# The Past Is Not Finished AI can make old letters, hidden cities and closed books readable again. It expands the past — and makes our standards of evidence more important. - Forfatter: Mikkel Freltoft Krogsholm - Type: Essay - Udgivet: 2026-10-05 - Opdateret: 2026-10-05 - Sprog: en - Emner: ai, history, archaeology, science, human control - Kanonisk URL: https://mikkelkrogsholm.dk/en/articles/fortiden-er-ikke-faerdig/ --- In 1809, a French general was isolated on the eastern side of the Adriatic. Auguste de Marmont had around 13,000 soldiers under his command. Austrian territory lay between him and the rest of Napoleon’s armies, and a new war was approaching. Napoleon therefore ordered his stepson Eugène de Beauharnais to send Marmont an overview of where the French and allied forces were located. The letter was to be written in code and carried by a trusted officer. It survived. The key did not. The only publicly known reproduction ended up as a grainy image in a French military history journal from 1969. At the top was a single sentence in ordinary French. Then came 24 lines of numbers, letters and hand-drawn symbols. For more than two hundred years, no one could read the rest. In September 2026, AI engineer Carter Church gave the image and a simple task to OpenAI’s AI system GPT-6 Astra. Around six hours of machine work later, the system had [reconstructed a reading of the letter](https://carter.church/writeups/the-letter-to-marmont/). Not because the model knew a single secret method that humans had overlooked. It completed an entire chain of different tasks. First, it had to read 1,300 characters from the poor image. It had to determine when two slightly different squiggles were probably the same symbol written by an inconsistent hand. It then found an old table online compiled by the French cryptology historian Daniel Tant. The table explained 33 of the letter’s 155 different symbols, but no one had previously connected it to this particular letter. That yielded roughly a third of the text. For the rest, the model built a program that tried different letters and words and measured how closely the result resembled French. Finally, it checked the proposed meanings against every place each symbol appeared in the image. It almost sounds like a story about a machine solving the past on its own. It was not. ## A Problem No One Could Afford Humans had already done much of the work. One researcher had found the letter. Another had assembled the partial key. Archives had digitised the journal. Napoleon’s correspondence had been published in book form as early as the nineteenth century. Marmont had written memoirs. Historians knew the course of the war. The model did not find traces that had disappeared. It assembled traces that were scattered. That also makes the solution unusually well suited to verification. The reconstructed key can be applied to every symbol in the letter, and Church has published both the material and the programs behind the reading. Napoleon’s original order to Eugène also survives. It names the same armies in the same sequence and with the same numbers that appear in the decoded letter. An incorrect key would therefore have to do more than produce random French sentences. By some miracle, it would have to produce a coherent military message matching another preserved order from the same moment. Historians are still necessary. As [Live Science describes](https://www.livescience.com/technology/artificial-intelligence/a-kind-of-rosetta-stone-ai-decodes-217-year-old-secret-letter-ordered-by-napoleon), the letter does not suddenly tell an entirely different story about Napoleon. It largely repeats what researchers already knew, and it has to be interpreted alongside the rest of our knowledge about the war. What is new is that we can now read the letter itself. Church calls it a problem no one could afford. A specialist could probably have solved it years ago. But it would have required weeks of old French, 1,300 indistinct symbols, statistics and source checking — all for a single plate in one military history journal. The people capable of doing it had more important questions on which to spend their time. The letter was therefore not unreadable in the strict sense. It was stranded in the gap between what was possible and what was practically possible. AI moved that boundary. ## Every Age Gets New Eyes It would be easy to tell the story as if AI were giving humanity access to a closed past for the first time. That would be wrong. We have always used the technological advances of the present to ask new questions of what has already happened. In the late 1940s, the American chemist Willard Libby developed a method for measuring the age of organic material using the radioactive carbon isotope carbon-14. Living organisms absorb carbon while they are alive. After death, the amount of carbon-14 declines at a known rate. By measuring how much remains, researchers can estimate when the organism died. It gave archaeology a new clock. Layers, tools and human bones could be placed in time in a different way than before. Libby received the [Nobel Prize in Chemistry in 1960](https://www.nobelprize.org/prizes/chemistry/1960/summary/) for the method. Later, satellites and aircraft gave us a new view of the landscape. Archaeologists had previously needed to move through deserts and jungles with old accounts, maps, local knowledge and well-founded guesses. With images and measurements from the air, they could suddenly see