Court and notarial records are some of the richest sources a historian or family researcher can find. A single notary's register can hold land sales, marriage contracts, wills, loans and inventories listing everything a family owned, right down to the cooking pots. Court records add the arguments, the witness statements and the verdicts, often in the actual words people spoke.
The problem is getting at them. These records were written fast, by professionals, in dense legal language crammed onto every inch of the page. Many run to hundreds of volumes, and reading them by hand can take years.
AI transcription has changed that. Software trained on historical handwriting can now produce a working draft of a page in seconds, which turns an impossible project into a manageable one. This guide explains how to do it properly: preparing your images, running the transcription, checking the results against the legal formulas, and turning the text into data you can actually use.
What Makes Court and Notarial Records Different
If you've transcribed letters or parish registers before, legal records will feel like a step up. A few features set them apart, and each one affects how you should approach AI transcription.
They were written by professionals at speed. Notaries, clerks and court scribes wrote all day, every day. Their hands are often fluent and fast, with letters run together and words shortened in ways that only another clerk would expect.
They follow fixed formulas. Almost every deed, will or court entry was built from standard phrases repeated thousands of times. That can make the text feel monotonous, but it's also a gift, because you know what should come next. More on using that below.
They're packed with names, numbers and places. Parties, witnesses, sums of money, property boundaries and dates are exactly the details researchers want. They're also exactly where transcription errors cause the most damage.
They mix languages and layers. A record might switch between Latin and the local language, carry summaries in the margin, and gain notes added years later by a different hand. Signatures and notarial marks sit alongside the main text.
There's a lot of it. Notarial archives and court series often run continuously for centuries. That volume is the main reason AI makes such a difference here: it lets you search and sort material that no one person could ever read in full.
Know the Records You're Working With
Before you upload a single page, it helps to know what kind of record you have. The type tells you how the page is laid out, which formulas to expect and which details matter most.
Notarial Registers and Protocols
In much of continental Europe, Latin America, Quebec and Louisiana, notaries recorded private agreements in bound registers, often called protocols or minute books. A typical volume holds sales of land and houses, leases, loans and debt acknowledgements, marriage contracts and dowries, wills, powers of attorney, apprenticeship agreements and post-mortem inventories. Each deed usually opens with the date and place, names the notary and the parties, sets out the terms and closes with witnesses and signatures.
Many notaries also kept an index or calendar for each year, listing the parties and the type of deed. These indexes are short, regular and easy for AI to read, so they make an excellent place to start.
Court Records
Court records come in more varied forms. You may meet pleadings, where each side sets out its case in writing, and depositions, where witnesses answered set questions under oath. Criminal courts produced indictments, recognisances and gaol calendars, while minute books and order books recorded what each sitting decided. Probate courts generated wills, inventories and administration bonds.
Court material is often messier than notarial work. Several clerks may have written in the same volume, entries can be crossed out or squeezed in, and later endorsements appear on the backs of documents. Expect to spend more time checking these pages.
Before You Upload: Access, Permissions and Privacy
Legal records often involve living people, family secrets and property that's still owned today. That's why many countries close them for long periods. In France, for example, notarial minutes generally become open to the public after 75 years, or 100 years if the deed concerns a minor. In Spain, notarial protocols pass to historical archives and open to everyone once they reach 100 years old. Court records have their own rules, which vary by country, court and type of case.
A few checks will keep you on the right side of the archive and the law:
Read the archive's terms. Some archives allow free reuse of their images; others limit publication or commercial use. Check whether uploading images to an online service is allowed, especially if you photographed them in a reading room.
Be careful with recent material. Records from the last century may name people who are still alive or their close relatives. If you plan to publish transcripts, think about what personal details really need to be shared.
Keep your source details from the start. Note the archive, the collection or fonds, the notary or court, the volume and the folio for every image. Rebuilding those references later is miserable work, and without them your transcript can't be checked or cited.
Capture Images the AI Can Actually Read
The quality of your transcript can never be better than the quality of your image. Legal volumes bring a few particular challenges.
Deal with the gutter. Notarial registers and court books are usually thick, tightly bound volumes. Text near the spine curves away from the camera and falls into shadow. Use a book cradle or support the volume so it opens without strain, and photograph each page separately, square to the camera, rather than shooting the whole spread at once.
Keep the margins in. Notaries and clerks often wrote the deed type and the parties' names in the margin. Those notes are some of the most useful words on the page, so never crop them off.
