Finding ancestors in handwritten registers: what AI does and where it goes wrong

The records exist, the index doesn’t

Most 19th-century church registers are digitised, but you can’t search them. The catalogue goes down to the file: parish, years, type of record. After that come hundreds of scans of handwriting with no text layer — 300–500 scans per register and up to 3,000 per volume. Search engines over digitised archives with text recognition do exist, but for one surname they return around 10,000 mentions, mostly noise: the surname matched a street name, and address books produced thousands of false hits.

So the search goes by geography: village → rural district → county → parish → file. Names and borders have changed since, so the old administrative map comes first. Once a request went to the wrong archive precisely because in those years the town belonged to a different county.

Access costs money and patience

Some archives give scans for free, in full resolution.
Others open them only after signing in through a government portal.
Others sell access by the day: a day costs about 100 roubles, a month about 3,000. Searching an undigitised file to order costs thousands of roubles and takes months.
Genealogy platforms limit views: the daily limit kicked in after anywhere from 20 to 90 pages. After bulk browsing the platform blocked the address for a day.

So the order is: first build a list of files with leaf numbers from the free catalogue, then buy a day of access for that list.

Techniques that work

Index first. Open the register’s title page and its index by year, work out the offset “scan number = leaf + k”, then go straight to the right year.
Strips. Cut one column out of each spread — “names of the born” or “parents” — and dozens of spreads can be read in minutes.
Read it all. To prove there were no other children you have to go through every birth in the parish for 15 years, not find one entry.
Age → year. An age at marriage, in a census or a confession list gives a birth year ±1. In the tree it is marked “about”.
Godparents and sureties. A marriage surety turned out to be the groom’s uncle, a godfather the father’s own brother. A baptism entry reading “son of the late so-and-so” gives an upper bound for the father’s death.
Names. A second son named after a dead brother gives away an unknown first-born. The saints’ calendar restores a month of birth that can’t be read. An entry may say outright that the person was known by another name in daily life — which explains why patronymics differ between documents.
No surname. Rural registers before the 1880s often leave out the surname, and a person is identified by “village + patronymic + spouse”.
Old and new style. A 13-day gap between a register and a gravestone is not a mistake but a calendar conversion. It became one of the proofs that both refer to the same person.
Margin notes. “Extract issued” with a date shows someone requested the document later. One such note led to a person who was doing paperwork half a century on.
Indirect dating. A photo was dated by the river boat in the background: the year it was built and the years it sailed. A letter was dated by its contents.

What AI does well

Reads handwriting, old spelling and abbreviations, and flags doubtful places.
Doesn’t tire on the hundredth scan and lists every entry with the surname, not just the first.
Cross-checks documents: a census against a marriage 30 years earlier, a baptism against a tax census.
Keeps a log for every file: which leaves were read, what was found, what was empty.
Writes letters and call scripts in other languages and builds composites for the tree’s cards.

Where AI goes wrong

Confident tone. In the first session a candidate great-grandfather was called “almost certain”, and a chain of generations grew on top of him. All of it had to be rolled back.
Namesakes. A matching surname and a plausible patronymic are not kinship. One diocese held dozens of families with the same surname.
Source conventions. In an address book a dash means “same surname as above”. Because of that, a stranger’s family briefly became ours.
Recognised text. Text recognition garbled the surname, merged neighbouring entries and gave one person two different ages. Everything is checked against the scan.
Legends. A model readily bends findings to fit the story it was told. For a long time the search followed a family legend about our origins, and a single baptism entry settled it: reading the registers cover to cover took one day.

Rules that keep a tree honest

Into the tree only with a document or relatives’ own words, with a reference to the file and scan.
Anything unproven starts with an “unconfirmed” mark. Next to it: what exactly is unconfirmed and which document would settle it. Links without a direct document carry a confidence percentage.
Namesakes without a proven link go into notes, not the tree.
Dates before 1918 in the old style, as in the register. Places as they were called then.
Scans are taken in full resolution from the archive’s own site. They are not “enhanced”: that only lowers the quality.
Empty results are logged too, so nobody opens the same file twice.

How the work is set up

AI searches, reads, cross-checks and enters what is confirmed into the tree. Signing in to archives, paying for access and deleting data happen only with the owner’s approval. The family’s answers changed the picture more than once, so questions for them are collected in a list as the search goes on.

When the registers run out

For the 20th century there are no church registers. Records older than a hundred years go to state archives, newer ones stay with the civil registry, and an extract from a civil record can be requested through the government portal. For the war years there are open databases of awards and losses; a negative result there tells you something too. Living relatives turn up through platform matches, forums, local history groups and old school photos. A grave paid for years ahead means there are living relatives, and the municipality can forward them a letter. But everything found this way goes through the same rule: without a document it is a hypothesis.