How AI Deciphers Ancient Texts, Lost Languages and Secret Codes

AI Is Decoding History: From Napoleon’s Secret Cipher to Lost Ancient Books

The first line was ordinary French. Below it stretched twenty-four rows of numbers, letters and curious handmade symbols. The message was meant for one of Napoleon's generals, far from the main French army, and it carried information an enemy would have loved to intercept. For more than two centuries, the code concealed almost everything after the greeting.

In September 2026, cybersecurity engineer Carter Church published a way to read it. Working from a low-resolution scan, an AI system transcribed the marks, tested possible substitutions and helped reconstruct an 1809 military briefing. The solution took roughly six hours of model execution and came with files that others could use to check the result. What emerged was not a sensational secret about Napoleon's private life. It was something historians often value more: an unusually detailed picture of armies preparing for war.

A different kind of breakthrough had arrived earlier that summer. Scientists announced they had digitally opened the surviving core of a Roman scroll charred by Mount Vesuvius in AD 79. And in another project, researchers pieced together a forgotten Babylonian hymn from cuneiform tablets scattered through collections. None of these discoveries involved the same technique. All three made historical evidence accessible in a way it had not been before.

These stories point to a larger change in historical research. AI can help recognize ink, sort fragments, propose missing letters, translate a known language or test a secret code. But the machine needs something to work with: scans, surviving signs, comparable texts and scholars who can tell a plausible answer from a defensible one.

From Napoleon’s encrypted military messages to ancient scrolls buried by volcanic ash, artificial intelligence is helping researchers recover knowledge once thought lost forever.

What does it actually mean to 'decode' the past?

A text can be unreadable for at least five reasons. It may be encrypted, with ordinary words deliberately replaced by secret symbols. The writing may be known but almost invisible because the ink has faded or the page is damaged. Pieces of an inscription may be missing. A surviving text may be in a language scholars understand, such as Latin or Akkadian, but still need translation. Or its script or language may remain unknown, with no agreed way to read it at all.

Those problems can look similar in a headline, yet they call for different solutions. Revealing ink on papyrus does not translate Greek. Guessing a missing Latin word does not prove that an ancient writer used it. Detecting patterns in Linear A does not mean we know the language that its signs recorded. Throughout this article, it is worth asking one simple question: did researchers recover something that survives in the evidence, or did they generate a hypothesis about what might once have been there?

Think of a damaged photograph. Software might reveal a faint detail that the camera genuinely captured, or it might invent a realistic-looking face to fill a gap. Both outputs could impress us, but only the first offers direct evidence of what was there. Ancient texts demand that same distinction. Readers curious about how models spot such patterns can find a plain-English primer in Next Horizon’s AI explained guide.

Napoleon's cipher: an isolated general and a missing key

In March 1809, war with Austria was approaching. General Auguste de Marmont commanded French troops in Dalmatia, cut off from the main French forces by the Adriatic and hostile territory. Napoleon needed him to know how French and allied armies were positioned. On March 16, he ordered his stepson Eugène de Beauharnais, the viceroy of Italy, to send Marmont a coded briefing by a trusted officer. The surviving message appears to answer that order. It was not a letter personally handwritten by Napoleon.

For modern codebreakers, the clues were frustratingly thin. The message survived in a plate published in a French military-history journal in 1969, later digitized at only 1,202 by 1,836 pixels. Some marks were scarcely larger than a few dozen pixels. A cryptology historian, Daniel Tant, had documented values for 33 signs, but there was no complete key and no ready-made transcription of the page.

Church used GPT-6 Astra in a multistep workflow described in his September 18, 2026 technical account. The model first had to identify symbols in the image. It then grouped inconsistent handwritten versions of the same sign, generated code to search possible substitutions and tested candidate readings against French and historical context. The striking part was not a chatbot suddenly 'knowing' a lost code: it was a system carrying out image analysis, programmatic search and historical cross-checking in one process.

There were about 1,300 encoded units and 155 distinct signs. The scheme was a homophonic substitution cipher, in which one common letter could have several alternative symbols. That frustrates the classic trick of matching the most frequent sign to the most frequent letter. Some signs also stood for entire words; Church identified 29 such word-signs. Even a human who knew French could stare at the page for hours without seeing a recognizable sentence.

The recovered text described French, Bavarian, Polish and other military formations as Europe moved toward the 1809 campaign. It also helped date the letter to late March 1809 rather than the older year attached to it in historical cipher listings. Most importantly, Church released the transcription, reconstructed key and scripts so that another researcher could test the reading sign by sign. That is what turns an impressive claim into a result open to scrutiny.

