Can AI Read Old Handwriting? Which Model to Use, and Why

The short answer (October 2026): yes, increasingly well, and which model you use matters as much as whether you use one. Leo's own model, ATR-1, is tuned specifically for historical documents, handwritten and printed. It reads the image at high resolution and is trained to transcribe what is on the page, not a tidier, modernised version of it. General AI models have improved quickly too: in tests published in November 2025, historian Mark Humphries found Google's Gemini 3 Pro made fewer errors than Claude and OpenAI's models on 18th- and 19th-century English handwriting. But general models, Gemini included, tend to quietly modernise spelling, and that is the error that survives proofreading. Older and harder hands, and material in other languages, vary much more. No published result tells you how a model will do on your page, which is why this guide ends with how to test one.

That variation is where most disappointment comes from. In a 2025 study, the same model scored under 2% character error rate on modern English handwriting and over 40% on a multilingual historical set. "Can it read?" is the wrong unit of measurement. The useful questions are which errors a model produces, whether you can see them, and which model suits which job. A system that garbles a word tells you it struggled. A system that quietly replaces a difficult word with a plausible one does not.

Which model for which job

Different models are good for different jobs, and in Leo you choose the model for the document from one menu in the transcription window. ATR-1 is the default. As of October 2026:

Handwriting, any period: ATR-1, Leo's own model, specialised for historical documents and built to keep the page's own spelling. A general model as a second reading, in its own tab.
Early-modern print, Fraktur, newspapers, or anything you'll quote diplomatically: ATR-1, or a stronger general model such as Gemini 3.8 Flash
Clean, legible print you just want the text of (a 19th-century book, a typescript): A small model such as GPT-6 Luna, or an OCR engine: fractions of a credit an image
A hard reading you need to check: Run a second model into its own tab and compare

ATR-1 is a Transformer fine-tuned to read images of historical documents. Where general models downsample the image, ATR-1 reads it at high resolution, weighing the visual evidence against context. It is trained to transcribe what is on the page rather than normalise it: a long s stays a long s, and archaic spelling survives. It works zero-shot, so there is no training on your hand first, and it costs 1 credit an image. The general models sit alongside it in the same menu, from 0.02 credits an image (GPT-6 Luna) to about 4. Gemini 3.8 Flash, for example, is 0.39. They are useful for cheap clean print, or as a second reading of a hard page. Leo keeps each run in a separate tab beside the image, so comparing two models means reading two tabs, not two browser windows.

Three different technologies get called "AI reading handwriting"

Before comparing results, separate the mechanisms, because they fail in different ways.

Optical character recognition (OCR)

OCR was engineered for machine-printed text with consistent typefaces. It classifies glyphs against a known type model: segment the page, isolate the character, match it. On a clean modern scan this is fast, cheap and highly accurate. Our comparison of HTR vs OCR sets out what that means for historical material.

Handwritten text recognition (HTR)

HTR treats a line as a sequence and decodes it jointly with a context model, because no two writers form a letter identically and the boundary between letters is often not visible at all. The architecture has developed in stages:

hidden Markov models
recurrent networks with CTC decoding
attention-based sequence-to-sequence models
Transformer encoder–decoders such as Microsoft's TrOCR, which handle visual understanding and text generation in one end-to-end model rather than classifying characters one at a time

ATR-1 belongs to this last family: a Transformer fine-tuned on images of historical documents, with a custom architecture rather than a general model behind a wrapper. If you want the fuller technical picture, our guide to handwritten text recognition covers how the field arrived here.

General vision-language models

The chat assistants read an image as tokens alongside text. They are capable generalists, and they are the thing most people reach for first. They are also the least constrained. Their fluency is a language-model fluency, and nothing in their design obliges them to prefer what is on the page over what the sentence probably said.

The distinction matters because the three technologies are often benchmarked as though they were doing the same job. They are not.

What current models genuinely manage

The good news is real and worth stating plainly.

A specialist model now reads historical hands without training on yours. Until recently, a good result on a historical hand meant training a model on that hand first. ATR-1 arrives already trained on historical documents in the Latin alphabet, manuscript and print, from roughly the last 500 years. Upload an image and it transcribes: there is no ground truth to prepare. It is strongest in English and strong across the other languages written in the same alphabet, French, German, Dutch, Spanish, Italian and Latin among them.

Modern handwriting is close to solved. On the standard research datasets of contemporary handwriting (IAM for English, RIMES for French), error rates are now in the low single digits. TrOCR Large reaches 2.89% CER on IAM. In a 2025 study in the Journal of Documentation (arXiv version), general models of the time (2024 versions of GPT-4o and Claude 3.5 Sonnet) were tested without fine-tuning. They came in around 1.7% CER on IAM and 1.6–1.7% on RIMES. Keep that result in proportion: it is one paper, on clean modern handwriting, and the same models collapsed on the historical sets in the same study.

