Even my mum couldn't read this. How AI helped me uncover my family

It began with conversations. Mum, an aunt, cousins: who came from where, who married whom, what Grandma's name was before the wedding. That became the skeleton of the tree on FamilySearch.

Then came the records, meaning entries from parish registers of baptisms, marriages and burials. I searched for them in indexes (searchable lists of such entries), opened scans from the archives and read the Polish-language entries myself. I fixed errors in the tree myself and added a source and a justification to each change. I also wrote to the authors of neighbouring branches.

First, where I searched. Then, where my skills ran out.

At some point searching alone was no longer enough. From 1868, records in the Kingdom of Poland were written in Russian, in handwritten Cyrillic, often in the hand of a clerk from over a hundred years ago. I don't know the language at all, so a record from 1903 looked to me like a cipher.

Mum knows Russian and Cyrillic, but she couldn't read many of these texts. That was when I first asked AI for help. I gave it a scan and asked it to pick out the first names, surnames, dates, parents and witnesses. One problem remained: how do I know the reading is right?

For every Russian record I check several things:

• Surnames and place names. I compare them with the indexes in Geneteka or Lubgens. This is the most important test, because names are the easiest thing to garble.
• Be careful with the index itself. Indexes make mistakes too, and surnames in them are sometimes garbled. I treat them as a second source, not an oracle.
• Consistency with other records. A child's parents should have the same first names in another document. Age, place and occupation should match as well.
• Uncertain passages. I tell AI to mark the passages it isn't sure about instead of guessing.

I admit this isn't full protection. Cyrillic has traps: handwritten т and ц, or п and с, look alike, and AI mixed them up too. That's why no Russian record goes into my tree on trust. First it has to fit the rest of the puzzle.

It follows the thread

From a child's record I take the parents' first names and the mother's maiden name, use them to find their marriage, and the marriage record gives me the next set of parents. That's how I go back generation by generation. By hand it's hours of clicking; AI does it in a few minutes and lists all the siblings as well.

This method wasn't obvious. At first AI searched for children by surname and a range of years, and often found nothing. I found the first marriage myself, by typing the parents' first names and the mother's maiden name into the Geneteka form. On one occasion a marriage was indexed under a garbled surname and only turned up this way too.

It knows spelling variants

A parish clerk wrote surnames by ear, so one surname can be written several ways. When I ask AI for similar-sounding variants, I get a list to try in the index.

It converts dates

In Russian-era records the dates are in the old style, 12 days behind our calendar. AI gives both versions for every record. Without that, an ancestor's birthday lands in the wrong week, and sometimes in the wrong month.

It works while I sleep

One day I said: “I'm going to sleep, keep working on your own.” When I woke up, the records it had read were waiting. Each had a number, a parish, a year, a link to the scan and a confidence rating. That lets me go back to any scan that caught my interest.

Nobody in the family had ever mentioned noble ancestry. I asked AI to count, from the data collected so far, how many townspeople and nobles there were in my family. The number of nobles came out at zero.

Something nagged at me, though. I had once browsed index entries and seen the word “szlachcic” (nobleman) next to one of my ancestors. I just couldn't remember which entry. I told AI it was about my ancestor, Jan. It checked Geneteka and found nothing, and I stubbornly insisted: maybe someone changed the entry later.

Only when I remembered it was a death record did I paste a screenshot of the tooltip on the icon in the “Uwagi” (notes) column, which appears on hover. It said: Wolica, kolonista, szlachcic, 61 l (settler, nobleman, 61 years old). Until then, AI had been reading only the visible text.

An index entry isn't proof, though, because it comes with no scan. Working from my screenshot, AI found the record itself in the Książ Wielki parish death register for 1861 and read it: “kolonista szlachcic żonaty” (settler, nobleman, married). I downloaded the full scan myself.

After that it went fast. In his son's birth record from 1832, Jan is “Urodzony” (in these records, a title used for the nobility) and an estate steward (ekonom). In the 1826 marriage record he is only an estate steward, with no status given.

That entry opened up a whole branch of the family I knew nothing about. From the marriage record I read the names of his wife's parents, from later records I learned who her father was, and under the marriage I found the signature of a brother-in-law.

This story shows how we complement each other. My memory was right, while AI confidently claimed there were no nobles. Only together, my “I saw it somewhere” and its search of the archives, produced a document. That's why the next section is about mistakes.

At first I wrote “find me my ancestors” and got back everything and nothing. AI turned out to be only as good as the context I give it. Over time I settled on a few rules:

• Tell it about your family. Which lines come from where, which villages, which surnames. With that knowledge AI knows where to look instead of guessing.
• Give it a search rule. The one from “following the thread”: “Search by the parents, not by the child.” Since I wrote that down, the results have improved noticeably.
• Treat the age in a record as an approximation. One ancestor was 26, 35 and 34 in records from 1801, 1812 and 1815, although 14 years passed between the first and the last. Since then AI records age as a range and looks for confirmation in another record.
• Have it mark confidence. Certain (I have the record), probable (fits, but unconfirmed), hypothesis.
• Keep notes. AI doesn't remember the previous conversation, so I keep findings, links and a to-do list in files we come back to every time.

Over time AI also started to suggest hypotheses on its own, for example that a marriage may have taken place in a neighbouring parish.

Each of us got our own job. I know the family stories and which aunt to believe. AI has patience for hundreds of thousands of rows and doesn't get tired at two in the morning.

What matters most to me is that AI doesn't tell me what I want to hear. When a scan is illegible, it should say so plainly. When it isn't sure, it should give a confidence level. When it makes a mistake, it should admit it and correct it.

I used to work with Gemini and had three problems with it in genealogy:

• Making things up. When it couldn't read an entry in a register, it didn't write “illegible”. It invented first names, surnames and place names that sounded plausible but weren't in the record.
• Flattery. It told me what I felt like hearing instead of what the source said.
• Circling and empty apologies. It latched onto one thread and kept coming back to it. When I pointed out that it had no basis in the source, it apologised and carried on doing the same.

In genealogy that is dangerous, because an invented surname looks as credible in a table as a real one.

That's why I stayed with Claude. It makes mistakes too, but differently. It worked on real data and didn't invent people. It got a line wrong, but when it hadn't seen something, it said so plainly. After I pointed it to the icon with the tooltip, it wrote (in Polish, translated here): “You were right. The death entry in Geneteka is described as noble, and I hadn't seen that before.”

To be fair: this is my experience with two tools, not a comparison test. Every model makes mistakes, so whichever tool you use, add this to your prompt: “If you don't know something, say so plainly.”