What AI Speaks for Your Community: Polling AI Agents for Public Opinion on Data Center Projects

AI summary

Overview

Research area: AI safety and ethics / computing and society (cs.CY) — specifically responsible AI infrastructure deployment and the use of large language models as synthetic opinion-polling agents.

Technical level: Intermediate. The paper combines accessible survey and public-opinion concepts with LLM agent design, population synthesis (Iterative Proportional Fitting), and conformal prediction, though the conformal prediction component is described but not implemented.

Scope: This paper proposes and tests a six-stage framework that uses LLM-based AI agents, sampled to match county demographics, to approximate community opinion on proposed data center projects in two U.S. counties.

What This Paper Is About

Data centers that power AI bring jobs and tax revenue to local communities but also consume electricity and water and can strain local infrastructure, yet traditional polling is expensive and usually happens too late to shape planning decisions. The authors ask whether large language models can be used as AI agents to poll a statistically representative "virtual community" about a proposed data center, producing early, scalable, region-specific sentiment estimates. The goal is not to replace real community voices but to offer a screening tool that lets developers consider local sentiment during early-stage planning.

Key Contributions

1. A six-stage AI agent polling framework for data center public opinion, combining regional context modeling, virtual agent construction from census data, LLM polling, optional conformal prediction calibration, and multi-level analysis (cross-model, cross-regional, and human-poll comparison). The authors state this is the first application of AI agent polling to public opinion on data centers.
2. A demographic sampling method that uses Iterative Proportional Fitting (IPF) to synthesize a joint distribution of correlated characteristics from U.S. Census Bureau 2023 American Community Survey 5-year estimates, then validates the generated agent population with chi-square goodness-of-fit tests.
3. A multi-model polling experiment across three LLMs — OpenAI GPT-5 (primary model), Google Gemini-2.5-Pro, and Alibaba Qwen-Max — with 1,000 agents per model or region answering a 13-question survey about a hypothetical 100 MW data center.
4. An explicit disclosure and topical comparison with real-world polling, including an acknowledgment of legitimacy, bias, and representativeness concerns, and a qualitative comparison against a national Heatmap News survey.

Main Findings

• Baseline attitudes in Taylor County, Texas (GPT-5, n=1000): 54.2% of agents expressed neutral attitudes toward the proposed data center, while 43.6% showed positive attitudes.
• Economic perception was largely mixed: 80% of agents viewed the project's economic effects as mixed and 20% perceived them as positive.
• Environmental concern was near-universal: 97% of agents expressed they were worried about environmental impacts.
• Government oversight attitudes: 60% of agents maintained neutral attitudes toward government oversight, while 40% expressed distrust.
• Top open-ended concerns: LLM-driven topic analysis of open-text responses found the top three themes to be water resource & protection, utility costs, and local jobs & employments. The full ten-topic breakdown was: Water Resource Protection, Utility Costs, Local Jobs & Employments, Clean Energy, Economic Benefits, Transparency & Public Reporting, Accountability & Enforcement, Grid Impact & Reliability, Housing Costs, and Taxes & Public Finance.
• Model-specific divergence on economics: Qwen agents were notably more optimistic, with 91% expressing positive attitudes toward the project's economic impacts, whereas GPT-5 and Gemini-2.5 agents primarily viewed them as mixed. Qwen agents almost uniformly prioritized tax revenue and job creation, while GPT-5 and Gemini-2.5 showed more diverse preferences.
• Model-specific divergence on governance: GPT-5 and Gemini-2.5 showed substantial distrust in government regulatory capacity (40% and 32%, respectively), while nearly no Qwen agents expressed distrust. Academic research was a top information source for all models, but Qwen agents showed a much stronger preference for local government as a trusted source.
• Cross-regional difference in support: Taylor County agents showed notably higher support than Loudoun County agents (43.6% vs 9.7%). Net support was approximately +41% in Taylor County (44% support vs 2% oppose) and -5% in Loudoun County (10% support vs 15% oppose).
• Conditions for support differ by region: Environmental protection was a top consideration for nearly all agents in both counties. In Taylor County, 94% selected lower utility bills and about 51% chose local job guarantees; in Loudoun County, approximately 90% selected stricter oversight.
• Top economic worry shared across regions: Communities in both counties identified higher utility bills as their top economic concern. Public service strain was a worry unique to Loudoun County, consistent with a cited report from a Loudoun County Supervisor about stress on the local power grid and public resistance to new development.
• Topical alignment with a national human poll: Because of methodological differences, the authors state that direct quantitative comparison is not feasible. Thematically, both the Heatmap national poll and the AI agent polling identified tax revenue as a top perceived benefit and water usage plus electricity usage (leading to higher utility bills) as top concerns. The divergence: the human poll's other top benefit was job creation, while the AI agents ranked infrastructure upgrades second — which the authors attribute to their survey offering a more granular set of options.
• National versus local patterns: The Heatmap poll showed net support of approximately +2% nationally (44% support vs 42% oppose), compared with the county-level results of +41% and -5%, suggesting support is highly location-dependent.
• Polling cost and runtime: Total API costs were $36.2 per run, reduced by at least 50% by cost-optimization strategies. Model-specific costs were $23.3 for GPT-5 (2,000 agents across baseline and regional comparisons), $11.2 for Gemini-2.5 (1,000 agents), and $1.7 for Qwen (1,000 agents). Each run took more than 24 hours.

