The First Mass Protest on Threads: Multimodal Mobilization and AI-Generated Visuals in Taiwan's Bluebird Movement

1. Problem Statement & Research Context

The paper investigates the 2024 Bluebird Movement in Taiwan, a youth-led protest against parliamentary reforms that threatened democratic checks and balances. This event is unique for three reasons:

• Platform Migration: It occurred primarily on Threads (Meta), a new microblogging platform that, unlike in most markets, captured 23.75% of its global traffic from Taiwan, effectively becoming the "Twitter Taiwan never had."
• The Algorithmic-Engagement Gap: Activists faced a divergence between algorithmic exposure (what the platform's feed ranking showed) and social engagement (what users chose to amplify). Meta had previously announced a tapering of political content, creating a tension between platform curation and user mobilization.
• Generative AI Integration: The movement saw the unprecedented, large-scale deployment of AI-generated imagery (specifically "cute" or kawaii aesthetics) used for both mobilization and partisan attacks, introducing a new form of political toxicity.

Research Questions (RQs):

1. What is the divergence between algorithmic exposure and user engagement across partisan lines?
2. Which textual and visual communication strategies predict virality?
3. How is generative AI deployed in protest and counter-protest contexts?

---

2. Methodology

The study employs a multimodal computational social science approach, combining large-scale data collection, Large Language Model (LLM) annotation, and machine learning interpretability techniques.

A. Data Collection

• Source: Meta Content Library (MCL).
• Dataset: 62,321 unique posts and 21,572 unique images collected between May 2024 and June 2025.
• Filtering: Posts were filtered for Traditional Chinese and the keyword "Bluebird" (qīngniǎo).
• Metrics: The dataset includes exposure metrics (views) and engagement metrics (likes, comments, reposts).

B. Data Annotation (LLM Zero-Shot Learning)

The authors utilized GPT-4o with zero-shot prompting to annotate the dataset without task-specific labeled training data.

• Text Annotation: Extracted identifiers (hashtags, locations), categorized content (e.g., Commemorative, Personal Testimony, Call to Action, Condemnation), and estimated partisan stance (pro/anti-DPP, KMT, TPP) with probability scores (0–100).
• Image Annotation: Categorized visual elements (real vs. AI, human vs. non-human), identified specific symbols (animals, plants), and coded political intent (mobilization vs. attack).
• Validation: Manual spot-checking confirmed 97% relevance to the movement; strict JSON schema validation ensured output consistency.

C. Machine Learning & Analysis

• Model: CatBoost (Gradient Boosting Trees) was used to regress engagement metrics (likes, reposts) based on textual and visual features.
• Interpretability: SHAP (SHapley Additive exPlanations) values were computed to determine feature importance and directionality (positive/negative impact on virality).
• Hyperparameters: The model used a 75-25 train-test split, learning rate of 0.01, 1,500 epochs, and tree depth of 12, optimized via grid search.

---

3. Key Results

RQ1: The Exposure-Engagement Gap (Partisan Asymmetry)

• Supply vs. Demand: While the volume of anti-DPP content (14.3%) and anti-KMT content (10.6%) was relatively balanced in the dataset, their circulation patterns diverged significantly.
• Algorithmic Bias: The platform's algorithm favored anti-DPP content, granting it 28% more views than anti-KMT content.
• User Amplification: Conversely, user behavior (reposts and likes) heavily favored anti-KMT and pro-DPP content. Anti-KMT posts received 45% more reposts than anti-DPP posts.
• Conclusion: There is a fundamental misalignment where the algorithm surfaces content that users do not necessarily endorse, while users actively recirculate content that reinforces their partisan identity or attacks opponents, creating "overlapping publics."

RQ2: Predictors of Virality (Textual & Visual Strategies)

• Textual Drivers: The strongest predictors of likes were:
  1. Commemorative content (referencing historical parallels like the 2014 Sunflower Movement).
  2. Personal Testimonies.
  3. Calls to Action.
  • Tone: Formal language and negative sentiment correlated with higher engagement, contrasting with theories that humor drives virality.
• Visual Drivers:
  • Human Photographs: Generated the highest views and replies (especially photos of legislators, which sparked debate).
  • Non-Human/AI Visuals: Generated significantly higher reposts. These served as low-friction mobilization tools rather than discussion starters.

RQ3: Generative AI and "Kawaii Toxicity"

• Mobilization Symbols: Pro-DPP/Anti-KMT groups used AI to generate cute animal symbols based on homophonic puns (e.g., Qing Niao = Blue Bird/Blue Island; Dong Lu = Winter Deer/East Road). These symbols (Blue Birds, Winter Deer, Whales) circulated widely as portable shorthands for collective identity.
• Toxic Symbols: Anti-DPP groups utilized AI to generate "cute" images of goblins and toads to attack opponents.
  • Goblins: Originally anti-TPP, shifted to anti-DPP following harassment scandals.
  • Toads: Used to mock President Lai Ching-te (phonetic pun on his surname) and DPP legislators supporting martial law.
• Kawaii Toxicity: The study identifies a new paradigm where extreme political toxicity is cloaked in "cute" aesthetics. This allows partisan attacks to bypass moderation filters and lower the psychological barrier for sharing toxic content.

---

4. Key Contributions

1. Theoretical Extension of "Algorithmic Publics": The paper empirically demonstrates that algorithmic exposure (what the platform shows) and social engagement (what users share) operate on distinct logics, creating fragmented information environments even within a single platform.
2. Concept of "Kawaii Toxicity": It introduces a novel framework for understanding how generative AI reshapes political communication by packaging partisan animosity in adorable, non-threatening aesthetics, effectively weaponizing cuteness for political attack.
3. Multimodal Mobilization Strategy: It distinguishes the functional roles of different media types: Human photos drive debate and accountability (high replies), while AI-generated symbols drive coordination and identity reinforcement (high reposts).
4. Censorship Evasion in Democracies: It highlights how activists in democracies (Taiwan) use linguistic and visual adaptations (homophones, puns) to navigate corporate censorship (platform moderation) rather than just state censorship.

---

5. Significance

• For Platform Governance: The findings suggest that platform algorithms may inadvertently amplify divisive content (anti-DPP) while users actively curate counter-narratives, complicating efforts to manage political discourse.
• For Social Movement Theory: The study updates theories of emotional contagion and visual mobilization, showing that in the era of Generative AI, the production cost of symbols has dropped to near zero, allowing for rapid, iterative, and highly specific visual campaigns.
• For Democratic Resilience: The Bluebird Movement illustrates how digital platforms can "bootstrap" existing constitutional remedies (like recall elections), transforming online symbolic protest into formal institutional oversight. However, the rise of "kawaii toxicity" poses a risk by normalizing the aestheticization of political vitriol.