Â鶹ÒùÔº

A centurion wearing sneakers and shades? AI image generators sometimes take serious liberties with history. Researchers in classical studies have now developed a tool that creates images of antiquity that are grounded in history, offering fresh insights and perspectives.

Click, and there they are: three Roman soldiers playing a game of dice. But wait—shouldn't the dice look different and be made of bone? Felix K. Maier, professor of ancient history, glances over at Phillip Ströbel, a computational linguist. "I'll adjust the LoRA scale and tweak the prompt," says Ströbel. Admittedly, the terms LoRA scale and prompt aren't ones you'd traditionally expect to hear in a course on history—but they're becoming increasingly common.

Maier and Ströbel have developed , a platform that uses artificial intelligence to generate images depicting the world of Ancient Greece and Rome. To create the platform, they combined three AI models, which they trained on research literature and ancient sources to generate scientifically sound images.

Most mainstream AI image generators draw on modern references and glossy visuals, which can lead to fairly absurd results. In one early test, the system generated an image of a triumphal procession parading through the streets of Ancient Rome—with spectators holding their smartphones up in the air.

Visual experience of history

"The AI may have used pictures of modern-day papal processions," Maier suspects. The people pictured were also extremely muscular and chiseled, and the ancient city of Rome looked far too clean. "It wasn't that easy to make the city look dirtier and get average-looking people," says Ströbel. The new platform, whose AI is fed with accurate reconstructions from researchers' work, can now create much more authentic and accurate images, adds the computer linguist.

The platform's interface is similar to ChatGPT but it has additional features. Users can choose from three different image models, each with its own benefits. In addition, it's possible to automatically refine the prompt, that is, the text that is entered to guide image generation. As Maier explains, the tool is designed for use in three key areas: research, education and museums.

"It enables historians to develop new perspectives," says Maier. "If we want to create an image of a Roman triumph, we now have to ask ourselves concrete questions. How clearly can the triumphator be seen among the jubilant crowd? How was victory staged in an effective way? And which route did the procession follow?" says Maier.

Twenty years ago, the historical TV drama "Rome" raised similar questions about the daily lives of the common people in Ancient Rome. Maier cites the influential historian R.G. Collingwood, who argued that historical understanding requires reenacting past experiences. This is the very idea the project builds on.

"The more deeply we engage with AI-generated imagery, the more our historical imagination is stimulated," says the history expert. He doesn't see this as a threat to human creativity: "Dealing with AI challenges us to continuously assess the plausibility of its output. AI provides us with visual hypotheses, which we then need to ponder."

Dive right into the action

The AI tool doesn't produce definitive visualizations but rather invites users to experiment with different scenes. This can also be particularly effective in educational settings. "When high school students are tasked with depicting the coronation of Charlemagne, the platform lets them dive right into the action and raises important questions: For example, how did the pope position himself to appear larger than the emperor?" says Maier.

The historian believes that tackling these kinds of questions leads to a deeper understanding of historical processes and the interests that drive them. These insights enable high school students to develop their own ideas, evaluate their assumptions and reflect on them. Of course, adds Maier, the use of the tool requires instruction and guidance. However, when used properly, it can help users realize that recording history always involves a degree of interpretation.

Even with fine-tuned models, AI can still make mistakes when prompted to generate images based on historical contexts. But the two researchers view these errors as opportunities to learn rather than as shortcomings. Even if the generated images aren't historically precise in every last detail, these inaccuracies and imperfections may provide valuable learning experiences. "Our method gives rise to new questions, even if not all Roman sandals in the image happen to be depicted accurately," Ströbel says.

It was evident early on that this approach would work, back when students of ancient history were asked to test various AI models. According to the students' feedback, the tool made them view Roman triumphs in an entirely new light. Maier and Ströbel hope to achieve a similar outcome through a planned collaboration with museums, in which visitors would be able to use the platform to generate their own images based on selected exhibition topics. In line with participatory museum practices, they could even create small "exhibitions within the exhibition."

Unconventional collaboration

The research project of Felix K. Maier and Phillip Ströbel brings together two very different disciplines, ancient history and computational linguistics. "Of course, it takes a great deal of communication," Maier says. "Sometimes Phillip has to explain to me over and over how these AI models work."

However, it's precisely this kind of unconventional collaboration that allows them to go further than previous forms of interdisciplinary cooperation. In their day-to-day work, both researchers find that new insights often emerge where seemingly unrelated fields converge. In today's digital world, Maier argues, such cross-disciplinary approaches are indispensable when it comes to tackling complex questions about the past.

The platform Re-Experiencing History, which is open to all members of UZH, is only one of many ongoing digital projects at the university. In the at the chair of Felix Maier, several people are currently working on an AI-supported database that translates classical texts. They also produce an AI podcast on topics related to antiquity.

Maier and Ströbel are convinced that the humanities need to actively engage with artificial intelligence—or risk getting steamrolled by it. With their Re-Experiencing History project, they have opened the gate to a new, visually powerful way of thinking about and experiencing history.