How AI May Reshape Which Historical Narratives Survive
A Bowdoin College honors student examined Wikipedia edits and AI systems to understand how digital platforms decide which versions of history become dominant. Wing Kiu Lau '26 studied the Kowloon Walled City-a demolished Hong Kong settlement-to investigate a question with implications for how future generations encounter the past.
Her research reveals a problem: as more people access historical information through AI rather than archives or books, the systems that generate those narratives may become invisible gatekeepers of collective memory.
The Physical City and Its Digital Afterlife
The Kowloon Walled City was razed between 1993 and 1994 and replaced with a public garden. But the settlement persists in films, documentaries, video games, and public conversation about Hong Kong identity.
Built in 1847 as a Qing dynasty garrison, the site became a jurisdictional anomaly after Britain leased surrounding territories in 1898. The Walled City remained under Chinese rule but surrounded by British-controlled Hong Kong. This ambiguity left it largely ungoverned.
By the late 1980s, it housed roughly 33,000 people across six and a half acres. The settlement had unlicensed businesses, migrant populations, improvised infrastructure, and tightly stacked apartment blocks. Criminal activity flourished. So did informal social networks where residents depended on each other.
"It has been called both a cesspool of iniquity and a fully functioning community," Lau said. "Neither account is wrong."
Twenty Years of Wikipedia Edits
In the first phase of her project, Lau analyzed two decades of changes to the Kowloon Walled City Wikipedia page. Early versions emphasized criminality and urban decay. Recent ones highlight community life and resident experiences.
Yet the platform's reputation for democratized knowledge concealed structural limitations. About 75 percent of sources came from nonlocal institutions-British archives, Western publishers, international media. Local oral histories and firsthand accounts from residents appeared less frequently because Wikipedia privileges formally published, easily verifiable sources.
"Even though Wikipedia was built to democratize access to information and let anyone be a contributor, it has historiographic limitations," Lau said.
The Opacity Problem With AI
The second phase tested how users interact with AI-generated historical narratives. Lau created her own AI systems-called Retrieval-Augmented Generation models-that drew from different source collections: some from oral histories, others from newspapers or academic publications.
She invited Bowdoin students to test these systems and evaluate trustworthiness. Some participants could see the underlying sources and trace where claims originated. Others interacted with systems that concealed those pathways.
A crucial finding emerged: even when the AI provided citations, participants judged credibility by tone and formality rather than source quality. "AI sounds authoritative whether its sources are reliable or not," Lau said.
Wikipedia makes its editorial process transparent but not accessible enough for most users to actually check it. AI makes its process invisible entirely through opaque algorithmic synthesis. In that gap between visibility and accessibility, accountability disappears.
Implications for Two Fields
Both Wikipedia and AI systems consolidate narratives rather than preserve competing perspectives-what Lau calls "flattening" historical narratives. This matters to computer scientists working on ethical AI and to historians grappling with how generative AI may reshape their field.
As more people encounter the past through AI's conversational interface rather than through archives or books, the technical choices built into these systems become defaults harder to challenge over time. "The curatorial choices built into these systems will harden into defaults that become harder to challenge over time," Lau said.
Historians increasingly ask how they should write and teach when archives are mediated by these systems. Generative AI tools like Claude and ChatGPT produce fluent, singular narratives without showing what evidence was marshaled or how reasoning was conducted.
"The question of who gets to tell that story is becoming a technology question as much as a historical one," Lau said.
Computer scientists developing AI research would benefit from adopting a historian's approach: incorporating multiple perspectives, showing traceable sources, and making tools more transparent and trustworthy. Without that shift, the systems that answer questions about the past may determine not just how history is presented, but which histories survive at all.
Your membership also unlocks:
