In spring 2024, more than 100,000 people protested in Taiwan’s streets. On Threads, a parallel fight was underway.
In May of that year, opposition lawmakers introduced legislation that would give Taiwan’s legislature broad new investigative powers, create fines and possible imprisonment for “contempt of the legislature,” and restrict access to legal counsel for those charged. The protest movement, called the Bluebird Movement, lasted for months and eventually triggered one of the largest recall elections in Taiwanese history.
Tracy Weener ’26 and Assistant Professor of Quantitative Social Science Herbert Chang ’18 set out to study the digital side of the protests. Their paper, published this month in Proceedings of the International AAAI Conference on Web and Social Media, identifies a new tactic in political warfare: AI-generated images that look cute and harmless but function as partisan attacks. Weener, the paper’s lead author, and Chang call it “kawaii toxicity”—one of their two findings about how Threads is reshaping political mobilization.
Threads launched in 2023 and quickly became, in Weener’s words, “the Twitter Taiwan never had.” Taiwan accounts for 24% of the platform’s global traffic.
Tracy Weener presents her research at the 2026 International AAAI Conference on Web and Social Media.
“Previous social media platforms in Taiwan were built on social networks, so users largely saw posts shared by people they knew,” Chang says. “Threads is driven largely by algorithms, making it close to the social media equivalent of a public square. Because the Bluebird Movement was the first big political movement in Taiwan after the rise of Threads, we wanted to see how people used it.”
Weener and Chang collected 62,321 Threads posts related to the Bluebird Movement from May 2024 to June 2025. They used AI tools to label the text and visual content in those posts, then cross-checked a subset manually to verify accuracy.
The study builds on earlier work from Chang’s lab examining how AI-generated content shaped the 2024 U.S. election, in which Weener also took part as a James O. Freedman Presidential Scholar, a program that pairs undergraduates with faculty in part-time research assistantships. She continued the research with Chang as a 2026 Hanlon Scholar.
Their first finding in the new study concerns the gap between what users share and what the platform actually shows them. Posts opposing the Bluebird Movement received 28% more views than posts supporting it. But Threads users were far more likely to share pro-Bluebird content, re-posting it 45% more often than anti-Bluebird posts.
“This points to a sharp disconnect between what is showing up on the feed, due to algorithms or other variables, and what users themselves were choosing to promote,” Chang says.
The researchers call this the “dual publics” phenomenon: two competing narratives running simultaneously on the same platform, one shaped by users and one shaped by the algorithm. The gap matters most in moments of political mobilization, when what people see can influence whether they act.
Kawaii toxicity
The second finding emerged from scrolling through the dataset, where Weener and Chang kept encountering AI-generated images that looked, at first glance, like harmless cartoons. Goblins with bluebird wings. Toads with feathers. They were cute, oddly specific, and proliferating.
In context, they were attacks. Opponents of the Bluebird Movement had taken to calling protesters “goblins.” The Taiwanese president’s name resembles the Chinese character for “toad.”
“Animal and plant symbolism were an essential component of the Bluebird Protests, imagery often based on homophonic puns in Mandarin,” Weener says.
Kawaii is a Japanese term for a sense of innocent cuteness. “Kawaii toxicity,” Chang says, “uses cute imagery as a Trojan horse for political polarization.”
“Generative AI tools have made it incredibly easy to generate ‘cute’ images that are actually mean-spirited or dehumanizing about any given group’s political opponents,” he adds. "That means this is something we will likely see more of and should pay attention to.”
The study was one of five from Chang’s lab accepted to this year’s International AAAI Conference on Web and Social Media (ICWSM), a flagship conference for computational social science. The other four undergraduates—Ben Shaman ’26, Sean Noh ’28, Annie Yuan ’28, and Mingyue Zha ’27—presented research on AI-generated visuals in the 2024 U.S. elections, human negotiation with AI, moral contagion on TikTok and Instagram, and mental health discourse on TikTok.
“It’s amazing to see undergraduates producing graduate-level research,” Chang says. “Our students’ original work tackles consequential questions about AI and society.”
Weener says she was honored to present at ICWSM as an undergraduate. After graduation, she’ll head to Taiwan on a Fulbright award to teach English, and she plans to keep the research going. “I aspire to follow this thread of inquiry further, from conducting multi-platform research to studying the policy implications of AI,” she says.