AI-Driven Voter Mimicry: Feasibility, Existing Initiatives, and Implications for Citizen Representation

I. Introduction: The Vision of AI Voter Mimicry and Citizen Representation

A. Addressing the Core Interest in AI Voter Agents

The prospect of developing artificial intelligence (AI) agents capable of faithfully mimicking individual voters—predicting their choices on political candidates, ballot measures, and other civic decisions with high accuracy—has captured significant interest. This exploration extends to the potential for such AI agents to eventually serve as a form of citizen representation. The allure of this technology is multifaceted, hinting at possibilities for a more nuanced understanding of public opinion, potentially leading to more responsive governance, or even pioneering novel forms of democratic participation.

B. Overview of the Current Technological Frontier and Socio-Political Context

Rapid advancements in AI, particularly in the domains of Large Language Models (LLMs) and agent-based modeling (ABM), have propelled concepts once confined to speculative fiction towards the realm of technical exploration.¹ This technological surge is occurring within a broader context of evolving democratic practices, ongoing challenges to traditional modes of representation ⁴, and the pervasive digitalization of civic life.⁵ This report aims to critically examine the technical feasibility of creating “digital twin voters,” survey existing efforts in this domain, and delve into the profound ethical, societal, and democratic implications that such an endeavor entails.

A fundamental consideration from the outset is the very nature of “representation” implied. The notion of an AI faithfully mimicking a voter encompasses at least two distinct ideas. Firstly, there is predictive representation, where the AI models voter choices to understand likely public responses or electoral outcomes. This aligns with established goals in social science and AI, where systems are built to forecast behavior.¹ Secondly, and more radically, there is the idea of delegative representation, where an AI agent might act on behalf of a citizen, perhaps even casting a preliminary or advisory vote, as conceptualized in some “Digital Twin Citizen” models.⁸ These two forms of representation, while linked by the aspiration of a “faithful mimic,” possess different technical prerequisites and carry vastly different ethical weights. An AI that excels at prediction may not be suitable, nor ethically justifiable, to act as a delegate. This report will navigate both facets, underscoring that proficiency in one does not automatically confer legitimacy for the other.

Furthermore, the ambition for AI to “faithfully mimic” voters to a standard of near-perfect accuracy immediately confronts the well-documented complexity, nuance, and inherent unpredictability of human political behavior. Voter decisions are not merely transactional calculations but are shaped by a confluence of deeply held values, emotional responses, cognitive biases, social influences, and the principle of bounded rationality, where individuals make decisions with limited information and cognitive resources.⁹ While contemporary AI, especially LLMs, can process and generate human-like language and identify complex patterns, capturing the full spectrum of human political reasoning—including implicit biases or rapidly shifting opinions not yet reflected in accessible data—remains a formidable challenge.¹ The pursuit of perfect mimicry might inadvertently lead to an oversimplification of the voter, focusing on easily quantifiable characteristics while missing deeper, less articulated motivations. This inherent tension between the ideal of faithful digital replication and the rich tapestry of human political identity will be a recurring theme throughout this analysis.

C. Report Roadmap

This report will proceed as follows: Section II will survey the current landscape of AI-powered voter simulation and analysis projects. Section III will delve into the specific concept of the “Digital Twin Citizen” and related ideas of AI as a voter proxy. Section IV will critically assess the feasibility of achieving near-perfect accuracy in predicting individual voter choices. Section V will explore the extensive ethical, societal, and democratic implications. Section VI will offer strategic insights for any project aiming to develop AI voter agents responsibly. Finally, Section VII will provide concluding thoughts on charting a responsible path for AI in democratic processes.

II. Current Landscape: AI-Powered Voter Simulation and Analysis Projects

A. Overview of AI in Political Simulation

The application of artificial intelligence in political science has expanded significantly, with LLMs and Agent-Based Modeling (ABM) being increasingly employed for tasks such as election prediction, public sentiment analysis, and the simulation of political behavior dynamics.¹ It is important to distinguish between general election forecasting, which typically aims to predict aggregate outcomes like vote shares or election winners ¹¹, and voter simulation. The latter, which is more aligned with the objective of creating AI voter mimics, focuses on modeling the decision-making processes of individuals or specific voter cohorts.¹ This section examines prominent projects and approaches in this burgeoning field.

