'I didn't believe it at first': OpenAI model transcribes, decodes Napoleon's cipher

Napoleon wanted the message to be encrypted and carried by an intelligence officer. This would make it harder for enemies to read if they captured it.

An OpenAI model reportedly solved a 217-year-old code used by Napoleon, working from just one scanned image. It recreated a battlefield report from March 1809 and corrected the date when the letter was published.

Here's what to know

The document was a letter Napoleon Bonaparte sent to General Auguste de Marmont. According to Popular Mechanics, Carter Church, an AI engineer at SentinelOne, asked OpenAI's GPT-6 Astra to work from a single, low-quality scan. The blurry image came from a French Army journal published in 1969. The model read and decoded 24 lines of symbols, numbers, and letters.

Church said Astra recorded about 1,300 hand-drawn marks and grouped them into 155 different signs. It then solved a type of code called a homophonic cipher. In this kind of code, several different symbols can represent the same common letter. Church said it took about six hours to check the results.

Church said the model connected the letter to part of a code key that had not been linked to it before. It then used a computer method called simulated annealing to finish decoding the message. The model also checked other letters from around the same period to correct the publication date. The editor of Cryptiana's list of unsolved historical codes has now marked the entry as solved.

"Reading the symbols and breaking the code are usually separate skills, often handled by different experts," he wrote. "Here, they were done as one process, and the original image was checked throughout: each suggested meaning was compared with every place that symbol appeared in the image."

More background

The surviving message is dated March 16, 1809. From Paris, Napoleon told his stepson Eugène, the Viceroy of Italy, to deliver a coded letter to Marmont. Marmont's troops in Dalmatia had been cut off from the rest of the French armies.

Napoleon wanted the message to be encrypted and carried by an intelligence officer so it would be harder to read if intercepted. Church wrote that the decoded message was basically a military report describing the positions of French and allied armies. It was sent before Austria invaded.

Church also warned against making the code sound harder to crack than it was. He said human codebreakers could have solved it "easily" if they had spent enough time on it. He told Live Science, "I didn't believe it at first. Everyone has seen a model make an unbelievable claim. But this result is different because it can be tested step by step. There is almost no chance that a wrong key would produce a message that makes sense and matches what we know about history."

"Every field has a list of unsolved, specialized problems that experts don't have enough time to work on," he wrote. "The lesson here is that more of these problems will be solved by people from outside the field who are simply curious."

AI systems that can process an old document from 1809 may also help utility companies predict energy demand and make better use of clean energy on the power grid. However, the growing use of computers behind these tools can require huge amounts of electricity and water.

As these systems become more powerful, they also raise concerns about misuse and cybersecurity. There are also worries that rising demand for data centers could increase household energy bills.

Where can I learn more?

These stories explore other ways AI is being used, along with some of the costs and challenges that come with it.

• Researchers have developed a new approach to artificial intelligence that could replace methods that use large amounts of energy.

• Scientists used AI to map ocean currents in much greater detail.

• At a nuclear cleanup site, scientists used AI to speed up research into removing radioactive iodine.

• Westinghouse and Google described a plan to use AI in nuclear power plant construction, which they called the fastest way forward.

• Researchers are warning about serious environmental problems as generative AI adds to electronic waste.