prove back into your tree — with sources, in a beautiful and permanent home that stays yours whether or not you keep paying us. The records were always there. They were just waiting for something that could finally read them.

Ready to begin?

Free for your first 100 people. Import an existing GEDCOM in seconds. Your tree stays yours, always.

Found your house →

Keep reading