EJAI: Judgment-Preserving Civic AI for Data Center Environmental Justice
Organizations: Department of Computer Science University of North Carolina, Chapel Hill, USA · *Society-Centered AI Lab · School of Data and Information Sciences University of North Carolina, Chapel Hill, USA
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.
Figures & tables
| Sub-Q Coverage | Faithfulness | Non-Collapse | Rebuttal Grounding | |||||
|---|---|---|---|---|---|---|---|---|
| Persona Configuration | % | ( ) | % | ( ) | % | ( ) | % | ( ) |
| No Persona (Baseline) | 100.0% | (27/27) | 98.8% | (83/84) | n/a | n/a | ||
| Environmental | 100.0% | (27/27) | 100.0% | (75/75) | n/a | n/a | ||
| Corporate | 85.2% | (23/27) | 90.9% | (70/77) | n/a | n/a | ||
| Environmental-first exchange | 100.0% | (27/27) | 96.2% | (150/156) | 91.7% | (11/12) | 100.0% | (12/12) |
| Corporate-first exchange | 96.3% | (26/27) | 95.3% | (143/150) | 75.0% | (9/12) | 100.0% | (12/12) |
| Design goal | Concern and capability | Features | Results |
|---|---|---|---|
| DG1: Couple map and AI | Distributive; engaging | Map, county overview, map-layer suggestions | Participants moved across maps, testimony, documents, and AI (P1, P15) |
| DG2: Integrate evidence | Recognitional; understanding | Testimony beside metrics, mixed corpus | P15 tied testimony to metrics; near-data codes lowest in EJAI (47) |
| DG3: Provenance and gaps | Epistemic; understanding | Citations, source labels, missing-data statements | Gaps acknowledged in 32 of 33 cases; verification codes 5, 12, 21; map metrics felt official (P1, P7) |
| DG4: Stakeholder comparison | Procedural; interacting and preparing to act | Personas, deliberation matrix | Rehearsal (P7, P9); 85.2% coverage for the Corporate Executive; conditional briefs |
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
| Company | Years | Environmental Reports | Corporate Disclosures |
|---|---|---|---|
| Apple | 2008–2025 | 18 | 0 |
| Amazon | 2018–2024 | 7 | 1 |
| 2016–2025 | 10 | 10 | |
| Meta | 2019–2021, 2023–2024 | 5 | 3 |
| Microsoft | 2003–2007, 2009–2022, 2024–2025 | 21 | 1 |
| NTT DATA | 2025 | 1 | 0 |
| Author(s) | Year | Publication Title |
|---|---|---|
| Marwah et al. ( Marwah et al., 2010 ) | 2010 | Quantifying the Sustainability Impact of Data Center Availability |
| Siddik et al. ( Siddik et al., 2021 ) | 2021 | The Environmental Footprint of Data Centers in the United States |
| Monserrate ( Monserrate, 2022 ) | 2022 | The Cloud Is Material: On the Environmental Impacts of Computation and Data Storage |
| Al Kez et al. ( Al Kez et al., 2022 ) | 2022 | Exploring the Sustainability Challenges Facing Digitalization and Internet Data Centers |
| Frenzel et al. ( Frenzel et al., 2023 ) | 2023 | Digital Infrastructures and Environmental Justice: Policies, Practices, and Visions |
| de Vries-Gao ( de Vries-Gao, 2026 ) | 2026 | The Carbon and Water Footprints of Data Centers and What This Could Mean for Artificial Intelligence |
| Publication Title | Year |
|---|---|
| Legal Tools to Advance Environmental Justice | 2022 |
| Technical Guidance for Assessing Environmental Justice in Regulatory Analysis | 2023 |
| Environmental Justice Strategic Plan | 2024 |
| EJScreen Technical Documentation v2.3 | 2024 |
| Environmental Justice and Civil Rights in Permitting FAQs | 2024 |
| County | State | Comments |
|---|---|---|
| Mecklenburg | North Carolina | 21 |
| Orange | North Carolina | 9 |
| Loudoun | Virginia | 14 |
| Organization | Publication Title |
|---|---|
| Brookings | Why Community Benefit Agreements Are Necessary for Data Centers |
| Center for Biological Diversity | How the AI Boom Threatens to Entrench Fossil Fuels and Compromise Climate Goals |
| Data Center Watch | Local Activism Threatens to Derail the U.S. Data Center Boom |
| Good Jobs First | Cloudy with a Loss of Spending Control: How Data Centers Are Endangering State Budgets |
| What to Ask When a Data Center Wants to Come to Town: A guide for public officials and community members on the most important questions to ask, long before approval | |
