Silencing the Silencing: The Uneven Amplification of Internet Shutdowns in India's English-Language Press
Revise & resubmitPolitical Communication
Abstract
India imposes more internet shutdowns than any other country, yet whether its national press places these events before readers has been measured only for single outlets, leaving open whether the silence belongs to the press or to a newspaper. I measure per-event coverage of 556 state-level shutdowns (2013–2024) across three English-language national dailies: the Times of India (TOI), India's largest, through a 1.5 million-article full-text corpus, and The Hindu and The Indian Express, titles editorially more adversarial to the current government, through 3.3 million archive headlines and 32,014 targeted full texts, all classified under one validated rule. In TOI, shutdowns in Jammu and Kashmir receive roughly 2.7 times fewer articles than comparable shutdowns elsewhere. The gap does not replicate: The Hindu shows no detectable J&K deficit, the Express less than half of TOI's, with both interaction estimates significantly above TOI under identical text exposure. During the first seven months of the 2019–20 Kashmir blackout, the longest ever imposed in a democracy, TOI published 23 shutdown articles against 102 in the Express and 118 in The Hindu. Because the outlets face the same reporting conditions, the divergence cannot reflect newsgathering constraints: the silence is a property of one outlet's editorial choices, not of the press field. Whether alignment or audience demand motivates those choices, the accountability reach of a formally free press depends on which outlet dominates its audience, a stratification that indices scoring a press system as a single unit do not register.
Electoral Accountability and the Channeling of Content Removal: Theory and Evidence from Google Transparency Reports
Under reviewwith P. Zhang
Abstract
How does regime type shape a state's intervention in its information environment? We argue that democracies and non-democracies differ not only in the quantity but, more importantly, in the institutional channel of content removal. To that end, we develop a political agency model and show that a democratic government would refrain from direct executive takedown and instead delegate to the court. This is because electoral accountability disciplines politicians' behavior by providing incentives for reputation-building. We test our hypotheses using data from Google transparency reports. Exploiting quasi-experimental variation in the timing of constitutionally scheduled elections, we provide supporting evidence that compared to non-democracies, takedown requests from democratic governments decline significantly as elections approach. This reputational discipline effect is not observed in authoritarian regimes or other types of requests.
Build, Borrow, or Just Fine-Tune? Empirical Guidelines for NLP Model Selection in Computational Social Science
Under review
Abstract
Researchers in computational social science increasingly face a consequential choice when adopting natural language processing tools: build a domain-specific model from scratch, borrow and adapt an existing one, or simply fine-tune a general-purpose model on task data? Each approach occupies a different point on the spectrum of performance, cost, and required expertise, yet the discipline has offered little empirical guidance on how to navigate this trade-off. This paper provides such guidance. Using conflict event classification as a test case, I fine-tune ModernBERT on the Global Terrorism Database (GTD) to create Confli-mBERT and systematically compare it against ConfliBERT, a domain-specific pretrained model that represents the current gold standard. Confli-mBERT achieves 75.46% accuracy compared to ConfliBERT's 79.34%. Critically, the four-percentage-point gap is not uniform: on high-frequency attack types such as Bombing/Explosion and Kidnapping, the models are nearly indistinguishable. Performance differences concentrate in rare event categories comprising fewer than 2% of all incidents. I use these findings to develop a practical decision framework applicable to any NLP-assisted classification task: when does the research question demand a specialized model, and when does an accessible fine-tuned alternative suffice? The marginal value of domain-specific pretraining, I show, is a predictable function of class prevalence, error tolerance, and available resources. The model, training code, and data are publicly available on Hugging Face.
“Why Aren't We Talking About This?” The Attention Trap in Networked Protest
Under review
Abstract
On 2 February 2021, the musician Rihanna tweeted about the Indian farmer protests to her 100 million followers, producing the largest attention spike in the movement's yearlong history. Using 1.02 million tweets classified by a large language model, I show that the intervention displaced the movement's own discourse rather than diluting it with newcomers. Benchmarked against 277 placebo dates, the celebrity share of conversation rose 33 percentage points while the policy share fell 8 points, and a within-user decomposition attributes most of the shift to established participants changing their own speech. The shift carried costs: participants who pivoted hardest toward celebrity content were half as likely to remain active a month later, and a coordinated pro-government campaign penetrated the movement's hashtags for the first time. Comparison with six domestic events shows the pattern is specific to the celebrity shock. I call this dynamic the attention trap.
Using Messy Text in Future LLM Annotations
Under reviewwith P.T. Brandt, P. Cuellar-Cuellar & X. Zhao
Abstract
When working with real-world text, researchers often inherit corpora and annotations with costly human judgments but were not collected, organized, or annotated for modern text-as-data workflows. This is especially common in social science domains where web-scraped documents are noisy, incomplete, or misaligned with the unit of analysis. We propose codebook attention: a codebook-guided turn-by-turn extractive summarization pipeline that reuses valuable annotations without requiring full re-annotation. For each victim in a corpus of disappearance reports from Mexico, the model processes documents sequentially, extracts verbatim evidence spans for each codebook category, and updates a running structured summary to produce a clean and evidence-traceable document-level representation. Using the codebook as the task schema focuses model attention on the fields that matter for downstream classification and reconnects prior human annotations to contemporary language model workflows. We evaluate the resulting summaries by classifying and comparing the predicted labels with human annotations across multiple open-source models including Llama 3.1, Gemma 3 and Ministral 3. Results suggest that hours of computation can approximate weeks of human coding effort while producing evidence-traceable summaries suitable for downstream human rights and event data research.