Paleography and AI: deciphering old handwriting, and checking what the machine read

Can AI decipher old handwriting in place of a paleographer?

An AI can produce a first reading of old handwriting, often accurate, including 17th-century manuscript deeds full of abbreviations. It does not replace the paleographer, though: it proposes a text and the researcher validates it. Paleography brings what the machine lacks: the abbreviations of a period, the formulae of a type of deed, the place names and family names of a region. AI saves the time of the first deciphering; paleography guarantees what you cite.

What has changed: conventional handwritten text recognition (HTR) required training a model on one hand or one collection before giving usable results. Recent AI models read the whole page without that prior training. The researcher's work shifts: less copying, more checking.

What makes old handwriting hard to read?

Old handwriting is hard to read for several reasons that compound: letterforms that are no longer ours, abbreviations, spelling that was not standardised, and damaged material. A modern reader, human or machine, stumbles on the same traps: look-alike letters, contracted words, proper names spelled several ways in the same register, faded ink or ink bleeding through the paper.

Which old scripts does wondarchive's AI read?

wondarchive's AI reads typescript, printed text and Fraktur with excellent accuracy. It transcribes handwriting with enhanced review, using its most accurate model: cursive hands in European languages, French included, in Latin, Cyrillic and Greek alphabets. Kurrent and Sütterlin, the German cursive hands used before 1941, are read too; results depend on the hand and the photo, and are best judged by testing on the 100 free pages.

wondarchive est un service en ligne qui transcrit, traduit et indexe les photos d'archives prises en salle de lecture (manuscrits, tapuscrits, imprimés), pour les historiens, les doctorants, les archivistes et les généalogistes.

• Manuscripts: letters, marginal notes, forms filled in by hand; transcribed and translated, high quality, with enhanced review.
• Languages: German, French, English, Italian, Spanish, Dutch, Latin, Russian and other European languages, transcribed in their own language and translated into yours.
• Blackletter prints: quality depends on how sharp the photo is more than on the typeface; doubtful words are given as read, not silently corrected.
• Blank page or empty verso: recognised as such; no text is invented.

The detail by document type is on what wondarchive reads.

How do you decipher a page of old handwriting with AI, step by step?

To decipher old handwriting with AI, you have the machine read the whole page, then check the text against the photo, starting with the most fragile items: proper names, dates, numbers, abbreviations. Paleographic effort then goes to the passages that need it, instead of copying the whole page by hand.

1. Photograph the page sharp and flat; a double-page spread can fit in a single shot and will be read left page, then right page.
2. Upload the folder as the camera wrote it: pages are set upright (the tool asks you to choose when it is unsure), then read from the full-resolution original.
3. Read in two panes: the straightened photo on the left, the text on the right, original or translated in one click.
4. Check first proper names, dates, numbers and every abbreviation.
5. Correct in the text: the machine's version is kept and each correction is dated (correcting the transcription).
6. Fix the vocabulary of the collection in the glossary, so an old term is translated the same way throughout.
7. Note your conventions: abbreviations expanded or not, spelling kept, illegible passages flagged, so you can cite consistently.

How do you check that an automatic transcription of old handwriting is accurate?

You check an automatic transcription of old handwriting by comparing it with the photo: first a few pages read in full, then a systematic check of the risky items, namely proper names, dates, numbers, abbreviations and rare words. A faulty reading is often spotted because it looks too modern: a well-formed present-day word in place of an old spelling, a common name in place of a local surname.

• Read the first pages by the same hand in full: you learn its letterforms and abbreviations.
• Compare the same name from page to page: full-text search finds proper names with their variants.
• Check every date on the photo, digits included.
• Be wary of passages that are too fluent where the original is faint or crossed out.

What makes a good transcription, and how to check it quickly: the reliability of automatic transcription.

Do you still need to learn paleography when you use AI?

Yes, paleography is still needed when you use AI, but differently: it is less about transcribing everything yourself than about judging a reading. Knowing the usual abbreviations, the alphabet of the script you study and the formulae of a type of deed is often enough to spot a machine error and correct it. For a doctoral student or a genealogist, AI also becomes an exercise: you compare your own reading with the machine's, on the same photo.

For deeds of sale, wills, inventories and marriage contracts, with their handwriting and legal vocabulary: working through notarial records.

Follow-up questions

How do you check and render the abbreviations in an old deed?

Do not assume how the machine rendered an abbreviation: check each one on the photo, displayed beside the text. Then correct it according to your convention, abbreviated form kept or expansion flagged; the machine's version is kept.

Does AI read archives in Latin?

Yes: printed Latin is read very well, and Latin manuscripts go through enhanced review. Abbreviations and proper names remain the points to check on the photo. See transcribing and translating Latin archives.

Can AI date a document from its handwriting?

No, wondarchive does not date a document from the shape of its script. It picks up the written date and the dates cited in the text; for an undated document, the content is what informs. See dating an undated document.

Are medieval scripts read?

There is no blanket promise for medieval scripts: the result depends on the script, the language and the photo. Try a few pages before committing a whole collection.

Are my photos used to train the AI?

No: your images are not used for any training. They are hosted in France, and the whole collection can be exported if you leave.

Try it on your own pages

The surest way to judge a difficult hand is to have it read. A free account gives 100 pages transcribed and translated, no credit card needed. To see the result first on real archives, from Humboldt's journal to lighthouse keepers' letters, open the demo.