A Vanished City
Growing up in Hong Kong, Lau lived near the site where the Kowloon Walled City once stood. By her childhood, the notorious settlement had been demolished—it was razed between 1993 and 1994—and replaced with a manicured public park, styled as a classical Chinese garden.
As a child, Lau was mostly unaware of what had existed before the park. It was once a dense enclave, with “buildings atop buildings” that housed roughly 33,000 people within six and a half acres. Over time, though, the Walled City began to resurface around her—in films, documentaries, video games, comics, and public conversations about Hong Kong identity.
“My generation's become more and more interested in this topic,” she said recently. “Though the city doesn’t physically exist anymore, people keep representing it in different ways.”
Those competing representations became the foundation of her honors project, the first interdisciplinary study at Bowdoin that combines computer science and history, Lau's two majors.
Faculty in both departments were initially skeptical about the feasibility of her project, said Visiting Lecturer in Computer Science Christopher Martin, who, along with Assistant Professor of History and Asian Studies Guo Jue, advised Lau. “History is a field so focused on studying and preserving different perspectives and viewpoints, while in computer science we strive to find objective answers to problems and...often leave behind the old 'sub-optimal' ways of thinking,” Martin said.
“Yet,” he added, his advisee allayed those worries and “not only successfully married these two opposites but harmonized them into something greater than either singular discipline's perspective could accomplish on its own.”
Lau's research focuses on the Kowloon Walled City not simply as a historical site, but as a case study in how digital technology constructs—and possibly distorts or “flattens”—history, a trend that could accelerate with AI.
“As AI becomes the primary way people encounter historical information because of ease of access, the curatorial choices built into these systems will harden into defaults that become harder to challenge over time,” Lau said.
Guo Jue agreed that Lau's project is not only innovative and ambitious, but timely. “As AI rapidly integrates into the fabric of our lives—including the digital tools we increasingly rely on to access, analyze, and understand the past—humanists, especially historians, are rightfully concerned. Technology is not neutral, nor is history.”
“The Walled City no longer exists physically. What remains are narratives. The generation that lived there is aging. The question of who gets to tell that story is becoming a technology question as much as a historical one. ”
—Wing Kiu Lau ’26
The Walled City
Constructed in 1847 as a Qing dynasty military garrison, the Kowloon Walled City evolved into a singular place. After Britain leased the surrounding territories in 1898, the Walled City remained under Chinese rule while enclosed by British-controlled Hong Kong. That jurisdictional ambiguity led to it becoming mostly ungoverned, and, by the late 1980s, it was one of the most crowded urban settlements in the world.
“It was known for criminal activity,” Lau said. “But recent reparative movements have been trying to highlight the more community-oriented aspects of how people lived there.”
The settlement had unlicensed businesses, large settlements of migrants, limited public services, improvised infrastructure, cheap clinics, tightly packed factories, and apartment blocks built on top of one another. It was dangerous and overcrowded.
But it was also a community where people took care of and relied on each other, creating informal but robust social networks. “It has been called both a cesspool of iniquity and a fully functioning community,” Lau said. “Neither account is wrong.”
Her project investigates how digital information tools decide which of those narratives becomes dominant. “My research asks: how do digital platforms construct historical narratives, and what does the shift from Wikipedia-style transparency to AI-style opacity mean for whose history gets told?” she said.
Research Phase I: Wikipedia
In the first half of her project, Lau examined twenty years of edits to the Kowloon Walled City Wikipedia page. Earlier versions of the article emphasized criminality and urban decay. More recent iterations highlight community life and residents’ experiences, reflecting broader cultural shifts in how Hong Kong citizens think about the city.
“Initially, a lot of the focus was on how dense it was, its criminal activity, the triad gangs, and sovereignty issues—the tension between China and the UK,” Lau said. “Later on, you start seeing a lot more emphasis on community aspects, with more residents talking about their life in the city.”
