4 AI Tools to Decipher Old Handwriting (2026)
I tested 4 AI tools on real parish registers from the 17th to 19th century. Transkribus leads on complex scripts; MyHeritage's Scribe AI is the real surprise since March 2026.
AI & Technology Writer, Incarn
In short
After testing 4 AI tools on 17th and 19th century parish registers: Transkribus is unmatched on complex scripts and damaged documents, but has a learning curve. MyHeritage's Scribe AI (March 2026) is the real surprise: it transcribes AND interprets historical context. Geneatique 2026 works if you're already in that ecosystem. Google Lens handles letters from the 1940s-1960s but falls apart before the 18th century.
TL;DR: After testing 4 AI tools on 17th and 19th century parish registers: Transkribus is unmatched on complex scripts and damaged documents, but has a learning curve. MyHeritage's Scribe AI (March 2026) is the real surprise: it transcribes AND interprets historical context. Geneatique 2026 works if you're already in that ecosystem. Google Lens handles letters from the 1940s-1960s but falls apart before the 18th century.
I'm looking at a birth record from 1743. Period calligraphy, browned ink, brittle paper. My great-great-grandfather is mentioned somewhere in it: the archivist confirmed it. But reading it alone, without training in paleography, is like reading Greek without knowing the alphabet.
AI tools are changing that. Not completely, not magically, but they shift the needle toward the reader and away from the specialist.
Here's what I found after several weeks testing four tools on real family documents.
Why standard OCR fails on your family archives
OCR, optical character recognition, is the technology that lets your phone photograph printed text and convert it into editable text. On modern documents, it works well. On historical archives, it fails almost systematically.
Three concrete reasons.
Historical cursive script isn't standardized. Every priest, notary, and secretary had their own hand. The same word could be written twenty different ways depending on the region, the era, and the scribe.
Documents are physically degraded. Tears, ink stains, faded ink, translucent paper: OCR doesn't predict what's missing: it fails the read and stops there.
Abbreviations and Latin. Parish records from the 17th and 18th centuries use abbreviated Latin, local notation conventions, repetitive formulas that only an algorithm specifically trained on this type of document can decode.
HTR, Handwritten Text Recognition, is the branch of AI designed for these problems. That's where everything happens in 2026.
HTR: what this technology actually does
HTR doesn't read characters one by one. It analyzes the writing flow as a whole, predicts probable words based on context, and fills in illegible areas with what similar documents from the same period would typically contain.
That's why a model trained on 18th-century Norman parish registers will be significantly better on your Norman 18th-century archives than on Provençal notarial records from the same era. The training data is the key variable.
The four tools I tested all use HTR, but with very different models, interfaces, and use cases. For a broader view of AI tools useful in genealogy (DNA, photo restoration, animating ancestors), see our comparison of the 11 best AI genealogy tools 2026.
The 4 tools tested: Transkribus, Scribe AI, Geneatique, Google Lens
Transkribus
Transkribus is the world reference in HTR for historical documents. Developed at the University of Innsbruck, it's used by national libraries and historical research projects across Europe.
What sets it apart: hundreds of pre-trained models, organized by language, era, and document type. You select a model calibrated for 18th-century French parish registers, and the results on well-preserved documents are striking.
What slows you down: the interface isn't intuitive. The learning curve is real. And the credit system can get expensive if you have many pages to process. Count 10 to 30 credits per page depending on length and image quality.
My verdict: indispensable for researchers handling large volumes or difficult documents. Not the right entry point for a first exploration.
Scribe AI (MyHeritage, March 2026)
In March 2026, MyHeritage® integrated Scribe AI into its toolkit. The comparison with what existed before is striking.
Scribe AI doesn't just transcribe handwritten text. It interprets it. It recognizes standard Latin abbreviations from parish records, conventional formulas, period spelling variations, and produces a modern-language summary of what the document says, with names, dates, and family relationships clearly highlighted.
What sets it apart: it's accessible directly from the MyHeritage® interface, with no additional configuration. If you already have a tree on the platform, extracted data can be integrated directly into your family tree.
What slows you down: on heavily degraded documents, Transkribus remains superior. And advanced features require a MyHeritage® subscription (Full plan, around 258€/year, according to MyHeritage®).
My verdict: the best option for the vast majority of occasional researchers. The contextual interpretation, not just raw transcription, genuinely changes the experience.
Geneatique 2026
Geneatique is a French genealogy software, published by the Société Généalogique et Héraldique de France since 1987. Its 2026 version integrates AI assistance for manuscript reading, accessible via a credit system. Since 2025, a Linux version exists, first time in 39 years.
What sets it apart: it's designed for users who already do their genealogy in Geneatique. No need to switch tools: the AI integrates into the existing workflow.
