Another day, another historical cipher broken by a frontier model. Yesterday, the security researcher Carter Church announced that he had used GPT-6 Astra (working on the problem for six hours) to break a Napoleonic-era cipher that had previously resisted all decryption attempts.
What I notice most about these forays into historical sleuthing using AI (and my own attempts at same) is the epistemological weirdness of how current frontier models now operate when given historical research tasks.
The core steps that led to what appears to be a previously-unnoticed Dutch report of hunting dodos dating to 1615:
- Start from an actual base of specialist knowledge to define a specific research question
- Identify a large, freely-available, well-edited corpus of historical sources (here, the GLOBALISE archive of Dutch East India Company archives)
- Download the sources and run them through an embedding model to allow semantic search
- Use semantic search to surface candidate passages and ask frontier models to read them and produce a ranked list for human review
- Iterate on promising passages or keep looking with new archives and search terms
The difference from solo research is that an AI agent like Opus 5.5 can spawn dozens of copies of itself to read through sources in multiple languages—and, with an API key, run its own embedding searches as new sources emerge.
A dodo in a haystack
Few historical creatures have been more widely studied than the dodo. That large identified source base made the problem tractable: most of what Opus 5.5 found was already known, but one manuscript source from 1615—a ship's log—seems not to have been noticed before.
The journal was probably written by Isbrant Cornelisz van Petten, who captained a Dutch East India Company merchant vessel called Wapen van Amsterdam. The Wapen made landfall on Mauritius in April 1615, where the crew collected water and food. Among other things, they "caught many tortoises, dodos [dodeersen], and some geese and parrots."
Relative to secondary literature such as Parrish's 2013 The Dodo and the Solitaire, this looks like a new addition to the timeline, which previously had a gap in the 1611–16 period.
The same search also found a probable new reference to another extinct Mauritius bird, the red rail: a 1638 account describing "field-hens" (velthoenderen), missed in part because a French scholar in 1890 mistranslated the word as perdrix (partridges). Opus 5.5 went back to the original manuscript and corrected this.
A still-speculative thread connects a Portuguese Jesuit's 1616 Mauritius "ostrich" (Mauritius has no ostriches) to the living dodo painted for Mughal Emperor Jahangir by Ustad Mansur—possibly the most scientifically important individual dodo depiction that survives.
Epistemological weirdness
The main thing contemporary AI can do for historical research is the digital equivalent of counting sheep: search enormous datasets for new evidence for existing claims (or potentially disprove them). They are worse at coming up with new ideas of their own. What works best is placing them on the boundary between two disciplines, giving a source base, and telling them to work methodically toward a human-generated question.
They are also notably bad at judging historical significance. Agents repeatedly drilled into minutiae and got lost in the weeds—sometimes fruitfully (an English captain stealing ebony and "two sea cows," plus ~20 land tortoises intended for St. Helena), sometimes into archives requiring specialist knowledge to fact-check.
The bottleneck will soon become not research findings themselves, but the attention of experts in niche topics. As a guest post on Terence Tao's blog put it: "We're gonna need a lot more mathematicians." I would add: we're gonna need a lot more historians and humanists.
Original: Res Obscura