AI Tools Help Decode Centuries-Old Encrypted Manuscripts

A 408-page manuscript known as the Borg cipher, held at the Vatican Library, remained unread for more than 400 years because its 34 symbols had no known key. Machine-learning models trained on historical scripts converted the handwritten pages into machine-readable text, allowing researchers to identify thousands of medical remedies such as drinking red wine or fermenting nutmeg to treat dysentery.

Beáta Megyesi, professor of computational linguistics at Stockholm University, participated in the project. She described the work as detective work that brings researchers closer to lost historical information.

Before decryption, handwritten documents must be transcribed. Michelle Waldispühl, professor of German linguistics at the University of Oslo, used the Transkribus platform to convert a 1637 letter written by nobleman Sigismund Heusner von Wandersleben into digital form.

The letter contained both encrypted sections using numbers and plain 17th-century German script. Cecile Pierrot, a cryptologist at the French National Institute for Computer Science Research, spent six months transcribing and decoding a three-page letter from Charles V that used 120 cipher symbols.

The decrypted text showed the Holy Roman Emperor's fear of assassination by an Italian mercenary.

Estimates indicate that about 1 percent of material in archives worldwide is fully or partially encrypted. Some ciphers date to ancient Greece and Rome. Documents include diplomatic messages, medical knowledge, and personal correspondence that people wished to keep private.

Megyesi and colleagues are developing additional AI models under the Descrypt project to handle unusual symbols and scripts. The models are trained across multiple alphabets and symbolic systems to reduce the time required for manual transcription.