Capture the backs of loose documents. Court papers were folded and endorsed on the outside with the case name, the date filed or the outcome. Photograph both sides and keep them together.
Aim for at least 400 dpi and even light. Avoid flash, which bounces off glossy parchment and washes out faded ink. Shoot in colour and keep the original files untouched; you can always make enhanced copies later.
If you're using existing scans, download the highest resolution the archive offers. Small online previews often look fine on screen but give poor results once the software tries to separate one letter from the next.
A Step-by-Step AI Transcription Workflow
This is the routine that works best for most legal record projects, whether you have ten pages or ten thousand.
Step 1: Run a Small Test Batch
Pick five to ten representative pages from different parts of the volume, including at least one difficult page. Run them through your transcription tool and check the results carefully. This tells you how well the software copes with this particular notary or clerk before you commit to the whole series.
Step 2: Check the Layout Detection
Before you look at a single word, look at how the software divided the page. Did it separate the margin notes from the main text? Did it keep each line in the right order? Did it treat signatures as text or ignore them? On legal records, most nonsense output comes from layout errors rather than misread letters, so fix these first.
Step 3: Process the Full Volume
Once the test pages look good, upload the rest in batches. Keep the files in page order and name them consistently, for example by volume and folio number, so every line of text can be traced back to its image.
Step 4: Correct the High-Value Details First
You don't need to perfect every word on every page straight away. Start with the details that drive research: names of parties and witnesses, dates, places, sums of money and property descriptions. Check each one against the image.
Step 5: Read Against the Formula
Compare each deed or entry with the standard pattern for that record type. Wherever the transcript breaks the formula, look closely. Either the AI made a mistake, or the clerk did something unusual, and both are worth knowing about.
Step 6: Mark What You've Checked
Keep a simple note of which pages are raw AI output, which have been spot-checked and which are fully verified. When you or anyone else comes back to the material months later, you'll know exactly how far to trust each part.
Use the Legal Formulas as Your Checking Tool
The repetitive language of legal records is your best friend when checking AI output. Once you know the standard openings and closings, you can spot a misreading almost instantly, and you can often fill a damaged gap with confidence.
Common Openings
English bonds and deeds often begin "Know all men by these presents", which in Latin documents appears as "Noverint universi per praesentes". Wills commonly open "In the name of God, Amen", followed by a statement that the testator is weak in body but of sound mind and memory. In Spanish notarial deeds you'll see "Sepan cuantos esta carta vieren", roughly "Let all who see this document know". French deeds typically start "Par-devant les notaires soussignés", meaning "before the undersigned notaries", followed by "fut présent" or "furent présents" and the names of the parties.
Common Closings
English documents usually end with a phrase like "In witness whereof I have hereunto set my hand and seal" and a line noting that the deed was "signed, sealed and delivered" in front of witnesses. Spanish notaries confirm the act with "doy fe", meaning "I attest", and sign beneath "ante mí". French deeds often close with "fait et passé" plus the place and date, then "dont acte".
Court Formulas
Indictments in English criminal courts often begin by stating that the jurors, on their oath, present that the accused committed the offence. The grand jury's decision was then written on the back: "billa vera" (a true bill) meant the case went to trial, while "ignoramus" meant it was thrown out. In depositions, each answer usually starts with a phrase like "to the first interrogatory this deponent saith", which makes it easy to match answers to questions.
When the AI produces something that doesn't fit these patterns, slow down. It's often a sign of a misread word, a skipped line or a jumbled layout.
Abbreviations, Money and Dates to Double-Check
These three areas cause the most AI errors in legal records, and they're also the details that matter most to your research.
Legal Abbreviations
Clerks shortened the words they wrote most often. In English records, "sd" is "said", "afsd" is "aforesaid", "decd" is "deceased", "exor" and "exrx" are "executor" and "executrix", and "admor" or "admx" stand for "administrator" and "administratrix". You'll also see "pd" for "paid" and "recd" for "received". Latin and other European records have their own sets, often marked by a raised letter or a line over the word. AI tools may expand these, leave them as written or misread them as different words, so decide early how you want them recorded and check them consistently.
Sums of Money
Old currencies rarely match modern decimal thinking. English sums are written in pounds, shillings and pence (£ s d), sometimes with "li." for pounds. French records use livres, sols and deniers, while Spanish documents may mention reales, maravedís, ducados or pesos. Amounts are often written out in words and then repeated in figures, so compare the two. If they disagree in your transcript, one of them has almost certainly been misread.
