---
title: "AI Handwriting Recognition for Genealogy"
description: "AI handwriting recognition can now give you a usable first draft of many old records, but where you meet it decides what it can read, and the script matters more than the brand. Here is how it works..."
canonical: "https://kleiobase.com/blog/how-ai-is-transforming-genealogy"
last_updated: "2026-10-18"
---

# AI Handwriting Recognition for Genealogy

AI handwriting recognition can now give you a usable first draft of many old records, but where you meet it decides what it can read, and the script matters more than the brand. Here is how it works, what the published tests show, and what it still cannot do.

Claims about  last verified 2026-10-18.


You have found the record. It is the right parish, the right year, possibly the right family, and you cannot read a word of it. The will is in secretary hand, or the baptism is an eighteenth-century Spanish entry dense with abbreviations, or the birth entry is in a Cyrillic cursive that drops letters the modern alphabet no longer has. The record survives. The trouble is the handwriting.

This is the problem AI handwriting recognition was built for, and it has moved quickly since general AI models learned to read images. It has also produced a lot of confident marketing, and a lot of people who tried one tool on one page and concluded either that AI reads everything or that it reads nothing.

**The short answer.** AI handwriting recognition, usually called HTR, learns to read old handwriting from pages people have already transcribed by hand, so it is strongest on the scripts and languages it was trained on and weakest everywhere else. You meet it in three places: inside an archive's own full-text search, in tools that transcribe a document you upload, and on specialist platforms such as Transkribus. In a 2024 study of 18th- and 19th-century English handwriting, general AI models misread 5.7 to 7 per cent of characters, and only a second correction pass brought that below 2 per cent. Either way, you check every name against the image. Treat what it gives you as a first draft that you correct.

## Handwriting recognition is trained, not programmed

Nobody writes rules for what a secretary-hand "e" looks like. An HTR model is shown thousands of page images alongside human transcriptions of those same pages, and it learns the mapping between the two. This is why it works well on some documents and badly on others.

FamilySearch describes its own process plainly in a 2026 guide by David Ouimette: transcribe many thousands of documents by hand in the target language and script, train the model on most of them, test it on the ones held back, and if the quality is inadequate, transcribe more and repeat. The guide gives 90 per cent word accuracy as an example of the line below which a model goes back for another round. That is a bar for searchability across billions of images, not a promise about your page.

**Two families of model now do this work.** The older one is specialist HTR, trained specifically on handwriting, often on a single script or archive. The newer one is the general multimodal model, the kind behind ChatGPT, Gemini and Claude, which was not built specifically for handwriting and needs no training on your script. They fail in different ways. Transkribus describes its general Text Titan models as built for languages in the Latin script, for example, so a Russian revision list needs a different model. A general model will attempt almost anything, and unless the tool is built to flag doubt, a misread word appears on screen exactly like a correct one.

## You meet it in three places, and only one reads your own documents

Most confusion about "AI for old handwriting" comes from treating three different things as one. They read different material, and the difference decides what each can do for you.

| Where you meet it | Examples | What it reads | What it cannot read |
|---|---|---|---|
| Archive full-text search | FamilySearch Full-Text Search, Ancestry full-text search | Machine transcripts of that organisation's own image collections | Anything the organisation does not hold, or has not yet processed |
| Upload-a-document transcription | Ancestry document transcription, MyHeritage Scribe AI, general chatbots | The single image you upload | Anything beyond that page |
| Specialist HTR platforms | Transkribus | Your uploaded pages, with a choice of trained models, or one you train yourself | Nothing in principle, but you pick the model and do the setup |

**Archive full-text search is the one most people have already used.** FamilySearch's guide says its Full-Text Search officially launched in August 2025 and covers about 2.5 billion transcribed images in more than 8,000 collections, with about a billion images added each year, and most of those collections have no fielded index at all. Ancestry runs the same idea over its own images. Its explanation of the science behind its full-text search notes that access requires an All Access subscription and that another 300 million record images are being prioritised for the searchable set. Both are very useful, but each reads only the images their owners hold, so the certificate in your grandmother's drawer is not in either. We cover that fence in detail in [which AI tools can actually find your ancestors' records](/blog/which-ai-can-find-your-ancestors-records).