Yet, despite Wikipedia’s reputation as a democratized knowledge platform, her findings revealed limitations. About 75 percent of the sources shaping the article came from nonlocal institutions, including British archives, Western publishers, and international media. Local oral histories and firsthand resident accounts appeared less frequently, in part because Wikipedia privileges formally published and easily verifiable sources.
Contributors must understand Wikipedia’s conventions and sourcing rules, and the result is that the site can offer imbalanced historical narratives.
“Even though Wikipedia was built to democratize access to information and let anyone be a contributor, my conclusion is that, despite its promises, it has historiographic limitations,” Lau said.
Research Phase II: AI
This led her to the second part of the project: an experiment examining how users interact with AI-generated historical narratives. “With AI, which is more popular now, you tend to get a single narrative—after prompting it—without any idea of what the sources are behind the narrative,” she said.
To test how AI is telling history and how users might be absorbing that information, Lau created her own Retrieval-Augmented Generation, or RAG, systems—AI models that tap more restricted collections of materials to answer users' questions. Some of her RAG systems drew from oral histories. Others relied on newspapers or academic publications.
She then invited a group of Bowdoin students to test the systems while evaluating how trustworthy they found the responses.
Lau's project benefited from the archival research she did in Hong Kong with support from the history department's Paul Nyhus Travel Grant. During the summer, she spent many days at the Hong Kong Public Records Office, scanning newspapers and collecting oral history materials from local exhibitions.
“My Chinese-language literacy enabled me to access sources that the English-language record largely excludes,” she said.
Some participants were allowed to see the underlying source materials and trace where the claims originated. She called these students the "informed population.”
Others were asked to interact with anonymized systems that concealed those pathways, becoming the “uninformed population.”
Core Findings: Digital Memory Making
Though in theory Wikipedia makes its sources known and its edits visible, users don’t tend to check this record, Lau argued. “So the visibility doesn’t necessarily lead to accessibility of information.”
AI systems compound the problem. Even when the models produced citations for the informed Bowdoin students, participants tended to lean on questionable indicators of credibility.
“With my study, I found that even when the AI model produced citations for the informed population, participants defaulted to whether the responses sounded authoritative, looking at the tone and formality of the response, which is how AI sounds whether its sources are reliable or not,” Lau said.
She concluded that, “Wikipedia makes its socially-mediated editorial process transparent, but not accessible enough for most people to actually engage with it. AI makes its process invisible entirely, through opaque algorithmic synthesis,” she said.
And in that gap between visibility and accessibility, accountability is lost, she said.
Implications for Computer Science and History
Both of the platforms Lau worked with tend to consolidate narratives rather than preserve plurality, she said, describing this process as “flattening” historical narratives.
The implications concern both of her majors. In computer science, researchers are working to make AI systems more ethically accountable. In history, scholars are grappling with how generative AI may reshape the field—especially as more people encounter the past through AI’s conversational style rather than through books, archives, or even Wikipedia.
Martin said that Lau’s work “dissecting the epistemic crisis of knowledge authorship, dissemination, transparency, and trust is crucial now more than ever as the world quickly and blindly shifts towards more generative AI use.”
He added that those working at the cutting edge of AI, who view history simply as a field AI can “do,” would be wise to adopt a historian’s approach by incorporating multiple perspectives and traceable sources, making their tools more transparent and trustworthy.
Guo Jue, too, said that generative and algorithmic tools like Claude and ChatGPT have enormous implications for historical research. “They produce seemingly fluent, neat, and singular narratives without showing what evidence is marshaled or how reasoning is conducted,” she said.
Historians like her are asking how they should write and teach about the past when archives are increasingly mediated by these systems. “We won’t find out without experimenting. Wing Kiu’s project is exactly such an experiment. It clearly demonstrates that historians must examine not only the politics and economics of technology—the usual subjects of history—but also the technical mechanisms through which archives are mediated and narratives generated.”
