AI labs urged to back historical research

New tests with GPT-6 Sol, GPT-6 Astra and Opus 5.5 suggest frontier models can help historians pursue tightly defined problems when the sources are digitized, accessible and open to verification. Promising areas include cryptography, tracing texts across translations and adaptations, and connecting findings scattered across niche fields. A key example involves GPT-6 Astra helping crack a 1941 Enigma message by finding relevant archival context rather than simply performing codebreaking.

Experiments ranged from John Dee’s Liber Loagaeth to Darwin’s research networks and the Hartlib papers. GPT-6 Astra found that Dee’s supposed angelic manuscript was largely nonsense syllables while identifying one meaningful reference. Opus 5.5 downloaded over 5,000 primary source files from Samuel Hartlib’s archive and identified parallels between Hartlib and Newton in coded references to Hungarian vitriol, suggesting a possible source relationship. Opus 5.5 also partially deciphered two 16th century Spanish letters linked to Emperor Charles V, later confirming that earlier human decipherments already existed.

The next step proposed is a coordinated effort among historians, archivists, AI researchers, labs and foundations. Priorities include digitizing more manuscripts, making restricted archival material freely accessible, and providing historians with free API access and compute. The likely gains are expected to be uneven, but potentially important for questions involving provenance, quotation, influence across languages and unresolved historical ciphers.