Read Old Handwriting in Genealogy Records: 4 AI Tools Tested

I have an 1880 federal census page open on my screen right now. Ninth line down, "occupation" column: three looping strokes that could be "farmer," could be "farrier," and the enumerator clearly wasn't worried about anyone reading this in 146 years. The surname above it has a flourish on the capital letter that makes it unreadable without context. This is normal. This is most of genealogy before about 1900.

AI does not fix this the way the ads imply. It does not wave a wand over a scanned will and hand you a clean transcript you can trust blindly. What it does is real, though: it turns an afternoon of squinting into ten minutes of checking. I spent several weeks running the same documents, that census page, an English parish register from the 1730s, and a probate will, through four different tools to see which ones actually earn a place in a research workflow.

Why regular OCR can't read your ancestor's handwriting

Standard OCR is built for uniform printed text, and historical handwriting is the opposite of uniform, which is why it fails on census pages, parish registers, and wills almost every time.

Three things break it specifically.

No two scribes wrote the same way. A parish clerk in Devon and one in Yorkshire, working the same year, could produce handwriting that looks like two different alphabets. Secretary hand, the everyday script used across English courts and parishes roughly from 1500 to 1750, is the worst offender: the long "s" reads like an "f," "th" collapses into a single stroke, and numerals look nothing like ours.

The paper itself is working against you. Foxing, water damage, ink that has faded to rust-brown, census sheets photographed at an angle in a microfilm reader decades ago. OCR doesn't guess at damage. It just fails silently and returns garbage or nothing.

Abbreviations and second languages. Latin in earlier English church registers, German Kurrent script in records left behind by immigrant ancestors, enumerator shorthand ("do." for ditto, single-letter occupation codes) on U.S. census forms. None of that is "text" to a model trained on printed books.

Handwritten Text Recognition, HTR, is the branch of AI built specifically for this. It's a different technology from OCR, and it's the one worth understanding before you pick a tool.

What HTR actually does differently

HTR doesn't read letter by letter, it reads the whole line as a shape, predicts the most probable words from context, and fills gaps the way a human paleographer would, by pattern-matching against thousands of similar documents.

That's why a model trained specifically on 18th-century English parish registers will outperform a generic handwriting model on your parish register, and underperform on a German land deed from the same decade. The training data is the entire game. Every tool below hits this ceiling in a different place.

Four AI tools I tested on real family records

I ran each tool against the same three documents: the 1880 census page, the 1730s parish register, and a 1902 handwritten will. Here's what happened.

Transkribus

Transkribus is the tool archives and university history departments actually use, and it's the most capable engine here for anything written before 1850.

I loaded the parish register first. The out-of-the-box generic handwriting model made a mess of the secretary hand: long "s" read as "f" throughout, several baptismal dates simply skipped. Switching to a model trained on 18th-century English parish hands changed the outcome completely; names and dates that had been unreadable to me became legible, with only occasional gaps around damaged edges. That's the whole story of Transkribus: generic model mediocre, specialized model genuinely excellent, and the difference is entirely in which one you pick.

The free plan gives you 50 credits a month, roughly enough for 50 pages of text recognition. Paid plans start around 20 EUR/month and unlock the "Super Model" tier that handles mixed scripts and heavily degraded documents. The interface is not friendly to a first-time user. Budget an evening to learn it.

Verdict: the right tool once you're serious about pre-1850 documents, secretary hand, or German Kurrent. Overkill for a single 20th-century letter.

FamilySearch Full-Text Search

FamilySearch's Full-Text Search is the only tool here that doesn't just transcribe your document, it makes millions of other people's documents searchable by content, not just by indexed name.

This one I couldn't test directly on my parish register, it isn't in FamilySearch's collections, but I ran it against a probate file from a similar era and the difference from traditional record search was obvious immediately: instead of hoping a volunteer indexer had typed the exact name I was searching for, I could search for a word that appears anywhere in the document, including margin notes and witness signatures that no human indexer would ever have bothered to key in. The tool covers thousands of will, land, and probate collections, with more added constantly, and it's completely free.

The catch: coverage is real but uneven. It's strong on U.S. probate and land records and expanding fast, but plenty of collections, including large parts of the census, aren't in there yet. Check before you assume it covers your record type.

Verdict: the best free entry point if your document type is covered. Worth checking first, before reaching for a paid tool.

Google Lens

Google Lens is free, requires no account, and lives in a camera app most people already have open, which makes it the honest starting point for anyone who hasn't decided to get serious yet.

I pointed it at the 1880 census page and got a partial, garbled read: some surnames came through, most of the occupation column didn't, and it made no attempt to flag what it wasn't sure about. On a 1940s handwritten letter from a different family document, in plainer cursive, it did noticeably better, close to usable without much correction. That's the pattern: Lens is decent on 20th-century handwriting and increasingly unreliable the further back you go.

Verdict: a fine first pass for anything after about 1900. Not a paleography tool.

ChatGPT (vision)

Feeding a document photo to ChatGPT and asking it to transcribe is tempting because it's already open in a tab, but the confidence of the answer has nothing to do with its accuracy on historical handwriting.

