Benjamin Breen, a historian researching animal extinctions in the early modern period, used Anthropic's Opus 5.5 model to search through the GLOBALISE archive of Dutch East India Company records and surface what appears to be a previously unnoticed 1615 ship's log reference to hunting dodos on Mauritius. The passage comes from the journal of Isbrant Cornelisz van Petten, captain of the Wapen van Amsterdam, who recorded that his crew "caught many tortoises, dodos [dodeersen], and some geese and parrots" during an April 1615 stopover. Cross-referencing with specialist literature including Parrish's 2013 book The Dodo and the Solitaire, the find fills a gap in the dodo timeline for the 1611-1616 period. The methodology matters more than the individual find. Breen's workflow is straightforward: start from a specific human-generated research question rooted in domain expertise, identify a large digitized archive (GLOBALISE contains millions of VOC records), run the corpus through an embedding model for semantic search, then deploy frontier models to read and rank candidate passages for human review. Opus 5.5 spawned multiple independent agents that searched across languages and followed citation chains autonomously, including running their own embedding searches on new sources that emerged during research. A second finding from the same search corrected an 1890 French mistranslation. A 1638 Dutch account of Mauritius described "velthoenderen" (field-hens), a term experts associate with the extinct red rail. A French scholar had rendered this as "perdrix" (partridges), and the error had persisted for over a century. Opus went back to the original manuscript and flagged the discrepancy — a task that requires cross-referencing the original Dutch against a published French translation, which is exactly the kind of tedious multilingual comparison AI excels at. The most tantalizing thread is speculative but well-framed. The Mughal Emperor Jahangir owned a living dodo, painted by his court artist Ustad Mansur in what biologists still regard as the most accurate surviving depiction of the bird. Nobody knows how the dodo reached India. Opus surfaced a connection: a Portuguese Jesuit described an "ostrich" on Mauritius in 1616 (Mauritius has no ostriches; experts already suspect this was a dodo), and a known chain of transmission already existed for exotic birds from Portuguese Goa to Jahangir's court — including an American turkey in 1612. The connection between the 1616 Portuguese dodo and Jahangir's bird has apparently not been previously made in the literature. Breen is honest about the limitations. The agents repeatedly got lost in minutiae, drilling into irrelevant detail without any sense of historical significance. One agent discovered an entertaining account of an English captain, Jonathan Hide, stealing ebony and "two sea cows" from the Dutch colony, with a carpenter threatening to split a Dutch official's head with an axe. Interesting, but not what was asked for. The models cannot generate research questions, judge significance, or distinguish a publishable find from trivia. They are, as Breen puts it, doing "the digital equivalent of counting sheep" — tireless scanning that never gets bored. This sits alongside Carter Church's use of GPT-6 Astra to break a Napoleonic-era cipher (six hours of autonomous work) as evidence that frontier models have crossed a threshold in archival research utility. The pattern is consistent: massive digitized corpus plus specific human question plus AI reading at scale equals genuinely new finds that human researchers missed because they couldn't read fast enough or widely enough. The epistemological shift is real — not because the AI is smarter than historians, but because it can spawn dozens of copies of itself to read through sources in multiple languages simultaneously. The generative potential here is substantial but structurally bounded. AI archival research amplifies existing expert knowledge rather than replacing it. The finds are real, the methodology is reproducible, and the limiting factor is now the quality and availability of digitized archives rather than human reading speed. The GLOBALISE project and similar digitization efforts become dramatically more valuable when frontier models can search them semantically. What remains firmly human is the research question, the significance judgment, and the connective reasoning that turns a ship's log entry into a hypothesis about Jahangir's dodo.