patterns that were invisible from the ground. Radar can penetrate dry sand under certain conditions. LiDAR can send laser pulses through small gaps in the canopy and create a precise elevation map of the ground below. In this way, roads, embankments, foundations and entire urban landscapes can emerge beneath desert or forest. [NASA describes](https://science.nasa.gov/earth/earth-observatory/searching-for-the-origins-of-space-archaeology/) how satellite and radar images helped researchers trace ancient trade routes in Oman, and how declassified images from American spy satellites later became an archive of landscapes that had since been built over, bombed or ploughed away. In Central America, LiDAR has shown that Maya cities beneath the jungle were far larger and more interconnected than excavations alone had revealed. The satellite did not write the history. It showed the human where to look. The archaeologists still had to travel, excavate, date and interpret. But they had acquired a new sensory organ. Ancient DNA gave us something else: a new map of kinship. By extracting damaged genetic material from old bones, researchers could investigate extinct forms of humans and their relationship to us. Svante Pääbo received the [Nobel Prize in 2022](https://www.nobelprize.org/prizes/medicine/2022/summary/) for his discoveries concerning the genomes of extinct hominins and human evolution. Carbon-14 told us more about *when*. Remote sensing told us more about *where*. DNA told us more about *who*. Each technology did more than provide better answers. It made new questions possible. ## Letters That Cannot Be Opened Sometimes several generations of technology must be stacked on top of one another before the past becomes readable. When Vesuvius erupted in AD 79, a collection of papyrus scrolls in Herculaneum was carbonised. They were found in the eighteenth century, but many now resemble black, compressed lumps. Attempts to physically unroll them can destroy them. You could possess the text without being able to read it. In 2023, high-resolution scans of the scrolls became the basis of the [Vesuvius Challenge](https://scrollprize.org/grandprize). Participants used three-dimensional models, image analysis and machine learning, among other things, to find barely visible ink inside the still-rolled layers of papyrus. Less than a year later, they could begin reading words that had not been seen for around two thousand years. AI was an important part of the breakthrough, but it did not stand alone. Without the highly detailed three-dimensional X-ray scan, the mathematical reconstruction of the curled layers, the papyrus experts and the open competition, the model would have had nothing to work with. That is worth remembering because we tend to make the newest tool the entire explanation. Mysteries of the past are rarely solved by a single instrument. They are solved when a new capability is added to a long chain of human care. ## From New Instrument to New Collaborator AI fits into this history, but it also has a slightly different character. A carbon meter can date a sample. A satellite can observe the landscape from above. DNA analysis can read biological material. Each instrument makes a particular kind of trace legible. AI can work across them. In the Marmont letter, the model moved from image to symbol, from symbol to a partial key, from the key to a statistical program, and from the program’s result to historical letters and memoirs. It was not merely using a new sense. It organised an investigation across several senses and specialist tools. It is the same ability I previously described in [*AI Finds the Connections We Overlook*](https://mikkelkrogsholm.dk/en/articles/ai-finder-de-forbindelser-vi-overser/). What is new often lies not in the individual building block but in the long chain: find a trace, connect it with another, test the connection, learn from the error and continue. Historical research is full of such chains. A damaged inscription does not merely have to be read. It may need to be located geographically from its dialect, dated from names and letter forms, and compared with thousands of other texts. Researchers have developed the AI system Ithaca to assist with ancient Greek inscriptions. In a [controlled experiment published in Nature](https://www.nature.com/articles/s41586-022-04448-z), in which parts of texts were hidden, historians working alone restored the characters correctly in 25 per cent of cases. Ithaca alone reached 62 per cent. When historians received the model’s suggestions as assistance, they reached 72 per cent. The best result came from neither human nor machine alone. The model could search for patterns in tens of thousands of inscriptions. The historian could assess language, society and context. The machine made the space of possibilities visible. The human could see what made historical sense. Perhaps this is how AI becomes a new tool for the past: not only as a measuring instrument, but as a collaborator able to move between the measurements. ## The Queue of Unsolved Questions In [*When the Dead Speak*](https://mikkelkrogsholm.dk/en/articles/naar-de-doede-taler/), I wrote about another scarce resource: attention. The past does not necessarily lack material. Letters, accounts, parish records, photographs, court rulings, maps and objects sit in archives and storage rooms. The problem is that a person can spend an entire working life collecting and understanding only a small part of it. We