It is tempting to call this 'AI breaking Napoleon's encryption.' A more accurate description is also more interesting. A powerful modern system used old-fashioned cryptanalysis, an incomplete historical key and careful verification to solve a specific paper cipher. Nothing in the case suggests that modern secure digital encryption has become easy to break.

In 2026, AI-assisted cryptanalysis helped decode an encrypted military message from 1809, demonstrating how modern technology can uncover secrets hidden in historical correspondence.

Inside the codebreaker's workshop

Imagine finding an unknown sign repeated dozens of times in a letter. Does it mean e, the most common letter in French? Perhaps. But if the writer used five different signs for e, a single frequency table will mislead you. The page might also contain code symbols for whole words—'army' or 'general,' for example—and meaningless fillers intended to confuse an interceptor.

An effective workflow therefore makes small, checkable moves. First, software maps the marks on the scan. Next it tests which signs might represent the same letter. A search algorithm tries combinations, scoring them against patterns in the known language. The strongest candidates are checked against every occurrence of the relevant signs. Finally, historians ask whether the names, military units, places and dates fit the period. If the reading falls apart elsewhere on the page, it goes back to the drawing board.

The model may also write the code that performs the search rather than relying on one fixed cryptography program. Readers curious about that broader capability can explore Next Horizon’s guide to what AI can—and cannot—do when it writes software.

There is no universal 'decrypt' button. Church's case was unusually tractable because the language was known, a partial key existed and the military setting supplied useful clues. Take away those anchors and a superficially similar cipher can be much harder—or impossible to solve with confidence.

A Roman scroll, opened without being touched

In AD 79, Mount Vesuvius buried Herculaneum in volcanic material. A villa there held a remarkable collection of papyrus books. Many rolls were carbonized: their surfaces blackened, their layers fused and their ordinary shape lost. For centuries, the most direct way to inspect a scroll was also one of the most dangerous. Physically unrolling it could reduce an irreplaceable text to fragments.

The Vesuvius Challenge, launched in 2023, took a different approach. At a synchrotron facility, researchers scanned the inside of the blackened roll using powerful X-rays. Computers traced the curled sheets through a three-dimensional volume and flattened them into a virtual writing surface. Machine learning then helped detect the weak signal left by ink against the papyrus. Specialists examined the resulting letters and words. No one had to force the original object open.

This matters because the AI is working with measurements of a real object. Its primary job is to reveal evidence of ink, not to compose a plausible passage by an ancient philosopher. Reading and interpreting the recovered Greek still requires papyrologists—specialists who study handwritten texts on papyrus.

In February 2025, Oxford researchers reported progress on PHerc. 172, a scroll kept in the Bodleian Libraries. The Greek word diatropē—'disgust'—was among the early readings. Oxford was explicit about the division of labor: the machine-learning tools detected ink, but they could not read the Greek language. The scroll's unusual ink chemistry also made some writing clearer in X-rays than it is in other Herculaneum rolls.

The bigger milestone came on June 25, 2026. The Vesuvius Challenge announced that PHerc. 1667 had been digitally opened and the surviving rolled portion read from end to end. It yielded about 1.4 meters of writing surface and roughly twenty-two columns of Greek. That phrase 'surviving portion' matters: earlier attempts to open the original roll had destroyed outer layers. What scholars recovered is a remarkable continuous text, not every word that once existed in the complete ancient book.

The text discusses ethics, human nature and how people should live. Its language and themes point toward Stoic philosophy, and one surviving column mentions Aristocreon, a nephew and pupil of the Stoic thinker Chrysippus. That is an invaluable historical clue, but it does not establish the work's author. The same 2026 research also identified the title and author of a different roll, PHerc. 139: Philodemus, On Gods, Book 8. Two sealed objects, and two quite different kinds of recovery.

The most exciting change is methodological. Researchers can now compare a virtual manuscript with the original scan and invite others to recheck the steps. The Vesuvius team released data, software and transcriptions, and a detailed preprint describes how the reading was produced and reviewed. That openness gives the discovery staying power beyond the announcement headline.

Hundreds of Herculaneum scrolls remain sealed, and not all will be as readable. Ink chemistry, torn layers and poor scan contrast still impose hard limits. Yet the idea of a lost Roman library has become less abstract: it now has an expanding catalogue of surfaces that teams may be able to inspect without destroying the books.

Using X-ray tomography and machine learning, researchers can virtually unwrap fragile scrolls buried by Mount Vesuvius nearly 2,000 years ago, recovering ancient writing without physically opening them.

