EJAI: Judgment-Preserving Civic AI for Data Center Environmental Justice

Evan Pačes-Wiles, Bryen Alvarez, Sara Gulabani, Akshara Kolipaka, Snehalkumar 'Neil' S. Gaikwad

TL;DR

EJAI is a novel multi-agent visualization system designed to help communities navigate the complex environmental justice issues surrounding data center development. By combining interactive maps with source-grounded AI personas that "debate" retrieved evidence, the system helps users perform the heavy lifting of evidence synthesis while preserving their own civic judgment.

What Problem Does This Paper Address?

As the AI boom drives a massive expansion of data center infrastructure, local communities are increasingly caught in high-stakes disputes over water use, energy demand, and land conversion. Residents trying to participate in public discourse face a "fragmented evidence" problem: data is scattered across technical metrics, corporate sustainability reports, policy documents, and public testimony.

The authors identify a phenomenon called premature closure, where current tools fail to support critical thinking:

  • Assumption-driven closure: Users of traditional dashboards often fill in data gaps with their own assumptions.
  • Authority-driven closure: Users of standard AI chatbots tend to accept "authoritative" summaries without verifying the underlying sources.

EJAI aims to solve this by creating "judgment-preserving civic AI"—tools that handle the "evidence work" (summarizing, finding gaps, comparing views) but leave the final interpretation and policy recommendations to the human user.

Methods

The EJAI system architecture (see Figure 4) uses a sophisticated coordination pipeline:

  1. Interactive Map: A bivariate choropleth map allows users to visualize environmental and socioeconomic metrics at the county level.
  2. Source-Grounded RAG: A Retrieval-Augmented Generation (RAG) pipeline pulls from a curated knowledge base. A "Manager" agent determines if queries are on-topic, and a "Retrieval Planner" selects specific tools to fetch metrics, documents, or local testimony.
  3. Multi-Agent Personas: Users can @-mention two distinct personas—an Environmental Advocate and a Corporate Executive. These agents receive the same evidence but interpret it through different priority lenses.
  4. Deliberation Matrix: The system summarizes the exchange between personas, highlighting shared evidence, conflicting interpretations, and missing information.
  5. Missing-Data Statements: Crucially, the system is prompted to explicitly state when the retrieved evidence cannot answer a specific part of a user's query.
Figure 4: EJAI system architecture. Interface interactions and active map selections are routed through a planning layer that determines which quantitative and qualitative evidence to retrieve. Retrieved data passes to the generation and deliberation layer, which produces and summarizes source-grounded responses and stakeholder exchanges.

Datasets

The authors built a specialized knowledge database for this study:

  • Documents: 107 files, including 77 corporate environmental reports (from companies like Google, Microsoft, and Meta), 19 advocacy reports, 6 academic papers, and 5 EPA documents.
  • Public Testimony: 44 transcribed public comments from town hall meetings in North Carolina and Virginia.
  • Metrics: County-level data from the EPA’s EJScreen and tracked data center facility locations.
  • Benchmark: A newly introduced functional benchmark of 21 queries (Factual, Multi-hop, and Understanding) used to test the system's faithfulness and gap-handling capabilities.

Key Insights and Ideas

  • The "Grounding Cost" of Personas: One of the paper's most significant findings is that persona alignment comes at a cost to evidence coverage. The "Corporate Executive" persona was found to omit evidence that conflicted with its assigned priorities (e.g., ignoring high water stress metrics). However, a two-persona exchange (see Figure 5) recovered this coverage, bringing more of the total record into view for the user.
  • Maps as "Official" Truth: In user studies, participants were highly skeptical of AI-generated text but often took map metrics "at face value" because they looked official, even when those metrics lacked clear provenance.
  • AI as a "Civic Rehearsal" Tool: Rather than just providing answers, the personas helped users "rehearse" for real-world town halls by simulating the types of arguments they might encounter from different stakeholders.
  • Surfacing the "Unknowns": By explicitly labeling missing data (e.g., "The evidence lacks site-specific water use figures"), the system empowered users to ask more pointed questions of local officials.
Figure 5: EJAI personas. Environmental Advocate and Cor- porate Executive agents deliberate over shared retrieved evi- dence. A deliberation matrix summarizes shared evidence, persona interpretations, and missing information.

Key Results

The system was evaluated through both a technical benchmark and a 15-participant user study:

  • High Faithfulness: In technical tests, 98.4% of the system's factual claims were supported by retrieved evidence.
  • Effective Gap Handling: The system successfully acknowledged missing evidence in 32 out of 33 test cases (97%).
  • Persona Performance: As shown in Table 1, two-persona exchanges achieved up to 100% sub-question coverage, significantly outperforming single-persona configurations.
  • User Behavior: Participants using EJAI were more likely to verify claims and move toward "conditional recommendations" (e.g., "I would approve this only if the company discloses its grid impact") rather than simple yes/no stances.
Table 1: Persona evaluation. Evaluation of single- and dual-agent configurations on non-factual queries (𝑁= 27 sub-questions, including manual scoring for reasoning tasks). Beyond standard coverage and faithfulness, dual-agent debate branches are assessed on Non-Collapse and Rebuttal Grounding (𝑁= 12 queries). Higher is better for all metrics.

Limitations and Future Work

The authors acknowledge several critical challenges:

  • Corpus Imbalance: The knowledge base is currently skewed toward corporate reports because they are more publicly accessible than community-led data.
  • Persona Oversimplification: "Stock" personas can flatten the complex, intersectional identities of real community members.
  • The "AI Paradox": There is an inherent tension in using resource-intensive, commercial LLMs to scrutinize the environmental footprint of the very infrastructure those models depend on.
  • Future Directions: The team plans to conduct longitudinal studies with residents in affected areas and explore "co-creation," where communities help build the document corpora and define the AI personas themselves.

Summary updated

Abstract

Communities in the United States and around the world are contesting data centers, the physical infrastructure of AI. These disputes pose environmental justice issues: residents must weigh unevenly distributed burdens using evidence scattered across metrics, corporate claims, policy documents, and public testimony. A formative study with 10 users showed that dashboards and single-agent AI can settle questions before people examine that evidence. Here we present EJAI, a source-grounded system that links an interactive map, retrieved evidence, missing-data statements, and two stakeholder personas that allow people to inspect evidence and form their own policy recommendations. In a small functional benchmark of 21 queries, retrieved evidence supported 98.4% of factual claims and the system acknowledged missing evidence in 32 of 33 cases. Persona alignment carried a grounding cost: the Corporate Executive omitted some evidence that conflicted with its assigned priorities. In a within-subjects study, 15 participants acting as residents prepared for a simulated town hall. Some checked cited AI claims while treating map metrics as official, some made conditional recommendations with safeguards such as water-use disclosure, and some found the personas simplified or biased. Our results suggest that civic AI can support preparation for public reasoning when it makes evidence, gaps, and disagreement open to inspection, and that corpus imbalance and two-sided personas remain open justice concerns.
BibTeX
@misc{paceswiles2026ejai,
  author = {Evan Pačes-Wiles and Bryen Alvarez and Sara Gulabani and Akshara Kolipaka and Snehalkumar 'Neil' S. Gaikwad},
  title = {{EJAI: Judgment-Preserving Civic AI for Data Center Environmental Justice}},
  year = {2026},
  eprint = {2611.13525},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.13525},
}
Figure