Methodology in Plain English

The researchers built a virtual version of a county's adult population and then surveyed it with LLMs.

1. Build the context. For each target county they gathered state-level data center electricity consumption from EPRI and LBNL reports, county profiles from the U.S. Census Bureau's 2023 American Community Survey 5-year API, and a standardized description of a hypothetical 100 MW data center. These three pieces together formed the fixed "global context" given to every agent.
2. Create representative agents. They pulled census marginals across social, economic, housing, and demographic characteristics (things like education, marital status, language at home, employment sector, income bracket, housing value and rent, vehicles, age, sex, race, and ethnicity). Because these were only one-dimensional distributions, they used Iterative Proportional Fitting (maximum 10 iterations, convergence threshold 10⁻⁹) to produce a joint distribution with realistic correlations, then sampled agents from it. Agents under 18 were discarded, and marital status was assigned afterward based on each agent's sex.
3. Verify the sample. Chi-square goodness-of-fit tests compared generated agent demographics against census data; p-values above 0.05 indicated the demographic structure was preserved.
4. Poll. Each agent received the static context as a system message (so the API provider could cache it) and its own demographic profile plus the questionnaire as the user prompt. The questionnaire had 13 questions — 12 single- or multiple-select items and one open-text question — spanning five domains: economic impacts, environmental concerns, community engagement, anticipated personal impacts, and overall project support. The survey ran through batch APIs, processing thousands of responses in parallel. Each experiment polled 1,000 virtual agents per model or region.
5. Analyze. The 12 multiple-choice questions were analyzed with quantitative frequency analysis; the open-text question was analyzed with LLM-driven topic analysis.
6. Calibrate (described but not done). The design includes a conformal prediction step: use a small set of real survey results to compute nonconformity scores, derive a threshold q̂ at a target confidence level α, and construct prediction intervals guaranteed to contain the true population probability with at least 1 − α probability. The authors state this step was not implemented in this preliminary study due to scope and lack of financial resources; validation instead relied on comparison with an existing national poll.

Notably, the paper reports implementing five of the six methodological stages.

Why This Matters

Impact on research. The paper opens a new application area at the intersection of LLM agent research and AI infrastructure governance: using synthetic agents not to predict elections or consumer preferences but to screen community sentiment about the physical facilities that make AI possible. It also models a transparency-first posture, with an explicit disclosure that synthetic opinions may not substitute for actual community voices and that biased training data could skew results.

Real-world applications:

• Early-stage data center siting. A developer could screen candidate counties for likely support level and for the specific conditions (environmental protections, lower utility bills, job guarantees, stricter oversight) that residents would want addressed.
• Pre-hearing community engagement. Instead of waiting for a formal public hearing after commitments are made, planners could use agent polling to anticipate concerns such as water consumption and utility costs before design decisions harden.
• Regulatory and municipal planning support. Local governments could use it to compare a proposed project's profile against the concerns their community is likely to raise.
• Extending to other infrastructure. The framework is presented as applicable beyond data centers, as the authors list other infrastructure domains as future work.

Industry relevance. The reported API cost of $36.2 per run and a runtime exceeding 24 hours suggest a technique that is cheap relative to traditional polling infrastructure, and the cross-model results warn that the choice of LLM materially changes the answers — Qwen agents were far more economically optimistic and far more trusting of government than GPT-5 or Gemini-2.5 agents. For any organization considering this approach, that means the model itself is a variable in the result.

Future Directions

• Implement the conformal prediction calibration with real-world survey data, which the current study describes but does not execute, in order to attach formal statistical guarantees to agent polling predictions.
• Extend the framework to other infrastructure domains beyond data centers.
• Develop site selection algorithms that incorporate community sentiment as an input.
• Simulate human deliberative processes and social interactions using multi-agent systems, and establish hybrid approaches that combine AI polling with traditional community engagement methods.

There are also open questions the paper raises about limitations: whether the approach generalizes to counties without existing data center infrastructure (both studied counties already host data centers), how to capture perspectives the agents may miss — such as neurodiverse perspectives or "engagement silence" — and how to handle the inherent biases in LLMs, including racial biases cited from prior work.

Target Audience

This paper is most useful to researchers and practitioners working on responsible AI infrastructure, AI governance, and computational social science. It will also be valuable to data center developers, utility and municipal planners, and policymakers who want to understand community sentiment earlier in the planning process, and to LLM researchers interested in the fairness and model-dependence of synthetic polling. Readers without a technical background can follow the main findings, but the methodology sections on IPF, chi-square verification, and conformal prediction assume some quantitative familiarity.

Authors’ abstract

The intense computational demands of AI, especially large foundation models, are driving a global boom in data centers. These facilities bring both tangible benefits and potential environmental burdens to local communities. However, the planning processes for data centers often fail to proactively integrate local public opinion in advance, largely because traditional polling is expensive or is conducted too late to influence decisions. To address this gap, we introduce an AI agent polling framework, leveraging large language models to assess community opinion on data centers and guide responsible development of AI. Our experiments reveal water consumption and utility costs as primary concerns, while tax revenue is a key perceived benefit. Furthermore, our cross-model and cross-regional analyses show opinions vary significantly by LLM and regional context. Finally, agent responses show strong topical alignment with real-world survey data. Our framework can serve as a scalable screening tool, enabling developers to integrate community sentiment into early-stage planning for a more informed and socially responsible AI infrastructure deployment.