Babylon: a forgotten hymn assembled from scattered fragments

Imagine that one verse of a poem survives on a broken tablet in one collection, while the next lines appear on another tablet thousands of kilometers away. Neither fragment is very useful alone. Together, they may reveal a work nobody has been able to read continuously for centuries.

That is what makes the 2025 rediscovery of a hymn to Babylon so appealing. Assyriologists Enrique Jiménez at LMU Munich and Anmar Fadhil at the University of Baghdad worked with an ancient text of about 250 lines, probably composed around 1000 BCE. AI-assisted matching helped the team identify some thirty other copies or related manuscripts. The hymn praises Babylon's buildings and the Euphrates, but it also offers unexpectedly human details: women serving as priestesses and citizens described as treating outsiders with respect.

Schoolchildren in ancient Mesopotamia copied literary compositions as exercises. That habit, which must once have seemed entirely ordinary, can now rescue lost literature: a student copy may preserve a line missing from a more formal tablet. Matching tools make the search across digitized collections faster. Assyriologists still decide whether two sequences genuinely belong to the same work and how damaged signs should be read.

The stakes go far beyond celebrated poems. Cuneiform tablets record commercial arguments, grain deliveries, marriage arrangements, astronomy, medicine and the routines of people who never appeared in a royal chronicle. One recovered sentence can change how we imagine a whole city.

There is a second, separate line of work: translating texts whose language is already understood. A 2023 PNAS Nexus study tested neural machine translation from Akkadian cuneiform into English. Results were promising for some short passages, but the researchers found hallucinations, omissions and a striking example in which the system missed a word of negation. A translation that turns 'does not go' into 'goes' reverses the event. Faster access to thousands of tablets is valuable; expert review remains essential.

The missing letters in stone: Ithaca and Aeneas

Now picture a stone inscription with two names broken away. The surviving line might tell us that citizens dedicated a monument, but not which citizens or to whom. No camera can recover letters carved into stone that no longer exists. Statistical models, however, can study thousands of similar inscriptions and offer candidates based on wording, place and historical period.

Google DeepMind's Ithaca, presented in Nature in 2022, was built to help with ancient Greek inscriptions. Researchers tested it by hiding characters that were known in otherwise surviving texts, allowing them to compare its guesses with actual answers. Ithaca restored about 62% of the hidden sequences correctly; historians working without it achieved roughly 25%, and historians assisted by it reached approximately 72% in the study's controlled exercise. Those figures are not a success rate for every damaged inscription in the real world.

In 2025, DeepMind and academic collaborators introduced Aeneas for Latin inscriptions. It looks for parallels in other texts, combines textual and visual evidence, and proposes restorations, dates and places of origin. In the Nature evaluation, participating historians found its retrieved parallels useful in 90% of the tested cases. Its dating estimates averaged a 13-year distance from the study's reference ranges—not a claim that the model can pinpoint the exact year in which any inscription was carved.

The lesson from both tools is simple: an explanation is worth more when it comes with comparisons. A scholar should be able to inspect the inscriptions that make a proposed restoration plausible—and reject it when better evidence appears.

The age hidden in handwriting: the Dead Sea Scrolls

Sometimes we can read the writing but are not sure when it was written. That can be just as consequential as deciphering a sentence. A date may determine whether a manuscript reflects a new political idea, an older tradition or a later copy of something much earlier.

In 2025, researchers published Enoch, a model that combined radiocarbon measurements with computer analysis of handwriting in the Dead Sea Scrolls. It was trained on 24 carbon-dated samples and then used to estimate the ages of 135 previously undated manuscripts. In the published evaluation, its mean dating errors against the radiocarbon-based reference were around 28–31 years. Some results suggested that certain texts may be older than scholars had assumed.

An error of three decades may sound large until you consider the problem: an ancient manuscript often lacks anything resembling a publication date. Enoch does not inspect a letterform and discover an exact year. It estimates a plausible period using measured samples, handwriting patterns and statistical uncertainty. That gives specialists an additional test—not a replacement for archaeology or radiocarbon dating.

AI tools such as Ithaca and Aeneas help historians analyze damaged inscriptions, suggest missing text and investigate their historical origins. Human experts remain essential for verifying these reconstructions.

Lost languages: where the clues begin to run out

Everything so far has involved writing that scholars can broadly understand once it is made visible. An undeciphered language is a different kind of mystery. We may recognize the signs on a tablet, copy them perfectly and still have no idea what the writer meant. Linear A, used on Bronze Age Crete, is one of the best-known examples.

