AI labs need to start funding historical research

Frontier AI models can now crack historical ciphers and trace knowledge across languages, but turning scattershot wins into systematic breakthroughs requires AI labs to fund digitization, grant compute access, and co-define verifiable “millennium problems” with historians.

Key Points

• Frontier models (GPT-6 Sol, Opus 5.5) now solve previously intractable historical problems: decrypting WWI/WWII ciphers, identifying Newton’s alchemical sources via anagram analysis, and partially deciphering 16th-century Spanish letters in Charles V’s secret code.
• The breakthroughs rely on models’ “spiky” strengths — multilingual reasoning, math/code writing, and autonomous search across massive digitized corpora — but only when problems are tractable: digitized data, clear verification, and expert-defined questions.
• Three case studies show real but niche gains: (1) GPT-6 tracing Darwin’s informant networks via the Darwin Correspondence Project; (2) Opus 5.5 linking Newton and Hartlib through distinct anagrams for Hungarian vitriol; (3) Opus 5.5 independently decrypting letters already solved in 1916, validating its method.
• Author argues AI labs and funders should pursue three interventions: (1) mass digitization of undigitized manuscripts; (2) free API/compute grants for historians; (3) historian-led identification of “millennium problems” (e.g., Voynich manuscript, Linear A, Dee’s Liber Loagaeth) that are verifiable and tractable.
• [AI Synthesis] The piece frames AI not as replacement for historians but as a force multiplier for provenance, influence tracing, and cross-lingual quotation detection — areas where human bandwidth is the bottleneck.