| Kapor Foundation | The Unequal Burden of Data Centers: An Examination of the Environmental and Public Health Impacts on Communities in California |
| Retrieval Coverage | Sub-Q Coverage | Faithfulness | Citation Accuracy | |||||
|---|---|---|---|---|---|---|---|---|
| Query Type | % | ( ) | % | ( ) | % | ( ) | % | ( ) |
| Factual | 100.0% | (9/9) | 100.0% | (9/9) | 97.7% | (43/44) | 77.3% | (34/44) |
| Multi-hop | 100.0% | (18/18) | 100.0% | (18/18) | 100.0% | (56/56) | 94.6% | (53/56) |
| Understanding | 100.0% | (6/6) | 100.0% | (9/9) | 96.4% | (27/28) | 96.4% | (27/28) |
| Overall | 100.0% | (33/33) | 100.0% | (36/36) | 98.4% | (126/128) | 89.1% | (114/128) |
| Negative Rejection Rate | ||
|---|---|---|
| Ablated Dimension | % | ( ) |
| No Testimony | 100.0% | (12/12) |
| No Documents | 91.7% | (11/12) |
| No Metrics | 100.0% | (9/9) |
| Overall | 97.0% | (32/33) |
| Configuration | Interp. share | Faithfulness | Citation acc. |
|---|---|---|---|
| No Persona (Baseline) | 7.7% | 98.8% (83/84) | 95.2% (80/84) |
| Environmental | 18.5% | 100.0% (75/75) | 88.0% (66/75) |
| Corporate | 24.5% | 90.9% (70/77) | 84.4% (65/77) |
| Environmental-first | 27.1% | 96.2% (150/156) | 81.4% (127/156) |
| Corporate-first | 25.4% | 95.3% (143/150) | 80.7% (121/150) |
| Metric | Items | Human–Human | GPT-4o–Human | Astra–Human |
|---|---|---|---|---|
| Faithfulness (supported vs. not) | 185 | 88.6% (0.86) [0.34] | 90.0% (0.89) [0.20] | 85.1% (0.81) [0.27] |
| Citation accuracy | 185 | 86.5% (0.81) [0.53] | 87.3% (0.83) [0.48] | 82.4% (0.75) [0.42] |
| Sub-Q coverage (judged rows) | 43 | 88.4% (0.86) [0.26] | 91.9% (0.91) [0.00] | 84.9% (0.82) [ 0.08] |
| Retrieval coverage | 11 | 10/11 | 19/22 | 19/22 |
| Negative rejection (gap) | 11 | 11/11 | 22/22 | 22/22 |
| Non-Collapse | 8 | 7/8 | 13/16 | 15/16 |
| Retrieval Coverage | Sub-Q Coverage | Faithfulness | Citation Accuracy | |||||
|---|---|---|---|---|---|---|---|---|
| Query Type | % | ( ) | % | ( ) | % | ( ) | % | ( ) |
| Factual | 100.0% | (9/9) | 88.9% | (8/9) | 90.9% | (40/44) | 72.7% | (32/44) |
| Multi-hop | 100.0% | (18/18) | 94.4% | (17/18) | 87.5% | (49/56) | 87.5% | (49/56) |
| Understanding | 100.0% | (6/6) | 88.9% | (8/9) | 96.4% | (27/28) | 96.4% | (27/28) |
| Overall | 100.0% | (33/33) | 91.7% | (33/36) | 90.6% | (116/128) | 84.4% | (108/128) |
| Sub-Q Coverage | Faithfulness | Non-Collapse | Rebuttal Grounding | |||||
|---|---|---|---|---|---|---|---|---|
| Persona Configuration | % | ( ) | % | ( ) | % | ( ) | % | ( ) |
| No Persona (Baseline) | 92.6% | (25/27) | 90.5% | (76/84) | n/a | n/a | ||
| Environmental | 100.0% | (27/27) | 89.3% | (67/75) | n/a | n/a | ||
| Corporate | 81.5% | (22/27) | 83.1% | (64/77) | n/a | n/a | ||
| Environmental-first exchange | 96.3% | (26/27) | 83.3% | (130/156) | 83.3% | (10/12) | 100.0% | (12/12) |
| Corporate-first exchange | 92.6% | (25/27) | 85.3% | (128/150) | 83.3% | (10/12) | 91.7% | (11/12) |
| Code | Definition |
|---|---|
| Actor-specific reasoning | User reasons from their own stake, identity, values, or lived experience |
| Third-person reasoning | User reasons from data/AI output alone, no personal stake invoked |
| Info: Far data | Institutional, quantitative, or national/general info surfaced |
| Info: Near data | Testimony, personal accounts, lived experience surfaced |
| Info: Perspective contrast | Two or more conflicting perspectives surfaced, presented by system or connected by user |
| Gap: User-Missing Knowledge | User flags own lack of domain knowledge |
| Code | Map-Only | Single-Agent | EJAI |
|---|---|---|---|
| Total coded action units | 937 | 1027 | 995 |
| Info: Far data | 148 (15.8) | 127 (12.4) | 126 (12.7) |
| Info: Near data | 74 (7.9) | 71 (6.9) | 47 (4.7) |
| Info: Perspective contrast | 5 (0.5) | 15 (1.5) | 29 (2.9) |
| Gap: User-Missing Knowledge | 33 (3.5) | 22 (2.1) | 19 (1.9) |
| Gap: User-Missing Evidence | 37 (3.9) | 17 (1.7) | 15 (1.5) |