The GPT-6 Astra neural network has helped decipher a message from French Emperor Napoleon Bonaparte whose contents had remained unknown for about 217 years. The letter was sent in 1809 to General Auguste de Marmont through Napoleon’s stepson, Eugène.
The message contained information about the movements and numbers of French and Austrian troops ahead of a military confrontation between the two states. The decryption was carried out by Carter Church, an artificial intelligence engineer at SentinelOne.
The neural network took about six hours to process roughly 1,300 characters and match them with elements of a cipher known to cryptographers. During the analysis, it used a simulated annealing method, in which the system tested different word variants and compared the results with other surviving texts written by Marmont.
The deciphered message instructed the general to prepare for a march and not to fear “a few detachments or a heap of rabble.” The letter also stated that Austria “will immediately attack from all sides.”
The interesting aspect of this story is not only that a message inaccessible for more than two centuries has become readable. What is also noteworthy is the way the system approached the old cipher: it did not simply offer a single ready-made version, but tested different combinations of words and compared them with other surviving texts by Marmont.
This demonstrates how artificial intelligence tools can be used to study historical materials, particularly in cases where handwritten or encoded texts have long been difficult to fully understand. At the same time, the result of this work is connected not only to technology but also to the documents whose surviving copies made it possible to verify different versions through comparison.