Dates
Legal dating can be tricky. English records may use regnal years, such as "the tenth year of the reign of King George the Third", or date the year from 25 March rather than 1 January before 1752. French notarial deeds from the mid-1790s to 1805 use the Revolutionary calendar, so you'll see dates like "25 ventôse an XI". Many deeds also name a feast day rather than a numbered date. Always record the date exactly as written and add your modern conversion separately.
From Transcript to Data: Getting More Out of Your Records
A full transcript is valuable, but the real payoff comes when you turn hundreds of deeds or cases into something you can search, sort and compare.
Because legal records follow such regular patterns, they're ideal for structured extraction. For each deed or court entry, aim to capture a consistent set of fields: the date as written and its modern equivalent, the type of record, the notary or court, every party with their role, the witnesses, places and properties mentioned, the sums of money involved, and the volume and folio reference.
Once that information sits in a spreadsheet or database, patterns start to appear. You can trace one family's land across generations, see who lent money to whom in a village, map where a notary's clients came from, or follow a single dispute through several courts. That kind of analysis simply isn't practical when the information is locked in thousands of handwritten pages.
Keep the link between each row of data and its source image. If a name or figure ever looks odd, you should be able to open the original page in a couple of clicks and check it.
Common AI Mistakes and How to Catch Them
Even the best transcription tools make predictable errors on legal documents. Knowing what to look for makes checking much faster.
"Correcting" historical spelling. Some tools quietly modernise old spellings or swap an unusual surname for a common one. Compare names letter by letter with the image, especially surnames that appear only once.
Mixing up margin notes and main text. If the software reads straight across the page, a marginal summary can end up spliced into the middle of a sentence. Garbled lines at regular intervals are a strong clue.
Misreading numbers. Old forms of 1, 7, 4 and 9 can look alike, and Roman numerals cause trouble too. Check every figure against the amount written out in words where there is one.
Dropping short lines and insertions. Words added above the line, short closing lines and signatures are easy for software to skip. Scan the image for anything that doesn't appear in the text.
Inventing plausible text. When the image is damaged, some AI systems fill the gap with words that sound right but aren't on the page. Treat smooth, confident text over a stained or torn area with suspicion, and mark any genuine gaps as [illegible] or [damaged].
A simple habit helps with all of these: never accept a key name, date or sum without looking at it in the original image yourself.
Frequently Asked Questions
Can AI read old court records accurately?
Yes, modern handwriting recognition trained on historical scripts can produce strong first drafts of court records, often accurate enough to search and sort large collections. Accuracy still depends on image quality, the clerk's hand and the condition of the page, so names, dates and sums should always be checked against the original.
What is a notarial record?
A notarial record is a legal document drawn up and kept by a notary, such as a land sale, will, marriage contract, loan or inventory. In many countries, notaries bound these deeds into yearly registers, which now survive in regional and national archives.
Can AI transcribe Latin legal documents?
Many AI tools can transcribe Latin, but legal Latin is heavily abbreviated. Expect to check expansions carefully and use the standard formulas of each document type to confirm the reading.
Do I still need to check an AI transcription?
Always. Treat AI output as a fast, searchable draft. For research or publication, verify the key details yourself and note which pages you've checked.
Why Scripily.com Is the Best Choice for Court and Notarial Records
For legal records, Scripily.com is the tool we'd recommend. It was built specifically for historical documents, so it's designed for the things that make court and notarial records hard: fast professional hands, centuries-old scripts, faded ink and archaic letter forms. Scripily reports 98.1% accuracy on historical material, and it can restore damaged pages as well as transcribe them. You can upload scans, photographs or PDFs of whole registers, read records in multiple languages, and use its field and table recognition to pull out exactly the details that matter in legal documents, such as parties, dates, places and sums. Results can be exported straight into your own spreadsheets or research notes, teams can work together in a shared workspace, and larger digitisation projects can connect through its API. Whether you're a genealogist tracing a family's land through a notary's protocols, a historian working through court depositions or an archivist opening up a whole series to researchers, Scripily turns years of reading into a fast, searchable first draft, leaving you free to focus on what the records actually reveal.
.jpg%3Falt%3Dmedia%26token%3D7a83f060-36aa-451a-a018-44ce0ee37650&w=3840&q=75)
.png%3Falt%3Dmedia%26token%3D18189173-4105-4e52-a0bc-16aaac745c42&w=3840&q=75)
