TL;DR
This study leverages open-source Large Language Models (LLMs) to mine a decade of Reddit discourse, revealing a widening "infrastructure divide" in public sentiment toward data center development. The researchers found that while opposition significantly outpaces support in both volume and engagement, the most provocative political framing drives visibility, while critical technical concerns regarding water and power remain buried in lower-level comment threads.
What Problem Does This Paper Address?
The rapid expansion of AI has triggered a massive wave of data center construction, yet this physical growth is hitting a wall of local resistance. Between May 2024 and March 2025, an estimated $64 billion in U.S. data center projects were blocked or delayed by community opposition.
The authors argue that traditional methods of gauging public opinion—like surveys and public hearings—are often "too little, too late," occurring only after plans are finalized. Furthermore, these methods don't scale well. There is a critical need for a systematic, real-time way to understand community concerns before they lead to project failure. This paper addresses that gap by using social media mining to characterize the evolution and drivers of public stance at scale.
Methods
The researchers developed a four-stage analytical pipeline:
- Data Collection & Filtration: They used the Reddit API to gather a decade of posts and comments, filtering for relevance and engagement.
- LLM-Powered Stance Detection: They compared four open-source LLMs (DeepSeek v3.1, DeepSeek R1, Llama 3, and Llama 4) to classify posts as Favor, Against, or Neutral. To improve accuracy, they used a Chain-of-Stance prompting framework, which forces the model to follow a six-step reasoning process before assigning a label.
- Engagement & Temporal Analysis: The team used statistical tests (like the Brunner–Munzel test) to compare engagement metrics (upvotes, comments) across different stances and time-series analysis to see if the "divide" was growing.
- Topic Modeling: They applied BERTopic to the explanations generated by the LLM rather than the raw Reddit text. This provided cleaner, more semantically rich data for clustering arguments.
Datasets
The study utilized a custom-built corpus of Reddit data:
- Initial Collection: 1,257 posts and 149,621 comments.
- Final Filtered Dataset: 798 posts and 102,328 associated comments spanning over 10 years.
- Validation: To ensure Reddit was a valid proxy for broader public interest, the authors compared Reddit activity volume against Google Trends search data for "data center," finding a strong alignment in the recent surge of interest.
Key Insights and Ideas
- The "Loudness" of Opposition: Opposition isn't just more common; it's much louder. Posts against data centers outnumbered favorable ones 2.6:1, but they generated 8.5 times more comments.
- The Visibility vs. Substance Gap: There is a clear divide in what drives engagement. The most "viral" posts (highest upvotes) relied on provocative political framing (e.g., calls for moratoriums). However, the actual substance of community discussion—technical worries about water usage and power grid strain—was dispersed across hundreds of lower-engagement comment threads.
- LLMs as "Human-Level" Annotators: By using domain-specific rules and reasoning chains, the open-source DeepSeek v3.1 model achieved 83% of the human inter-reviewer agreement ceiling. This suggests that researchers can monitor public sentiment without relying on expensive proprietary models or manual surveys.
- A Widening Divide: The gap between opposition and support is not just a temporary spike; time-series analysis confirmed it is a persistent, non-stationary trend that has accelerated since 2025.
Key Results
- Stance Distribution: 57% of posts were neutral/informational, 31% were against, and only 12% were in favor.
- Opposition Clusters: Topic modeling resolved opposition into four main pillars:
- Resource consumption and scarcity (42.8%): Concerns over water and electricity being "unfairly allocated" at the community's expense.
- Contested siting and secrecy (34.2%): Distrust in the approval process and lack of transparency.
- Political opposition (13%): High-level calls for construction pauses.
- Corporate accountability (8%): Questions about who bears the long-term infrastructure costs.
- Engagement Metrics: Against-posts attracted significantly higher upvote scores (Cliff’s $\delta = -0.62$, a large effect) compared to favor-posts.
Limitations and Future Work
The authors acknowledge that Reddit users are not a representative sample of the general population, which may skew the data toward certain demographics. Additionally, labeling a whole post with a single stance can oversimplify the diverse opinions found within a single comment thread.
Future research could involve:
- Geographic Linking: Connecting social media discourse to specific counties or utility service areas to see if online chatter predicts formal legal opposition.
- Comment-Level Modeling: Analyzing individual comments to capture the nuance that post-level labeling misses.
- Policy Impact: Testing whether increased transparency measures (like publishing incentive agreements early) actually shift community stance from opposition to partnership.
Summary updated