In February this year, Google released an upgraded version of the Gemini artificial intelligence model. This quickly turned into a propaganda disaster, as it was discovered that requests for photos of Vikings produced tough-looking Africans, while photos of Nazi soldiers included Asian women. Demands for racial diversity produce absurd inaccuracies.
Academic historians are confused and shocked. “They clearly didn’t consult historians,” said Benjamin Breen, a historian at the University of California, Santa Cruz. “Everyone who cares about the past wonders, ‘What happened?'”
Rewriting the past to fit contemporary political fashions is simply not how historians envision artificial intelligence. Instead, machine learning, large language models (LLMs), machine vision, and other artificial intelligence tools offer the opportunity to develop a richer and more accurate view of history. Artificial intelligence can decipher damaged manuscripts, translate foreign languages, discover previously unrecognized patterns, make new connections and speed up historical research. As teaching tools, artificial intelligence systems can help students understand how people in other eras lived and thought.
Brin believes that historians are uniquely suited to harness artificial intelligence. They are used to working with texts, including vast works that are not subject to copyright, and they know not to believe everything they read. “The main thing is to be completely skeptical of the source text,” Brin said. When it comes to using artificial intelligence, he said, “I think that’s part of the reason why the history students I work with are more sophisticated from the beginning than the random celebrities I see on Twitter.” Historians take a closer look Check the results for errors, just as they would check claims in 19th-century biographies.
Last spring, Breen created a customized version of ChatGPT for use in his medieval history class.
He wrote detailed system prompts, resulting in a chatbot that could interact with three characters living during the outbreak of the Black Death in 1348: a traveler passing through Damascus, a notorious Parisian apothecary, and a Honest councilor of Pistoia. The simulation works like a much more complex version of a text-based adventure game – the great-great-great grandchild of the 1970s classic Oregon Trail.
Each student chooses a character—a Parisian apothecary, for example—and receives a description of their setting, followed by a question. The apothecary looked out the window and saw a group of penitents whipping themselves with belts. what does he do? Students can choose from a range of options or improvise a unique answer. Based on the response, the chatbot continues the narrative.
After the competition, Brin asked the students to write essays analyzing how accurately their simulations described historical context. This comprehensive exercise immerses students in medieval life while also teaching them to be wary of the illusions of artificial intelligence.
This is a pedagogical triumph. Students responded with extraordinary creativity. Brin writes in his Substack newsletter that one of them “made a heroic effort to stop the spread of the plague with perfume, like an Italian doctor named Gilbert,” while the other ” He fled into the forest and became an itinerant hermit.” Others “became leaders of successful and failed peasant uprisings”. The students, who usually sat at the back of the classroom looking bored, enthusiastically participated in the game. The engagement, Brin wrote, was “unlike anything I had ever seen.”
For historical research, ChatGPT and similar LL.M.s can be powerful tools. Their ability to translate old text is better than professional software like Google Translate because their training material includes context in addition to language. As a test, Breen asked GPT-4, Bing’s Creative Mode, and Anthropic’s Claude to translate and summarize a passage from a 1599 book on demonology. The text was written primarily in “a highly learned form of Latin” that included some Hebrew and ancient Greek. Results were mixed, but Brin found that “Claude did an outstanding job.”
He then gave Cloud a chunk of the same book and asked it to create a chart listing the types of demons, what they were thought to do, and the page numbers on which they were mentioned. This chart isn’t perfect, mostly because the page numbers are difficult to read, but it’s useful.Such charts, Brin writes, “ultimately will be a game-changer for anyone doing multilingual research. This is not about making artificial intelligence replace you. Instead, it asks AI to act as a knowledgeable research assistant, providing you with clues.
The LL.M. can read and summarize articles. They can read old patents and interpret technical diagrams. They find useful nuggets in tedious texts, for example, identifying each of the diarist’s journeys. “It’s not going to do everything right, but if it’s narrow enough in its focus, when you give it documents to study, it’s going to do a good job in this kind of historical research,” said Steven Lubar, a historian at Brown University. Pretty good. “I found this very useful. “
Unfortunately, the LL.M. still can’t decipher the old handwriting. They are not good at finding resources on their own. They are not good at summarizing debates among historians, even if they have the relevant literature at hand. They are unable to translate impressive patent explanations into believable illustrations. When Lubar asked for a picture of the binder described in a 19th-century patent, what he got instead was the opening of a briefcase showing the steampunk mechanism used to write music scores. “It’s a beautiful picture,” he said, “but it has nothing to do with patents, it explains it very well.”
In short, historians still have to know what they are doing, and they have to check the answers. “They were tools, not machines,” said Lubar, whose research includes the history of tools. Machines operate on their own, while tools extend human capabilities. “You don’t just push a button and get results.”