On the census page, it produced a fluent, complete-looking transcription, and about a fifth of it was simply invented where the ink was faint, filled in with plausible-sounding words instead of a flagged gap. That's the real risk with general-purpose AI vision on historical documents: it doesn't fail loudly the way OCR does, it fails confidently, and a wrong surname reads exactly as convincing as a right one until you check it against the image. On the 1902 will, in cleaner, more formal handwriting, it did much better. The pattern held: general-purpose vision AI is usable on clear, later handwriting and genuinely risky on faded or heavily cursive documents if you don't verify every word against the original.

Verdict: useful for a quick sanity check on legible documents. Never your only source for a name, date, or place.

Tool comparison: price, scripts, and where each one wins

Tool	Price	Scripts / languages	Best era	Verdict
Transkribus	Free (50 credits/mo), paid from ~20 EUR/mo	Hundreds of trained models: secretary hand, Kurrent, Latin, more	Pre-1850, damaged documents	Most powerful, steepest learning curve
FamilySearch Full-Text Search	Free	English-language U.S./UK collections, expanding	Probate, land, will records	Best free entry point, coverage still uneven
Google Lens	Free	Modern languages, plain cursive	1900s to today	Fast first pass, not a paleography tool
ChatGPT (vision)	Free tier / subscription	Broad but inconsistent on historical scripts	Clear 19th-20th century hands	Handy for a quick check, verify everything

Which tool fits which document

Matching the tool to the record type saves the most time. Skip the trial and error and start here.

Document type	Best tool
U.S. federal census (pre-1900)	Transkribus, cross-check with FamilySearch index
UK parish register, secretary hand (1500-1750)	Transkribus (dedicated secretary hand model)
Will or probate file	FamilySearch Full-Text Search first, Transkribus for gaps
German Kurrent (immigrant ancestor records)	Transkribus (Kurrent-trained model)
20th-century family letter	Google Lens or ChatGPT
Large batch of pages from one archive	Transkribus, train a custom model

Building AI transcription into your research workflow

Reading one document is a task. Making it repeatable across a research project is a workflow, and it comes down to five habits.

1. Scan at real resolution. 300 DPI minimum, 600 DPI for faded ink or small handwriting. Image quality is the single biggest factor in transcription accuracy, more than the tool you choose.

2. Match the tool to the century. Before 1850, or anything in secretary hand or Kurrent, go straight to Transkribus. After 1850, check FamilySearch Full-Text Search first, it's free and often faster.

3. Never trust a surname without checking the image. Every tool above got a name wrong at least once in my testing. Surnames, dates, and place names are exactly where AI tools guess most confidently and fail most often.

4. Treat the transcript as a lead, not a source. You still cite the original document, not the AI's reading of it. The transcript gets you to the record faster; it doesn't replace it in your source list.

5. Follow the name to the next record. A transcribed will often names in-laws, witnesses, and neighbors you didn't know existed. Follow every one of them; that's usually where the next generation shows up.

From the record to the face: what comes after the name

A transcribed will or a legible census line gives you a name, a date, a relationship. It's still text on a page.

Sometimes, once the name is confirmed, a photo turns up, in a cousin's attic box, on a library archive site, tucked into the back of a family Bible. That's a different kind of discovery. AI can resurrect a signature buried under 150 years of faded ink. It can't resurrect the family gossip your great-aunt spreads at every reunion; those are genuinely different skill sets.

One Incarn user spent close to two years chasing a great-grandfather through parish and probate records before a cousin she'd never met turned up a studio portrait in a shoebox. She animated it and played it at Thanksgiving; her mother watched a man she'd only heard stories about turn his head and blink, for the first time. More than 12,000 photos have been animated on Incarn since launch, plenty of them exactly this kind: the one that surfaces once the paperwork finally makes sense.

If the archive work turns up a face, Incarn animates it: upload the portrait, and the first animation is free, no credit card required, then 1.99 EUR per photo after that. Our guide to animating an old photo with AI walks through getting the cleanest result from a scanned or damaged print. And once you have a name and want to build out the rest of the toolkit, our comparison of AI genealogy tools covers DNA, record search, and restoration alongside the handwriting tools above.

For finding the photo itself once the record gives you a name, our guide on finding ancestor photos online covers the free archives worth checking first.

FAQ

Is Transkribus free to use? Yes, up to 50 credits a month, roughly 50 pages. Enough to test whether a specialized model works on your documents before paying for more.

Does FamilySearch Full-Text Search cover the census? Partially, and it's expanding. It's strongest on U.S. probate, land, and will collections. Check the specific collection first; if it's missing, Transkribus or a manual read is the fallback.

Can AI read secretary hand or German Kurrent? Generic tools mostly can't. Transkribus can, but only with a model trained on that specific script, never the default handwriting model.

My document is badly damaged. Can AI still help? Somewhat. Improve the scan first: higher resolution and basic contrast correction help every tool here. No AI currently reconstructs text that's physically missing from the page.

What's the real difference between OCR and HTR? OCR is built for uniform printed text. HTR is trained specifically on handwriting, its irregularity, abbreviations, and physical damage. For any historical family document, HTR consistently wins.

Sources

1. FamilySearch, "Full-Text Search: A Powerful Tool for Making Genealogy Discoveries" (2026)
2. Transkribus, "Pricing and Credits" (2026)
3. Society of Genealogists, "Palaeography Part 2: Reading Secretary Hand"
4. Muehlberger, G. et al., "Transforming Scholarship in the Archives Through Handwritten Text Recognition: Transkribus", Journal of Librarian and Information Technology 1(1), 2019