A trained model on a known hand does very well. Where a research team has invested in ground truth for a specific hand, the numbers are research-grade. A TrOCR study on sixteenth-century Latin-language manuscripts (the Gwalther corpus, 142 pages) reached 1.86% CER for a single model and 1.60% with an augmentation ensemble.

Those results carry over: the same architectures, trained on the same alphabet, handle vernacular hands written in it.

Modern joined-up writing and historical hands are different problems. Current models largely handle a letter from the 1960s well, but secretary hand, court hand and German Kurrent have unfamiliar letterforms, dense abbreviation and far less training data, which is why we look at whether AI can read cursive in its own guide.

Historical print is often handled better than people expect, though not by the tools built for invoices. Large-scale newspaper digitisation under IMPACT and Europeana Newspapers reported character accuracy in the 85–94% band. That is usable for search, but thin for scholarship. A model trained on historical print, like ATR-1, reads the long s, ligatures and archaic spelling as what they are.

So, on the right material, with the right tool, machine transcription is now a serious research instrument. The failures are equally specific.

Where current models fail

Failure 1: the corpus cliff

The most misleading thing in this field is a bare percentage. The 2025 Journal of Documentation study tested models then current, and the results show how far one model can swing:

Modern versus historical. Claude 3.5 Sonnet scored 1.75% CER on IAM (modern English) and 41.19% CER on the ICDAR2017 set (multilingual historical documents). Same model, same prompts, a roughly twenty-four-fold spread.
GPT-4o came in around 61% CER on the same historical material
Other historical sets. On nineteenth-century Italian handwriting (the Leopardi corpus) Claude landed at 26.43% CER, and on the LAM historical dataset at 20.55%.

Newer general models have done much better on English historical handwriting since. The lesson about generalisation still holds. A model that has seen a great deal of modern cursive has not thereby seen court hand, Kurrent, or a notarial cursive from 1690. That is the case for a model trained on historical documents in the first place. LLM vision capability also moves fast enough that every one of these figures is a dated snapshot of a specific model version.

Failure 2: fluent rewriting rather than visible error

This is the failure mode that costs scholarship the most, because it survives proofreading.

A purpose-built recogniser that misreads produces a wrong character or a wrong word. The result looks visibly odd, and you can check it against the image. A general vision-language model under pressure tends instead to produce a better reading than the page supports:

Lee et al. A preprint from Lee et al. puts it bluntly: these models "do not always act as faithful transcribers". When the text is imperfect they rewrite it into a more plausible form. The rewriting is gated by how similar the model's internal representation of the observed word is to a familiar one.
Vesalainen et al. Vesalainen et al.'s comparison of TrOCR and a vision-language model on eighteenth-century English print shows the same behaviour concretely. The VLM achieved a lower CER partly because it normalised historical orthography: "Antient" silently became "Ancient".
Today's general models. The newest general models, Gemini included, still tend to do this. Their remaining errors are often a more probable spelling in place of the one on the page.

Consider what that does to a corpus. Spelling variation is evidence: of date, of region, of scribe, of literacy. A tool that improves it away has not transcribed your document. It has edited it, without telling you.

This is the failure ATR-1 is built against. It is trained to transcribe what is on the page rather than normalise it, so it does not over-correct or silently modernise: an archaic spelling stays as written. When it does misread, the error is the recoverable kind, a wrong letter or word you can check against the image beside the text, not a fluent improvement that reads correctly.

We treat this at length in Fluent but Wrong: understanding LLM transcription errors. To see it happen on a real page, our tests of Gemini on manuscript hands and of ChatGPT on historical handwriting show the pattern in detail.

Failure 3: general OCR is not attempting this task

Sending a manuscript to a print-OCR engine is not a marginal choice. It is a category error. Tesseract was built for printed text. One practitioner benchmark puts it at 38.75% average accuracy on handwriting, and independent practitioner reviews report a similar 20–40% band. These are tertiary sources, so treat the figures as indicative rather than definitive. Cloud OCR services exceed 99% accuracy on clean printed text and degrade sharply on cursive.

Historical print breaks these engines too, for structural reasons:

the long s is read as f
ligatures are split or dropped
macrons and brevigraphs stand for absent letters, and a glyph classifier has no representation for them
uneven inking and show-through are read as character evidence
post-processing "corrects" archaic spelling

Even print-specialist historical pipelines are the wrong instrument for manuscript. OCR4all, for instance, is a print pipeline for early books and Fraktur, not a handwriting tool.

Failure 4: hands and languages nobody has benchmarked

Published figures cluster around a handful of well-studied corpora. Several kinds of material have no widely cited error rates at all:

court hand and chancery hand
Spanish notarial cursive
Italian mercantile hands
much of Scandinavian and Central European chancery material

Two clarifications are worth making here, since they cause persistent confusion.