A notable trend is the methodological evolution from traditional ABM, which often relied on explicitly programmed rules and statistical models to govern agent behavior, towards LLM-driven agents. These newer models learn behaviors by ingesting and processing vast quantities of data, predominantly textual information sourced from platforms like social media. Traditional ABM has long been a “mainstream method” for election simulation, utilizing “heuristic-like rules or mathematics functions” to define agent actions.¹ In contrast, newer initiatives like ElectionSim are explicitly “based on large language models” and leverage extensive social media datasets.¹ While LLMs offer the potential for richer behavioral modeling due to their capacity for nuanced language understanding and generation, they also introduce new challenges, such as the potential for inherited data biases and a lack of transparency compared to rule-based systems. This shift represents an attempt to capture the complexity of individual behavior through data-driven learning rather than predefined heuristics.

The projects reviewed also exhibit diverse ultimate objectives. Some, like ElectionSim and the framework proposed by Zhao et al., concentrate on predicting voter choices to enhance understanding and forecasting capabilities. Others, such as the services offered by companies like Resonate, focus on prediction for targeted influence in political campaigns. A third category, exemplified by AI-GOV’s “Olivia,” positions AI as an active participant or actor within the political sphere. These differing goals entail distinct ethical considerations and metrics for success. An AI adept at academic prediction might not be suitable or ethically sound for direct political action, and the converse is also true. The aspiration to “represent the citizen” leans towards the AI-as-actor model, but achieving this responsibly likely necessitates the predictive accuracy sought by academic projects, coupled with robust ethical safeguards.

A critical observation across several prominent LLM-based voter simulation projects is their heavy reliance on social media data for constructing voter profiles and inferring characteristics. For instance, ElectionSim utilizes a “million-level voter pool sampled from social media platforms” ¹, and AI-GOV draws upon “real-time chatter from X and Telegram”.¹⁶ This reliance is significant because social media users do not perfectly mirror the entire electorate demographically or psychographically. Certain population segments may be overrepresented or underrepresented, and online discourse can often be more polarized or performative than offline opinions. While data processing techniques such as language filtering and user cleaning are employed ¹, the foundational data source inherently carries the biases of the platform and its user base. Consequently, AI “digital twins” constructed primarily from such data may be more adept at mimicking a voter’s “online persona” rather than their complete, offline self, potentially skewing predictions and any subsequent attempts at “representation.” This highlights a crucial challenge in data representativeness for achieving truly “faithful mimicry” of the broader electorate.

B. Specific Project Examinations

1. ElectionSim

• Description: ElectionSim is presented as an innovative election simulation framework powered by LLMs. It is designed to facilitate accurate voter simulations, support customized demographic distributions, and includes an interactive platform enabling users to dialogue with these simulated voters.¹
• Methodology: The framework processes raw data, including user information and historical activities like posts and comments, collected from social media platforms. This data undergoes user-level aggregation, followed by demographic tagging performed by specialized classifiers. Data preprocessing includes language filtering (to focus on American users for U.S. election contexts), filtering users based on activity levels (e.g., retaining users with over 30 posts and sampling 30 of them), and user cleaning to assess content repeatability and remove duplicates.¹
• Data Sources: The core data source is a “million-level voter pool sampled from social media platforms”.¹ The demographic classifiers are trained using a semi-automated method involving annotation by commercial LLM APIs (such as GPT-4o, Claude-3.5, and Gemini-1.5) followed by manual verification by human annotators.¹⁵
• Stated Objectives & Performance: The primary aims are to model the preferences of specific demographic groups, forecast potential real-world political trends, and support accurate individual-level simulations.¹ In voter-level simulations for vote-related tasks, ElectionSim reports achieving a Micro-F1 score of 0.812.¹ While a significant benchmark in this complex domain, this level of accuracy is considerably distant from the “almost 100%” correctness envisioned in some ideal scenarios.
• Relevance: ElectionSim directly addresses the interest in projects building AI agents to mimic voters and predict their choices. The interactive dialogue feature is particularly pertinent to the idea of AI “representing” citizens by allowing engagement with simulated voter personas.

2. Theory-Driven, Multi-Step Reasoning Framework (Zhao et al.)

• Description: This research introduces a framework that integrates demographic, temporal, and ideological factors to simulate voter decision-making at scale. It utilizes synthetic personas that are calibrated against real-world voter data.¹⁸
• Methodology: The framework employs LLMs within a multi-step prompting pipeline. This process first infers a persona’s political ideology based on demographic information and party policy agendas. In the second step, this inferred ideology is integrated into an extended persona, which, along with contextual information (current year, party agendas, candidate backgrounds), is used to simulate voting preference. This structured decomposition aims to mitigate model biases and enhance predictive accuracy.²⁰
• Data Sources: The system uses synthetic personas generated via the Sync framework, which reconstructs individual profiles from aggregated public datasets. These personas are aligned with real-world voter data from sources like the American National Election Studies (ANES). Temporal factors, such as evolving candidate policy agendas, are also incorporated.²⁰
• Stated Objectives & Performance: The goals include improving simulation accuracy, mitigating LLM biases inherited from training data, providing a scalable framework for modeling political decision-making, and offering insights into the capabilities and limitations of LLMs in political science research. The framework aims to approximate average voter behavior and is validated against real-world state- and county-level election data.²⁰
• Relevance: This project offers an alternative approach that emphasizes theoretical grounding and structured reasoning to improve the robustness and reduce bias in LLM-based voter simulations, addressing key challenges in creating faithful voter mimics.