What slows you down: transcription quality falls below Transkribus and Scribe AI on complex documents. The credit-based model can feel opaque.
My verdict: relevant if you're already a Geneatique user. Not a reason to adopt it for the AI function alone.
Google Lens
Free, immediate, no account required. Photograph the document with your phone, point at the text, Lens attempts a transcription.
What works: handwritten letters from the 1930s to 1970s, printed documents with handwritten corrections, signatures on 20th-century records. Lens is surprisingly effective for these cases.
What fails: almost anything before 1900. 19th-century cursive, parish records, handwritten Latin, Lens consistently falls short.
My verdict: a free starting point for recent documents. Not a paleography tool.
Which tool for which document type
| Document type | Recommended tool |
|---|---|
| Parish register (17th-18th c.) | Transkribus |
| Notarial record (19th c.) | Scribe AI or Transkribus |
| Family correspondence (20th c.) | Google Lens or Scribe AI |
| Census, civil records (19th-20th c.) | Scribe AI |
| German gothic script (Kurrent) | Transkribus (Kurrent model) |
| Large volume of pages | Transkribus (custom-trained model) |
Integrating AI into your genealogy workflow: 5 steps
Deciphering a document is one step. Integrating it into your research is another.
1. Digitize at high resolution. 300 DPI minimum, 600 DPI for damaged or fine-text documents. Image quality is the number one factor determining transcription accuracy. A poor scan produces a poor transcription, even with the best tool.
2. Choose the tool by century. Before 1850, go with Transkribus using a model matched to the region and period. After 1850, Scribe AI is often sufficient and faster.
3. Verify proper nouns systematically. AI tools often transcribe surnames with errors, especially regional or uncommon names. Always compare the transcription against the original image on critical elements: first names, surnames, dates, places.
4. Use the transcription as a lead, not a source. AI transcription assists reading: it doesn't replace the original record. Always cite the primary document, not the tool's interpretation.
5. Cross-reference with online archives. Once a document is transcribed, search the extracted names on Geneanet, FamilySearch, or French departmental archives online. One record often opens onto other family branches. Our genealogy tips for beginners walk through this method step by step.
To find photos connected to the ancestors you locate, see our guide on how to restore old photos with AI.
From the archive record to the photo: giving a face to the ancestor
Deciphering a birth record means finding a first name, a date, a family. But it stays text.
Sometimes, digging through family albums, photo archive sites, distant cousins: a photo eventually surfaces. A studio portrait from 1910. A group photo in front of the family farmhouse. A face, finally.
A user wrote to us after animating the photo of a great-grandfather found in notarial archives. He'd searched the registers for three years. The photo was in a box at a cousin's he'd never met. The animation was sent to the whole family for Christmas: the first time the children "saw" this ancestor move.
That's where Incarn comes in. The tool animates old portraits: from a still photo, it generates a short clip where the ancestor moves, breathes, turns their head slightly. Free first try, then 1.99 per photo. And if you want to check which tools let you animate a photo for free, no heavy sign-up or credit card, our comparison of 6 options lays it out.
To understand how to prepare your photos for the best animation results, our guide on restoring old photos with AI covers that step.
Frequently asked questions
Is Transkribus free? It offers a free tier with a limited number of monthly credits, enough to test quality on a few documents. For regular use, paid plans start around 10€/month.
Does Scribe AI handle Latin documents? Yes, that's one of its strengths. It recognizes the standard Latin formulations in parish records and automatically translates them into a modern-language summary. That's what sets it apart from raw transcription tools.
Can AI read German gothic script (Kurrent)? Transkribus is the best option for that. It offers specific models for 19th-century Kurrent, useful if your ancestors come from Alsace, the Moselle region, or areas with strong Germanic influence.
My documents are heavily damaged. Can AI still help? Partially. Start by improving image quality with a restoration tool, Transkribus gives better results on enhanced images. Our article on restoring old photos with AI details the options available.
What's the difference between HTR and OCR? OCR is designed for uniform printed text. HTR is specifically trained on historical handwriting, with its variability, abbreviations, and physical degradation. For family archives, HTR is always superior; standard OCR isn't suited to it.
Does FamilySearch also offer this type of tool? FamilySearch developed its own assisted transcription system (Full-Text Search), which made hundreds of millions of historical documents searchable. It's complementary to Transkribus: FamilySearch works on its own archives, Transkribus lets you work on yours.
AI & Technology Writer, Incarn
Thomas covers AI and machine learning applications for creative tools. Former research engineer with a focus on computer vision and video generation.
Deep Nostalgia® and MyHeritage® are trademarks of MyHeritage Ltd. Incarn is not affiliated with MyHeritage. Comparisons (features, pricing) are provided for information only and reflect publicly available offerings at the time of publication; they are subject to change.