Script is not language. Latin script is the alphabet used to write English, French, German, Dutch, Spanish, Italian and Latin alike, and a model's real constraint is the writing system, not the language on the page. Non-Latin scripts (Greek, Cyrillic, Hebrew, Arabic, Indic and East Asian systems) are a different problem requiring different models.
Transcription is not translation. They are separate operations, and a tool that does one does not automatically do the other.

Our guide to which languages and scripts HTR can read explains why marketing language-counts mislead.

Failure 5: the photograph

Every model is downstream of your image. A NIST study on machine print found that degrading image quality lifted character error rates from 1% to as high as 74%, and that was on the easy task. Blur, raking light, keystone distortion and heavy JPEG compression cost you more accuracy than any model choice will win back. That is why how you photograph documents in the reading room is part of the transcription question, not a preliminary to it. If the problem is the ink rather than the camera, see Can AI read faded documents?

How to read an accuracy claim before you trust it

Character error rate is the edit distance (insertions, deletions, substitutions) between output and reference, divided by the length of the reference. Word error rate is the same calculation at word level, and it is always higher, because one wrong character condemns the whole word.

What counts as good depends entirely on what you are doing with the text. For reading, quoting or editing, you want a character error rate of about 2% or below. At 10%, one character in ten is wrong, which means an error in roughly every second word: enough to find a keyword, not enough to read from.

Published guidance is less consistent than you might expect. Transkribus's CER explainer reserves "publication-ready" for under 2% and marks 5–10% as needing review, while its help centre describes results with a CER of 10% or below as "very efficient for automated transcription". Check which bar a claim is being held to.

So when you meet a number, ask four things:

which corpus
which hand and period
which model version
against whose ground truth

A figure without those is decoration. And a low CER can hide the costliest error: a model that silently modernises spelling can score well against a modernised reference while changing your evidence. Our piece on what AI transcription accuracy actually measures goes further into why confidence scores and headline percentages mislead.

Test it on your own document

Every figure on this page describes someone else's documents. The only test that settles the question for you is a small one on your own material, and it takes under an hour:

Choose three to five images that represent the series. Include the hardest one: the worst hand, the faintest ink, the page with tables or marginalia.
Transcribe a short passage yourself. A few lines from each image is enough to judge against.
Run ATR-1, then a second model. In Leo, ATR-1 is the default. Add a general model (Gemini, say) on the same images for comparison. Each run lands in its own tab, so you can read them side by side.
Check the fields that carry the risk. Look at names, dates, sums and spellings first. A fluent sentence with a wrong surname is worse than a clumsy one with the right surname.
Decide by error type, not by impression. Visible misreadings you can catch are cheaper than quiet normalisations you can't.

Our guide on how to test AI transcription on your own documents sets out the full method, including how to keep the comparison fair. Leo's free plan gives you 10 credits a month, which is ten images on ATR-1, or more on the cheaper general models. That is enough to run this test before you commit to anything.

The training question: who pays the ground-truth cost

Specialist HTR platforms achieve their best numbers by fine-tuning on your material, and that is also their price of entry. Transkribus documentation asks for at least 25 transcribed pages before training, and roughly 5,000–15,000 words in practice. eScriptorium requires you to supply ground truth before Kraken can learn your hand.

For a funded project with a homogeneous series, that investment pays back. For a researcher with eighty pages of one clerk's hand and forty of another's, it never does: you would transcribe the corpus by hand before the model was ready. The alternative that large heritage backlogs have historically used is volunteers:

Zooniverse finds two to five transcribers per image sufficient for a generally accurate result
The Library of Congress authenticates through second-volunteer review

That approach is accurate, but limited by how many volunteers are available. The consensus best practice is hybrid: a machine first pass, then human review and correction.

Where a ready-made specialist model fits

This is the gap Leo was built for: a specialist recogniser for Latin-script material that arrives already trained. There is no per-corpus ground-truth step, and no per-office or per-collection model to maintain. ATR-1 reads any language written in the Latin alphabet, in manuscript hands and in print, historical or modern. That includes English wills, French notarial acts, German parish books, Dutch notarial registers and Latin-language texts.

Three things set ATR-1 apart from a general model on this material:

It reads the image closely. The task needs high-resolution images, which general models downsample. ATR-1 weighs the high-resolution visual evidence against context, the way a reader uses the rest of a line to settle an unclear letter.
It keeps the page's own text. Source integrity is the design commitment that matters most for the failure modes above. Archaic spelling survives, strikethroughs and marginal additions are recorded, and expansions are marked rather than smoothed. Nothing is silently regularised into modern prose.
It checks its own output. Any text-generating model can fall into repetition loops or produce text the image does not support. ATR-1 checks its output for failure patterns, hides and retries suspect results, and refunds the credit if it cannot succeed. That is a mechanism, not a promise of perfection. The errors that get through are the recoverable kind: a wrong character you can check against the image beside the text.