Just knowing that these new tools are possible unlocks historical resources and allows for new questions and approaches. Get the map. Thousands of map series exist, regularly documenting the environment, and many have been digitized. They show not only terrain, but also buildings, railroads, roads, and even fences. Maps of the same place can be compared over time, and in recent years historians have begun using big data from maps.
Katherine McDonough is a historian at Lancaster University in the United Kingdom, and her thesis is on road construction in eighteenth-century France. She was attracted to digital tools but frustrated by their inability to solve her research problems. Map data comes primarily from 19th and 20th century series from the United States and the United Kingdom. Anyone interested in old maps of France is out of luck. McDonald hopes to find new methods that can be applied to a wider range of maps.
In March 2019, she joined a program at the Alan Turing Institute, the UK’s national center for data science and artificial intelligence. Macdonald knew that the National Library of Scotland had a large collection of digitized maps and suggested looking at them. “What can we do with access to thousands of digitized maps?” she wondered. The team worked with computer vision scientists to develop software called MapReader, which MacDonald describes as “a way to ask map questions.”
She and her colleagues combined maps with census data to examine the relationship between railroads and class-based residential segregation. “The real power of maps is not necessarily looking at them in isolation, but being able to connect them to other historical data sets,” she said. Historians have long known that upper-class Britons lived closer to passenger stations than to railway stations. The rail yard seemed like an obvious nuisance due to noise and smoke, and the lower-rise neighbors lacked better options. Pairing the map with census data on occupation and address shows a more subtle effect. People who live near a railroad station are likely to work at a railroad station. Not only do they save on rent, they also reduce commuting time.
MapReader does not require extreme geographic accuracy.It leverages techniques used in biomedical imaging to divide maps into sections called patch. “When historians look at a map, we want to answer the question, we want to know how many times does something like a building appear on this map? I don’t need to know the exact pixel location of each building,” says Mike Donner said. In addition to simplifying calculations, the patchwork approach encourages people to remember that “maps are just maps. They are not the landscape itself.”
In a nutshell, this is the answer historians can teach us about artificial intelligence. Even the best responses have their limits. “Historians know how to deal with uncertainty,” MacDonald said. “We know that much of the past is gone.”
Before photography, everyday images were scarce. Journalism didn’t exist before printing. Life was not recorded on paper, business records were shredded, courthouses were burned, and books were lost. The conquerors destroyed the chronicles of the conquered. A natural disaster strikes. But the tantalizing traces remain. Artificial intelligence tools could help recover new fragments of the lost past, including troves of ancient texts.
When Mount Vesuvius erupted in AD 79, it buried the seaside resort of Herculaneum, near modern Naples and the larger ancient city of Pompeii. The town’s wonders were rediscovered in the 18th century and include a magnificent villa believed to have belonged to Julius Caesar’s father-in-law. Early excavators discovered more than 1,000 papyrus scrolls there—the largest surviving such collection in the classical world. Archaeologists believe that thousands of people may still be alive in the still-buried parts of the villa. Historian Garrett Ryan writes: “If these texts are discovered, and if even a small portion remains readable, they will transform our understanding of classical life and history on a scale not seen since the Renaissance. Literary understanding.
Unfortunately, the Herculaneum Scrolls were carbonized by volcanic heat, and many of the scrolls were damaged during early readings. Only about 600 of the items initially found were intact and looked like charcoal blocks or charred logs. In February this year, one of the volumes, which had not been seen for nearly 2,000 years, began to be read.
This milestone represents a triumph for machine learning, computer vision, international collaboration, and the age-old lure of wealth and glory. The quest began in 2015, when researchers led by Brent Seales at the University of Kentucky figured out how to use X-ray tomography and computer vision to “open” ancient scrolls. This technology creates a computer image of the page’s appearance. But distinguishing letters from parchment and clay required more progress.
In March 2023, Seales, along with emerging investors Nat Friedman and Daniel Gross, announced the Vesuvius Challenge, offering huge prizes for key steps in reading the Herculaneum Scrolls. The challenge attracted international talent and was an almost immediate success. By the end of the year, a team consisting of students Youssef Nader, Luke Farritor, and Julian Schilliger had cracked the first volume (approximately 2,000 characters) and won the $700,000 grand prize. “We couldn’t have done this without the technical staff,” enthuses Richard Janko, professor of classical studies at the University of Michigan. wall street journal.
Although only about five percent of the text has been read to date, this is enough for scholars to determine the scroll’s point of view and themes. “Epicureanism greets you with words full of music, food, sensuality and joy!” enthuses Federica Nicolardi, a papyrologist at the University of Naples Federico II. This year, the program promises a $100,000 prize to the first team to crack 90% of four different reels. Recovering the lost scrolls of Herculaneum is the most compelling example of how artificial intelligence (future technology) promises to enhance our understanding of the past.