therefore know the past not only through what was preserved. We also know it through what someone had the time and reason to investigate. Napoleon received enough attention. Even so, one of the letters in his enormous legacy remained unread. Imagine, then, the long tail of less famous people, local archives, fragmented inscriptions, forgotten maps and finds that no specialist can justify spending months on. Some mysteries are certainly unsolved because the evidence is gone. No model can conjure it back. But there is another category: questions where the traces are already there but distributed across too many images, archives, languages or disciplines. Questions that are possible to solve but too expensive to pursue. This is the queue AI can begin to shorten. It could change who gets to explore. Carter Church was not an expert in Napoleon’s codes. He was a curious AI engineer who could put a powerful model to work and then publish a solution that others could verify. When a specialist’s abilities can partly be made available through a system, more people can test a qualified question. Not as a replacement for experts, but as a way of opening more trails than the field itself has time to follow. It is the democratisation of curiosity. Not everyone will be right. But more people will have the opportunity to investigate something properly. ## The Plausible Past There is also a danger in that very ability. A language model — an AI system trained to process and produce language — is very good at creating coherence. That is useful when the coherence truly exists. It is dangerous when the gaps in the material are so large that the model must bridge them with what is merely likely. The past cannot object. A general cannot tell us that the model has misunderstood his abbreviation. A woman from the nineteenth century cannot correct the voice we have constructed from her letters. A ruin cannot explain whether a straight line in the earth was a road, a wall or a natural feature. The more vivid and coherent the result becomes, the easier it is to forget the difference between trace and story. That is why the Marmont letter matters for another reason besides the dramatic name. It has receipts. The proposed key can be applied to the entire text. The program can be run again. The symbols can be compared with the image. The result can be checked against Napoleon’s preserved order and the known military situation. Uncertain symbols can be marked as uncertain instead of being smoothed away. It is a model for how we should use AI on the past. Not by asking: Can the machine tell a convincing story? But by asking: Can it show us the path from the trace to the claim? The carbon measurement has a margin of uncertainty. The satellite image requires verification on the ground. The DNA sample may be contaminated. A proposed text must match the actual characters. AI does not remove the old demands of evidence. It makes them more important because it can produce so much more that sounds right. Church put it precisely: Potential solutions can be automated. Responsibility cannot. ## History Opens Again History books can make the past appear closed. The events stand in sequence. The dates are printed. The cities have been found. The letters have been quoted. The dead have been given their places. But the past is not finished in that sense. The events themselves do not change. Our access to them does. A new chemical clock moves the dating of a settlement by several centuries. A map made with laser light shows that a city continued far beneath the forest. A DNA sample connects populations archaeologists had previously kept separate. A scan opens a book without anyone physically touching its pages. An AI finds a cipher key and turns 1,300 silent symbols back into French words. Every new technology gives humanity new senses. Some we turn towards the future. Others we turn backwards. AI is not a time machine. It cannot visit 1809, and it cannot recover what left no trace. But it can help us see more of the traces we already have and follow connections that were previously too slow, too narrow or too expensive to pursue. That means some of humanity’s unresolved past may not be waiting for new discoveries. It may be waiting for new ways of reading the discoveries we already have. The past does not change. But what humanity can make it tell us does. --- ## Sources and Further Reading - Carter Church: [*Breaking the Marmont Cipher, 1809*](https://carter.church/writeups/the-letter-to-marmont/), 2026. - Live Science: [*“A kind of Rosetta Stone”: AI decodes 217-year-old secret letter ordered by Napoleon*](https://www.livescience.com/technology/artificial-intelligence/a-kind-of-rosetta-stone-ai-decodes-217-year-old-secret-letter-ordered-by-napoleon), 2026. - NASA Earth Observatory: [*Peering through the Sands of Time: Searching for the Origins of Space Archaeology*](https://science.nasa.gov/earth/earth-observatory/searching-for-the-origins-of-space-archaeology/). - Nobel Prize: [*The Nobel Prize in Chemistry 1960*](https://www.nobelprize.org/prizes/chemistry/1960/summary/). - Nobel Prize: [*The Nobel Prize in Physiology or Medicine 2022*](https://www.nobelprize.org/prizes/medicine/2022/summary/). - Yannis Assael et al.: [*Restoring and attributing ancient texts using deep neural networks*](https://www.nature.com/articles/s41586-022-04448-z), Nature, 2022. - Vesuvius Challenge: [*2023 Grand Prize awarded: we can read the scrolls*](https://scrollprize.org/grandprize), 2024.