How AI Is Reading Old Archives, and What It Finds

Millions of handwritten pages from the 1600s and 1700s have been turned into searchable text, and AI can now read them faster than any person could. This month a software engineer, Jesse Waites, pointed an AI pipeline at the archives of the Dutch East India Company and reported a forgotten meteorite fall, three live rhinoceros shipments and three volcanic eruptions missing from the standard list. Here is how it worked, and which parts of the method carry over to other work.

Where it started

Waites was inspired by historian Benjamin Breen, who used AI to search transcriptions produced by the GLOBALISE project and found a 1615 ship's journal that appears to be a new eyewitness record of the dodo. Waites set out to widen the approach: many questions at once, and a chain of models to filter the results. He worked with the company's archive (about 4.35 million pages), Dutch and American newspapers, and some ship logbooks.

How the pipeline worked

• Search by meaning, not spelling. Old handwriting and print come out of text recognition full of errors, and 17th-century Dutch spelled "rhinoceros" in about fifteen ways. Waites converted passages into numerical representations of their meaning, so a search finds text that is about the topic even when the words differ. One overnight run on a home graphics card covered roughly 5.7 million passages.
• Let a cheap model do the first read. A single search can return tens of thousands of hits. The first screening went to Jev, a small, fast decision model that answers narrow yes-or-no questions. By his account it read 59,000 elephant mentions for about three dollars. We explain what that kind of model is in What Is Jev? The AI Model That Refuses to Write Text.
• Escalate only what survives. A larger model read the flagged passages closely, translated them and pulled out dates and places. An agent then opened the scan of each original page to compare it with the transcription.
• Test the search first. Before looking for anything new, the pipeline had to rediscover events already known, such as the dodo record and the 1783 and 1815 eruptions. Waites compares it to burying your own watch in the garden to check a metal detector works.

What it reported finding

• A meteorite fall, 1812. A letter describing a heavy, iron-rich stone falling during thunder near Pandharpur in what is now Maharashtra was reprinted in an English-language newspaper in Batavia that December. Waites found it in none of six catalogues he checked. If confirmed, he says, it would be the earliest recorded fall in Maharashtra, 26 years before the current earliest.
• Three Javan rhinoceroses, 1738 to 1740. They were caught near Batavia as gifts for the King of Kandy in Sri Lanka. One reached Colombo and died there. The other two died at sea on a single voyage, and the crew gave sworn statements about it.
• Three eruptions. The pipeline surfaced 769 eruption reports. After cross-checking against the Smithsonian's volcano list, three remained: Gamkonora (1722), Ciremai (1712) and Slamet (1780).

What carries over

Four ideas are useful far beyond history: search by meaning when the text is messy, use a cheap model to screen and an expensive one to read, prove your search can find something you know is there before trusting an empty result, and make every claim end at an original page someone can check.

The same logic applies to any pile of material that is too large to read, such as family letters, old photographs with captions, or voice recordings. Recordings become searchable once they are text, and OffgridScribe turns them into text on your iPhone.

What it doesn't show yet

• These are candidates. Waites says they are checked against the original pages and the catalogues he could access, but not yet reviewed by the specialists who maintain them. He had contacted curators and was waiting for replies.
• The AI was often wrong on its own. In one case it blamed an ash fall on a volcano about 1,500 kilometres away, so each candidate had to be placed using the letter's own geography.
• Some searches found nothing. He could not locate the large unexplained eruption of 1808 or 1809, and thinks the reports simply are not in these sources.
• It was a guided process. He describes it as AI-assisted research, with a person deciding where to look and when to stop.

We are working from his published write-up and have not checked the archives ourselves. Treat the findings as claims awaiting confirmation.

What to watch

• Whether the curators and volcanologists confirm the meteorite and the eruptions.
• Whether Waites's open-source toolkit, Antiquity, lets other people repeat the method on different archives.
• How many more archives get transcribed and opened up, since that is the real limit.

Sources

• Jesse Waites, "I Pointed AI at 400 Years of Historical Archives" (8 October 2026)
• Benjamin Breen, post on a new eyewitness account of the dodo (1 October 2026)
• GLOBALISE project

For disclosure: we make iPhone apps, OffgridStem, OffgridScribe, OffgridVox and OffgridCam.

More from Offgrid Studio

What Is Jev? The AI Model That Refuses to Write Text
Why Cloud AI Bills Are So Hard to Budget