3. AI-GOV and “Olivia”

• Description: AI-GOV is a political agent framework featuring “Olivia,” an autonomous AI agent designed to inform policy-making using hard data and community voices. Launched in 2025 by individuals described as libertarians and ex-politicians, it aims to operate free of lobbyists and partisan influence, emphasizing open reasoning and accountability.¹⁶
• Methodology: Olivia collects real-time citizen input from social media platforms like X (formerly Twitter) and Telegram. Advanced algorithms are used to process this data, validate claims with external evidence, and transform raw input into actionable policy proposals. These proposals are then tested against real-world statistics and refined based on user feedback. Olivia also engages in advocacy by submitting policies to Congress and rallying public support online.¹⁶
• Data Sources: Data includes real-time social media conversations, alongside information from official sources such as Census.gov, Department of Justice reports, and live X trends.¹⁶
• Stated Objectives & Performance: The overarching goal is to transform governance through AI, combat corruption, and promote transparency. Olivia is designed to engage directly with the public through one-on-one chats, group discussions, and live appearances on platforms like Twitter Spaces and podcasts.¹⁶ Success is currently framed by engagement metrics (e.g., over 200,000 engagements from more than 40,000 unique individuals on X within a week of activity) rather than predictive accuracy of individual voter choices.¹⁶
• Relevance: While AI-GOV’s primary focus is on policy generation and advocacy rather than individual voter choice prediction, Olivia embodies an AI agent actively participating in political discourse and attempting to “represent” a collective voice. This aligns with the facet of the user’s query concerning AI agents that can represent citizens.

4. Agent-Based Modeling (ABM) in Elections

• Description: ABM is a well-established computational social science methodology. In ABM, autonomous “agents” are endowed with specific attributes (e.g., preferences, demographics) and rules of behavior. These agents interact within a simulated environment, and their collective actions lead to emergent macro-level phenomena, such as election outcomes or shifts in public opinion.¹⁴ ElectionSim acknowledges ABM as a mainstream method for election simulation but also points to the limitations of traditional ABMs, and even LLMs, in having sufficient personalized input to simulate the nuanced behavior of individuals.¹
• Methodology: ABM involves defining agent properties (e.g., political preferences, emotional states), rules governing their interactions (e.g., influence from neighbors, exposure to media messages), and the characteristics of the environment in which they operate. Simulations can run synchronously (all agents update simultaneously) or asynchronously (agents update in a sequence), which can affect outcomes.¹⁴ An example is the model by Kollman, Miller, and Page on party competition.¹⁴ More recently, Kelter et al. have used ABM to investigate Quadratic Voting (QV) by modeling more realistic, less-than-perfectly rational voting behaviors.²¹
• Data Sources: ABMs can draw on a variety of data sources, ranging from theoretical distributions of agent characteristics to empirical data derived from surveys or demographic records, which inform the initial setup of the agents and their environment.¹⁴
• Stated Objectives & Performance: The main purpose of ABM in this context is to understand how micro-level specifications (individual agent behaviors and interactions) can generate macro-level political outcomes. It allows researchers to observe patterns emerging from these interactions and explore the co-evolution of multiple processes, such as voter turnout and ideological preference shifts.¹⁴ The accuracy and validity of an ABM depend heavily on the soundness of its theoretical assumptions and the quality of its input data.
• Relevance: ABM provides important historical and methodological context for the newer LLM-based approaches to voter simulation. It highlights the long-standing challenge of accurately modeling individual nuances, a challenge that LLMs are now attempting to address with their advanced language processing capabilities.