Historical linguists compare sound patterns, repeated endings and possible family relationships between languages. AI can quickly test candidate connections and search large collections for patterns that people might miss. But machine learning cannot conjure a bilingual inscription or an unknown vocabulary from thin air. Without an independent anchor, even a beautifully patterned solution can remain guesswork.

MIT researchers have explored computational methods for this problem, testing approaches against languages whose decipherment is already understood, such as Ugaritic. Their 2020 work also examined possible relationships for the ancient Iberian script and ways to detect word boundaries where no spaces survive. These are research tools for generating and checking hypotheses, not a demonstrated full reading of an undeciphered language.

Linear A shows why that distinction is so important. The surviving collection is limited, many inscriptions are short, and the language behind the script is uncertain. A University of Melbourne project is investigating deep-learning approaches, including comparisons with potentially related languages. A 2024 systematic review catalogued the obstacles facing these efforts. As of October 2026, there is still no widely accepted decipherment of Linear A.

A model can spot recurring signs and group nearly identical handwriting. It may infer that a certain ending behaves like a grammatical marker. But if the surviving corpus is too small, many different interpretations may fit the same evidence. The missing piece may not be computing power. It may be one more inscription with a known name, a clear archaeological context—or the ancient equivalent of a translation written beside the original.

That is why the Rosetta Stone remains such a powerful metaphor. Connected versions of a known text gave scholars a way to test their understanding of Egyptian scripts. An AI can help find patterns that resemble such clues, but it cannot guarantee that a true Rosetta Stone exists for every lost language.

The Voynich manuscript: perhaps we have the pages in the wrong order

The Voynich manuscript looks as though it was designed to torment future readers: unfamiliar writing winds around strange plants, astronomical diagrams and other enigmatic illustrations. Proposed solutions appear regularly. None has yet persuaded specialists that the whole text has been reliably deciphered.

A 2026 study by Colin Layfield and Lisa Fagin Davis approached the mystery from an unusual direction. Rather than announce a translation, they compared textual patterns and examined the manuscript's physical construction. Their hypothesis is that sheets now bound together in a conventional arrangement may originally have been intended to form a different sequence of individually folded leaves. If so, researchers who assume the present page order is original may have been comparing passages in the wrong order. This is a testable scholarly proposal, not a settled reconstruction or a reading of 'Voynichese.'

The attraction of that idea is that it changes the question. Before demanding that software tell us what a manuscript says, perhaps we should ask whether we have reconstructed the book correctly. Not every meaningful advance arrives in the form of a translation.

The quiet revolution in handwritten archives

Not all lost writing is ancient. An archive may hold tens of thousands of letters or diaries in a familiar language, each impossible to search because the pages exist only as images. Handwritten text recognition (HTR) can convert those scans into draft transcriptions. Unlike cryptanalysis, the task is not to discover a secret code, but to recognize letters formed by particular people, pens and writing habits.

One encouraging example is Ukrainian cultural heritage. In a 2024 study, Aleksej Tikhonov and Achim Rabus developed a general HTR model using material connected with Taras Shevchenko, Lesya Ukrainka, historical correspondence and a diary held in the Holodomor Museum collection. It achieved a character error rate of 4.2% on its validation data. That number does not guarantee the same performance on a stained notebook or an unfamiliar hand, but it shows how much accessible historical material could be made searchable.

For a historian searching thousands of pages for one family name or place, that shift can be transformative. It also introduces a new responsibility: keep the scan beside the transcript. A misread date, surname or negative can lead to an entirely false story.

Before the current AI wave: the letters of Mary, Queen of Scots

The story of computer-assisted codebreaking did not begin with chatbots. In 2023, George Lasry, Norbert Biermann and Satoshi Tomokiyo revealed the contents of 57 encrypted letters connected to Mary, Queen of Scots, written during her captivity in England between 1578 and 1584. Many had remained unidentified in French archives. They described the conditions of her imprisonment, her health and her diplomatic hopes. The team combined algorithms with painstaking manual transcription and cryptanalysis. It was a discovery made by people using computers, not a case of a general-purpose language model solving everything alone.

That history gives the Napoleon story a useful perspective. Frequency analysis, specialist databases and custom algorithms have supported cipher research for decades. Today's AI systems can bring more of those steps together and lower the technical barrier, but the older scholarship remains the foundation. The team's peer-reviewed Cryptologia paper is an excellent example of what transparent historical codebreaking looks like.

The biggest risk: an answer that looks too good

A translation can be grammatically flawless and still tell the wrong story. The 2023 Akkadian translation study contained a revealing case: the source stated that a king did not go to a temple, while the machine translation said that he went. One missing negative reversed the action. On a fragmentary tablet, the difference might influence an entire historical interpretation.

