DeepSeek, an artificial intelligence research lab in China, made waves last year by releasing a low-cost, high-performance AI model and chatbot that matched and even surpassed others in the industry.
The DeepSeek model quickly topped the global charts, becoming the most-downloaded app on Apple and Android phones worldwide. Its success quickly raised concerns about narrative bias and foreign interference through AI-driven information environments.
This led Tracy Weener '26 to ask a simple question: Will language models—AI systems capable of understanding and generating human language—lie for their home countries?
Weener, a Hanlon scholar who is pursuing a double major in quantitative social science and computer science with a minor in French, is interested in understanding how misinformation impacts electoral politics and applying data science to government through work in policy analysis.
To dissect the loyalties of language models, Weener and a team of researchers from the Program in Quantitative Social Science audited four of the world's top models—DeepSeek, France's Mistral, and U.S.-based GPT4 and Grok.
Their study, published in the Harvard Kennedy School Misinformation Review, investigated whether large language models created by different companies and countries have different national biases toward world leaders and countries and whether this leads to increased agreement with misinformation. And it revealed surprising nuances.
Their first main finding, for instance, is that "DeepSeek favors Western leaders but rates Xi Jinping of China higher relative to other models, especially in simplified Chinese."
DeepSeek and GPT-4o were asked to rate several regions and world leaders on a five-point scale ranging from very unfavorable to very favorable. Compared to DeepSeek, GPT-4o rates Xi Jinping much lower.
When researchers tested whether the language of interaction mattered, they found that DeepSeek's favorability scores varied significantly between English, simplified Chinese, and traditional Chinese—with simplified Chinese scoring Xi the highest.
Similarly, the model's rating of Western countries and leaders is more favorable when queried in English, reinforcing the link between language models' viewpoints and their data sources.
Correlation between bias and misinformation varied depending on the type of narrative. Models more readily agreed with positively framed misinformation about leaders they favored while diminishing or suppressing negative falsehoods, often shaped by the models’ underlying guardrails—DeepSeek censors negative misinformation about Xi and French President Emmanuel Macron.
"This feels like the opposite of humans on social media, where things like toxicity or animosity can spread faster," says Weener.
The study demonstrates how prompt-based interrogation can reveal whether models' biases are driven by data or induced by guardrails programmed in by developers. "We can see the favoritism of each model and get a glimpse into the internal thoughts of the AI as well," says Weener.
Her previous research has focused on investigating U.S.-Taiwan relations in the digital age. Weener, who hails from Boxford, Mass., and has Taiwanese heritage, traces her interest in the topic to the first time she visited Taiwan, right before the start of the 2024 presidential election in that country.
"As a student, researcher, and journalist, I was inspired by the elections, and I really wanted to know what the Taiwanese electorate was thinking, what young people were thinking," Weener says.
She connected with Herbert Chang ’18, an assistant professor of quantitative social science, and Yusaku Horiuchi, who was a professor of government and the Mitsui Professor of Japanese Studies at the time, and worked with them to examine misinformation narratives on social media in the lead-up to the Taiwanese election through survey experiments and analysis. Weener's work won an award from the North American Taiwan Studies Association.
The U.S. presidential election later that year, which coincided with the rise in popularity of generative AI tools, gave the researchers an opportunity to study how AI-generated images and social media shape political engagement online. They found that AI-generated content alone does not increase exposure and engagement, but when combined with memes, there is an increase in its virality.
In the years to come, Weener hopes to continue studying and contributing to the emerging field of Taiwan studies as well as broader political science research that can help policymakers and citizens alike as they navigate online misinformation, potentially fueled by AI.
As one of the top targets in the world for misinformation, she says, Taiwan offers many examples about how to counter false narratives.
"I've always enjoyed understanding more about policy and using tech as a tool for social good, whether it's policy analysis, or now understanding the implications of an AI-driven information environment," says Weener. "I think we're just literally beginning to ask the questions; forget understanding what the answers are."