**Upload tools read the page you give them.** Ancestry announced the beta of its document transcription feature on 30 April 2025, for JPG and PNG images of single documents attached to a public tree. At launch, PDFs were not supported and access started at the World Explorer membership level. The same post notes that Ancestry had previously used handwriting recognition to transcribe the 1950 US Federal Census and the 1921 Census of England and Wales, which is the archive-search use of the same technology. MyHeritage launched Scribe AI on 5 March 2026, describing it as transcribing, translating and interpreting an uploaded image of a document or photograph. A general chatbot does the same job with no genealogy framing at all.

**Specialist platforms give you control in exchange for setup.** Transkribus offers a catalogue of trained models and lets you train a private one on your own corrected pages. In our judgement, for a single long register in one clerk's hand, that is usually the strongest route. For forty documents from six parishes in three languages, the setup rarely pays off.

## The script matters more than the brand

Which AI is best at handwriting depends on the handwriting. A model's quality is a property of its training data, and training data is wildly uneven across the scripts genealogists meet.

**English secretary hand.** The everyday hand of sixteenth- and seventeenth-century England, in which several letters bear little resemblance to their modern forms. Crosilla, Klic and Colavizza found that general models favour English, which they attribute to what the models were trained on. That helps, but early modern letterforms are still a different problem from modern cursive.

**Spanish, Portuguese and Italian records.** After English, these are where FamilySearch's archive search is largest. FamilySearch lists its top languages by volume as English, Spanish, Italian, Portuguese, French and Dutch, and its top countries include Italy, Spain, Portugal, Brazil, Colombia and Mexico alongside the United States, the United Kingdom, Canada and Australia. Its guide is candid that most of its models are trained on Latin script, with work under way on Chinese, Arabic, Cyrillic, Japanese, Korean and other scripts.

**Pre-reform Cyrillic.** Russian records written before the orthography reform of 1918 use letters the modern alphabet dropped, such as ѣ, і and ѳ. A tool reading these has to recognise characters a modern Russian reader rarely sees. Documents like the [Russian Empire revision lists](/blog/russian-empire-revision-lists-revizskie-skazki) are where this bites.

**German Kurrent.** The old German handwriting is a script of its own, and a tool that knows German the language has not necessarily seen much Kurrent the script. Some vendors blur the two. We went through what each tool's own model catalogue shows in [which AI reads German church records](/blog/which-ai-reads-german-church-records-kurrent-latin).

In practice, if your record is a Mexican notarial act, an American deed or a register from Italy, the archive may already have read it for you. If it is a Cyrillic revision list or anything else outside the Latin script, such as a Greek parish register, you will probably be the one uploading it.

## The published accuracy figures disagree, and all of them are honest

Accuracy is normally reported as character error rate: the share of characters the machine got wrong against a human-checked transcription. Three serious tests in the last two years point in different directions.

**General models won on English.** A team led by the historian Mark Humphries, posting to arXiv in November 2024, tested commercial multimodal models from OpenAI, Anthropic and Google on 18th- and 19th-century English handwritten documents. They reported character error rates of 5.7 to 7 per cent and word error rates of 8.9 to 15.9 per cent, improvements of 14 and 32 per cent over specialist HTR software such as [Transkribus](/compare/transkribus-alternatives). When the models were used to correct existing transcriptions, character error fell as low as 1.8 per cent.

**Nobody won consistently on historical European material.** Crosilla, Klic and Colavizza, publishing in the *Journal of Documentation* in 2025, benchmarked proprietary and open-source models against Transkribus on English, French, German and Italian handwriting. They found the general models strong on English and modern handwriting, weaker on other languages, poor at correcting their own mistakes, and with no consistent advantage over Transkribus on historical documents.

**The specialist won on its own archive.** In June 2025, Transkribus benchmarked its Text Titan I ter model on 2,000 pages drawn from a corpus of 7 million pages across about 9,000 archival sources, excluded from training. It reported 10.8 per cent character error on handwritten text for its model against 35.0 per cent for the best general model it tested, Gemini 2.5 Flash. That is a vendor testing its own product on a set it chose, which is data rather than a verdict, and Transkribus has since released a successor model, Text Titan II.

Those figures measure different corpora and should not be set against one another, but the pattern is durable even as the numbers age: general models are strongest where the training data was richest, and specialist models are strongest on the material they were built for. Neither figure tells you about the one word you care about, which is why a 5 per cent error rate can still mean the surname is wrong. We take the metric apart, and set out a checking routine, in [how accurate AI is at reading old handwriting](/blog/how-accurate-is-ai-handwriting-transcription).

## What AI handwriting recognition cannot do

The limits are where your own work starts.

**It cannot judge the source.** A document might be a contemporary register or a later copy, an official entry or a family Bible note with no provenance. The model reads what is on the page. Whether the page deserves to be believed is a question about the record, not the handwriting.

**It cannot settle a contradiction.** One record says 1847, another 1849, a third "about 1850". An AI can line them up and flag the conflict. Deciding which is right means asking who supplied each date and how close to the event they were.

**It does not always tell you when it is guessing.** A general model can produce a misread surname as fluently as a correct one. Unless the tool is built to flag uncertainty, nothing on screen tells you which names to check, so you check all of them.

**It cannot read what was never digitised.** If the register burned, or sits unfilmed in a diocesan archive, there is nothing to recognise.

## What KleioBase does with handwritten records, and what it does not

We are a tool in this space, so read this section as disclosure.

KleioBase is an AI research companion that turns scattered historical records into one connected family archive. Reading the handwriting is the first step, not the product. When you upload a JPG, PNG or PDF page, the AI produces a transcription in the document's original language and script, which is never translated, alongside a translation and narrative in your chosen language, and it pulls every named person, date, place and relationship into structured fields. Processing takes about one to three minutes. You add optional context first, such as the language, place or approximate year, which guides the reading.

Nothing joins your archive until you have [reviewed and confirmed it](/docs/uploading-records). Confirming creates or links person profiles across every record that mentions them, and starts [duplicate matching](/docs/finding-duplicates) in the background, so the Maria in an 1890 baptism and the Marya in a later census can surface as one possible person with the evidence shown. The [research companion](/docs/research-companion) can then answer questions across the whole archive, such as what is still missing for a great-grandfather, by searching your people, records, families and timeline.

For difficult handwriting there is a deep scan, which uses a stronger model and costs 5 credits instead of 1. When a deep scan is unsure of a name, it says so: the review pane shows a warning and tags the individual names to verify against the image.

There are also things we do not do. We have never published an overall accuracy percentage, because handwritten records range from colonial American deeds to Italian registers to Cyrillic revision lists, across several centuries, and an average over all of them would say little about any one document. If you have hundreds of pages written by the same clerk, training your own model in Transkribus on corrected pages from that register will probably read it better than we do. And we only work with the images you upload. The companion's web search reads public pages and cannot search inside FamilySearch's or Ancestry's collections, so if your record is in one of those, use their full-text search to find it.

## Where to start with the record you cannot read

If the image came from FamilySearch, try Full-Text Search before anything else. Search for the parish, the place or a neighbour as well as your ancestor's name, since the page may already have a machine transcript you can search.

If it did not, work out which script you are looking at before you pick a tool, because the tools differ most by script. For census pages, our guide to [reading old census records](/blog/reading-old-census-records) covers the column-by-column habits that also help when checking a machine reading. It is also worth running the same page through two different tools: words they read the same way are probably right, and words they read differently are the ones to check yourself.

Whatever tool you use, compare the names, dates and places with the image before you put them in your tree, and keep a note of which record and page each one came from. The errors that matter rarely turn up in the standard phrases of a register. They turn up in the handful of words that identify your family.

Canonical: https://kleiobase.com/blog/how-ai-is-transforming-genealogy