Analysis (correction, modernisation, translation, summary) happens as a separate Transformation, written to its own tab, which leaves the base transcription intact. And as users correct transcriptions in Leo, those corrections feed back into training, so ATR-1 keeps improving over successive releases.

Leo is also a model library. Alongside ATR-1, the default, are models from OpenAI, Google Gemini and Anthropic Claude (34 at release, in October 2026), all chosen from one menu in the transcription window. Use a small one for cheap clean print, or run a general model beside ATR-1 on a hard page as a second reading, each in its own tab.

If you are choosing between tool classes rather than reading about them, our comparison of handwriting transcription software is the fuller treatment.

What this changes about how you work

Machine transcription has moved the bottleneck, not removed it. First-pass reading of a long series is no longer the expensive part. Verification is, and verification is a paleographic skill, not a software feature. You need to know three things:

which letterforms are systematically confusable in the hand in front of you
which abbreviation marks a model might resolve without saying so
which fields in your document type carry the risk: names, dates, sums, boundary calls

That means the discipline holds:

Keep the base transcription separate from any interpretation of it
Check names and numbers against the image every time, whatever the CER
Learn enough of the hand you are working in to catch a reading that is too good to be true

Our guides to reading old handwriting and to whether historians still need paleography are a place to start. If your documents are Tudor or Stuart, Leo Academy's free secretary hand course teaches the hand in short lessons from real scribes' writing. The machine now gives you a draft of the page in seconds. Deciding what the page actually says is still, as it has always been, your job.

Frequently Asked Questions

Can AI read old handwriting?

Yes, and increasingly well, though which model you use matters. Specialised models trained on historical documents, such as Leo's ATR-1, read the image at high resolution and are built to transcribe what is on the page rather than modernise it. General AI models have improved quickly too, but they tend to quietly normalise old spelling to a more probable modern form, an error that survives proofreading because it reads correctly. Results vary widely with the hand, period and language: a 2025 study found the same general model could score under 2% character error rate on modern handwriting and over 40% on a historical set. The useful question is which errors a model makes on your documents, and the reliable way to find out is to test it on a few of your own images.

Which AI model is best for reading old handwriting?

It depends on the document, and the answer changes as models are released. A model specialised for historical documents is the place to start: Leo's own model, ATR-1, is tuned for historical handwriting and print, and trained to transcribe what is on the page rather than normalise it. General models such as Gemini, ChatGPT and Claude read many hands well but tend to modernise spelling, so they are most useful as a second reading. In Leo you can run ATR-1 and a general model on the same image, each in its own tab, and compare where they disagree.

How can I test whether AI can read my documents?

Pick three to five images that represent your series, including the hardest one, and transcribe a few lines of each yourself. Run ATR-1 and a second model on the same images and compare their output with your reading. Check names, dates, sums and spellings first, because those are where quiet errors do the most damage. In Leo, each model's run lands in its own tab beside the image, and the free plan's 10 credits a month cover ten images on ATR-1.

Can OCR software read old handwriting?

Not reliably: OCR was engineered for machine-printed text and matches isolated glyphs against known typefaces, so it degrades sharply on old handwriting, where every writer forms letters differently and the boundaries between them are often invisible. Sending a manuscript to a print-OCR engine is a category error; a handwriting model that reads the whole line in context is the right instrument.

What is a good character error rate for historical transcription?

It depends entirely on what you are doing with the text. For reading, quoting or editing, aim for about 2% CER or below. At 10%, roughly every second word contains an error: enough for keyword search, not for reading. Published guidance varies: Transkribus's help centre calls 10% or below "very efficient for automated transcription", while its own CER explainer reserves "publication-ready" for under 2%. Character error rate is the edit distance (insertions, deletions, substitutions) between output and reference, divided by reference length. Word error rate is the same calculation at word level and is always higher, since one wrong character condemns the whole word. Always ask which corpus, hand, period, model version and ground truth a figure describes.

Why do AI models change the spelling in old documents?

Because general vision-language models are not obliged to prefer what is on the page over what the sentence probably said. Under pressure they produce a better reading than the page supports, rewriting imperfect text into a more plausible form. One comparison on eighteenth-century English print found a vision-language model achieved a lower error rate partly because it normalised historical spelling: "Antient" silently became "Ancient". The newest general models still tend to make this kind of quiet normalisation. Spelling variation is evidence of date, region, scribe and literacy, and a tool that improves it away has edited your document rather than transcribed it. Leo's ATR-1 is trained the other way: to transcribe what is on the page, without over-correcting or silently modernising it.