Reconstruction carries a similar risk. A language model can suggest a phrase that would fit a missing patch of Latin or Greek. But if the original surface has been destroyed, there may be no surviving ink against which to test that phrase. Responsible editions distinguish visible letters, uncertain readings and scholarly supplements. AI-assisted editions must preserve those distinctions rather than smoothing them into one seamless paragraph.

Good research makes checking possible. A cipher reading should fit every encoded sign and relevant independent records. An ink-detection system needs physical scans and tests on known material. A dating model should report uncertainty rather than one attractive year. A translation should be open to correction by specialists. In every case, readers deserve a route back to the evidence, not just a fluent final answer.

Access matters too. Scans of Mesopotamian tablets, Greek papyri and Ukrainian manuscripts should not be treated merely as training material extracted from places far from the institutions that hold them. Local researchers, archives and the communities connected with those collections need a role in deciding how images, transcriptions and interpretations are shared. Making knowledge searchable is valuable; making it responsibly accessible is better.

There is a useful parallel with the larger world of generative AI. As Next Horizon explains in How ChatGPT Actually Works, fluent language is not proof that the model checked a fact. The same rule applies when a confident paragraph claims to be 2,000 years old.

What might the next discovery look like?

It may not be another imperial secret. A battered shipping receipt could reveal a forgotten trade route. A student's practice tablet could preserve the missing line of a poem. A batch of newly searchable letters might tell researchers how a town lived through a famine or epidemic. Better scans, more connected archives and more precise recognition tools are making such discoveries increasingly practical.

Many advances will come from hardware as well as software: synchrotron scans, multispectral photographs, chemical analysis and meticulously curated databases. The AI can rank candidate readings or spot overlooked matches, but it cannot replace a missing artifact. The strongest future systems may be those that show a historian not just a suggested answer, but exactly which pixels, letters or parallel texts led to it.

There is something moving about that possibility. History is often told through emperors and battles because their records were more likely to survive. With better tools, the voices returning to the page may belong instead to merchants, students, librarians, doctors and families whose lives rarely entered the official chronicles.

From scanning fragile manuscripts to recognizing forgotten handwriting, artificial intelligence is transforming how historians preserve, search and interpret the written records of human civilization.

Conclusion: the past is still waiting to be read

Napoleon's secret message and the Herculaneum scroll offer two very different images of lost knowledge. One document was deliberately disguised; the other was physically trapped. AI helped recover both, but neither result came from asking a chatbot to improvise ancient prose. One needed cryptanalysis and documentary checks. The other needed X-ray evidence, painstaking virtual reconstruction and experts who could read Greek.

Some texts will remain silent. A burned page may have lost too much of its surface; an unknown language may leave too few clues for any unique solution. That is not a failure of imagination. It is the boundary between evidence and invention.

Yet every newly readable fragment changes the size of the historical record. The next discovery might not rewrite the fate of an empire. It might simply give us back the words of someone who never expected to be heard two thousand years later. That is reason enough to keep looking.

FAQ: AI and the decipherment of history

Did AI really break a 217-year-old Napoleonic cipher?

Yes. In September 2026, Carter Church published a reproducible reading of a ciphered 1809 military letter connected to Napoleon’s orders. The source was a letter sent to General Marmont, not a note handwritten by Napoleon himself.

Can AI read scrolls without opening them?

In some cases. The Vesuvius Challenge combines X-ray tomography, virtual unwrapping, ink detection and expert transcription. In June 2026 the team reported the digital reading of the entire surviving rolled portion of PHerc. 1667—not the destroyed parts of the original scroll, and not every scroll from Herculaneum.

Has AI deciphered an entirely unknown ancient language?

There is no general breakthrough that makes this routine. Computational research helps identify patterns and test language relationships, but writing systems such as Linear A still lack a widely accepted decipherment.

Has AI solved the Voynich manuscript?

No broadly accepted translation exists. A 2026 study proposed that its pages may have been bound in the wrong order, but the suggested reconstruction remains a hypothesis, not a decipherment.

What is the difference between decoding and translating?

Decoding a cipher reconstructs a deliberately hidden message. Translating converts meaningful text from one language into another. Recognizing faint handwriting or restoring a missing inscription is a third, separate challenge.

Will historians and linguists still be necessary?

Yes. They design meaningful questions, identify reliable reference material, check proposed readings, evaluate physical evidence and explain cultural context. AI can speed up parts of that work without replacing those responsibilities.