I Spent 30 Days Using AI to Trace My Family Tree — Here's What Actually Worked

Family history is the one hobby where being slightly wrong is worse than knowing nothing. A guessed date quietly becomes a fact, the fact gets copied into someone else's tree, and within a year a stranger you are not related to is hanging off your branch. So when I gave AI thirty days on my own family tree, I was not looking for shortcuts. I wanted to know exactly which parts of the work it could take off me, and which parts it would get confidently, plausibly wrong.

Both answers turned out to be sharper than I expected. There is real, unglamorous value here — the kind that saves an afternoon rather than solving a mystery. There is also a failure mode so convincing that I think it is the single biggest risk facing amateur genealogy right now. Here is the honest account of both.

What AI genuinely solved for me

The wins were almost all about friction rather than discovery. Nothing found me an ancestor. What it did was remove the small obstacles that turn twenty minutes of research into an entire evening.

Handwriting was the big one. Anyone who has stared at a nineteenth-century parish register knows the specific misery of one illegible word standing between you and a name. Photograph the page, hand it to a capable assistant, and you generally get a usable transcription back — not perfect, but enough to break the deadlock. It read a badly faded occupation for me that I had been squinting at on and off for weeks. It turned out to be a trade I had never heard of, which is exactly why I could not read it.

Second was context. Old records are full of vocabulary nobody has used for a century: obsolete occupations, legal phrasing in wills, Latin in older church entries, county boundaries that have since moved. Asking for a plain-English explanation of a phrase is fast, and — importantly — this is the category where the answer is easiest to check, because you can verify a definition in seconds.

Third, and the one I underrated going in, was direction. Describing what I had and asking what I was missing produced genuinely useful suggestions about which record set to try next. It is not magic; it is a well-read assistant with no ego, available at eleven at night. If you have never worked this way, my complete ChatGPT playbook for beginners covers the prompting habits that make this kind of back-and-forth productive rather than vague.

The failure that should worry everyone

Now the part that matters more than any of the above.

Somewhere in the second week I asked for help placing a great-grandfather I had partial information about. What came back was excellent. A birth year, a parish, a marriage, a plausible occupation for the time and place, and — this is the bit that stopped me — a citation. A named record set, formatted correctly, exactly the sort of reference I would have written myself.

None of it existed. Not the record, not the reference, not the man. It had produced a completely fluent, entirely fictional ancestor, and it had done so in the same confident register it used for the things that were true.

"The danger is not that AI gets genealogy wrong. It is that it gets it wrong in precisely the format you were expecting to be right."

This is not a flaw you can prompt your way out of, because it is what the technology does. A language model produces what a plausible answer looks like. Given a thin question about an ordinary person from 1880, a plausible answer is a name, a date and a parish — so that is what you get, whether or not such a person ever drew breath. I wrote at more length about where this tendency shows up in everyday use in my piece on using AI honestly for creative work, but genealogy is where it does the most lasting damage, because a fabricated ancestor does not stay in your notes. It propagates.

The workflow I ended up with

After a month of trial and error, the process that survived is unremarkable, which is probably why it works.

I gather first, without AI anywhere near it — census images, certificates, parish entries, whatever I actually hold. Then I use AI on the documents I have, rather than asking it about people I am looking for. That distinction is the whole thing. Transcribing a record in front of me is a task with a right answer that I can check. Asking who my great-grandmother married is an invitation to invent one.

Once a body of verified material builds up, feeding it into a tool that works strictly from sources you supply is far safer than open-ended questioning. That is the approach I described in the guide to building a personal research assistant with NotebookLM — it answers from your documents rather than from a general impression of the nineteenth century, and it will tell you when your papers do not contain the answer. For family history that constraint is not a limitation. It is the entire point.

Last comes the writing. Turning a verified life into a few readable paragraphs for relatives who will never open a census index is where AI is straightforwardly good, and where nothing is at stake — the facts are already established, and it is only prose.

Is it worth doing at all?

Yes, with the guardrail. Thirty days in, my research moves noticeably faster and my tree has not gained a single unverified name. The time saved is real but narrow: it is the hour you no longer lose to one unreadable word, the dead end you avoid because something suggested a record set you had forgotten, the Latin you no longer have to look up twice.

What has not changed at all is the verification. That still takes exactly as long as it always did, and it should. Anyone selling you a version of this hobby where the checking gets automated away is selling you a tree full of strangers.

If you want the wider view of which assistants are worth your time and what each is actually good at, I keep that current on the AI tools hub. And because a month of notes did not fit in one article, I put the full method — the prompts, the record-by-record workflow, and the verification checklist I use — into Tracing Your Roots With AI.

Questions people actually ask

Can ChatGPT actually read old handwritten records?

Partly, and it is worth being precise about this because the tools are unreliable narrators about their own abilities — ask ChatGPT whether it can read handwriting and the answer you get back may not match what it actually does. In practice you can paste a photo or scan of a record into any of the major assistants and ask for a transcription, and on a clean, well-lit page it will usually get you most of the way there. Faded ink, tight cursive and the abbreviations that census enumerators used are where it starts guessing. Purpose-built handwritten text recognition such as Transkribus is trained specifically on historical documents and handles connected cursive and tabular layouts far better. My habit is to use a general assistant for a first pass and a specialist tool when the page matters.

Is AI accurate enough to rely on for genealogy research?

Not on its own, and this is the part people learn the hard way. A language model is built to produce a plausible answer, and a plausible-looking ancestor with plausible-looking dates is exactly what it will hand you when it has nothing solid to work from. It will also cite records that do not exist, in convincing detail. Treat everything it produces as a lead rather than a fact. The moment a name, date or place matters, it has to be confirmed against a parish register, a census image, a civil registration index or an archive catalogue. AI narrows where you look; it does not establish what is true.

Which AI tools are most useful for family history?

Three kinds, and they do different jobs. General assistants such as ChatGPT and Claude are best for transcription first passes, translating a foreign-language record, explaining an unfamiliar occupation or legal term, and suggesting where to look next. Dedicated handwriting recognition like Transkribus is better on difficult historical documents. And the genealogy platforms themselves — FamilySearch and MyHeritage among them — now build AI into record matching, photo restoration and DNA connections, which is the one place the AI has your actual tree and real records underneath it rather than a general impression of history.

Can AI build my family tree for me?

No, and expecting it to is the fastest route to a tree full of strangers. It can accelerate almost every individual step — reading a document, spotting a pattern across records you have gathered, drafting a readable summary of a life, proposing which record set to search next. What it cannot do is decide that this particular John Wilson is your John Wilson. That judgement rests on evidence you have checked, and it stays yours. The tree is only as trustworthy as the weakest link you accepted without looking.

Does using AI make genealogy research faster?

Substantially, but not in the way people expect. It did not find me ancestors. What it removed was the friction around the finding — the hour lost to a single illegible word, the unfamiliar Latin in a parish entry, the shapeless feeling of not knowing which record set to try next. Over thirty days the research itself moved a good deal quicker while the verification took exactly as long as it always did. That is the honest trade, and it is still a very good one.