C. Other Relevant AI Applications in Politics

Beyond direct voter simulation, AI is being applied in related areas of political analysis:

• Predictive Analytics for Voter Targeting (e.g., Resonate): Companies like Resonate utilize AI to analyze vast datasets encompassing voter psychographics, preferences, behaviors, demographics, and media consumption patterns. The goal is to help political campaigns identify, analyze, target, and communicate with voters more effectively, often by predicting what issues or messages will resonate most strongly with specific segments.⁷ While this application is geared towards influencing or persuading voters rather than creating a “digital twin” for neutral representation, it demonstrates sophisticated AI use in understanding and predicting voter segments.
• AI in Political Polling: AI techniques are increasingly used to process large datasets, identify subtle patterns in voter behavior, analyze social media sentiment, and potentially enhance the accuracy of traditional political polling methods.⁶ This application is more focused on aggregate predictions and understanding broad public opinion trends rather than individual-level simulation.

Comparative Overview of Key AI Voter/Citizen Modeling Projects

To provide a clearer picture of the diverse efforts in this domain, the following table summarizes key characteristics of the discussed projects and approaches.

This comparative overview illustrates the varied methodologies, data dependencies, objectives, and notions of success across different projects. It underscores that while the goal of “mimicking” voters is shared to some extent, the interpretation and application of this concept diverge significantly, impacting the relevance of each approach to the specific aim of creating highly accurate, representative digital twins of individual citizens.

III. The “Digital Twin Citizen”: Exploring AI as a Voter Proxy

The concept of a “digital twin” – a dynamic virtual representation of a physical entity or process, continuously updated with real-time data ²² – has primarily been applied in engineering and medicine. However, its extension to human beings, particularly in the context of civic participation, has given rise to notions like the “Digital Twin Citizen.” This section explores these ideas, their proposed functionalities, and their connection to broader concepts of AI-enhanced representation and liquid democracy.

The fundamental idea of a digital twin involves a physical entity, its digital representation, and a continuous data feedback loop connecting them, enabling the digital version to simulate and predict the behavior of its physical counterpart.²² While creating a digital twin for a jet engine or a human heart ⁸ is complex, applying this to the multifaceted and evolving nature of a citizen’s political identity presents a qualitatively different order of challenge. Nevertheless, projects like Twin4Dem, which simulates political systems to understand democratic backsliding ⁸, indicate a move towards applying such modeling to socio-political realms.

A. Conceptualizing the Digital Twin Citizen

Gersbach & Martinelli’s “Digital Twin Citizen”

A prominent proposal comes from macroeconomics professor Hans Gersbach and his colleague César Martinelli. They envision that every citizen would possess a personal AI twin.8 This Digital Twin Citizen would be designed to learn the political views and ethical values of its real-life human counterpart over an extended period.

• Proposed Mechanism: The core function of these twins would be to participate in a novel two-round referendum system. In the first round, the digital twins would cast preliminary, non-binding votes based on their learned understanding of their human’s preferences and extensive factual information. The results of this AI vote would then be published and openly discussed, serving as an indicator of informed public opinion. In the second round, real-life citizens would cast their own binding digital votes, having had the opportunity to consider the AI twins’ collective decision and the ensuing public debate. Voters could then adjust the preliminary choice made by their digital twin.⁸
• Information Sources: These digital twins would ideally obtain information from the same sources their human counterparts consult. Martinelli suggests that the AI could even be trained to detect misinformation, though this capability is not guaranteed.⁸
• Intended Purpose: The primary goals are to ease the cognitive load on citizens, particularly in jurisdictions with frequent referendums (like Switzerland or California) where individuals may struggle to stay informed on numerous complex issues. By providing an informed preliminary vote, the system aims to foster more considered decisions. It is also hoped that such a system could “save local democracy” and enhance the global attractiveness of direct democracy.⁸ Crucially, proponents stress that humans would retain the final decision-making power, allowing for emotions and information not considered by the AI to be factored in.⁸

Cesar Hidalgo’s Vision

Taking the concept of AI representation further, Cesar Hidalgo, a professor at the University of Toulouse, has proposed the more radical idea of replacing politicians entirely with digital twins of citizens.8 This vision implies a much deeper level of delegated authority to AI agents.

These varying conceptualizations highlight a spectrum of potential “representation.” Gersbach & Martinelli’s model focuses on informed input and deliberation enhancement, with the AI serving an advisory role. Hidalgo’s idea, and some interpretations of AI in liquid democracy, lean towards delegated authority, where AI might make binding decisions. This distinction is critical, as the technical and ethical feasibility shifts dramatically depending on where on this spectrum a specific project aims to operate.

A core challenge for any “Digital Twin Citizen” concept, particularly one aspiring to the technical definition of a digital twin, is the “data feed” requirement. True digital twins rely on a “continuous data feedback loop” from their physical counterparts.²² For a voter, this would necessitate a constant, comprehensive, and inevitably privacy-invasive stream of data encompassing their evolving beliefs, values, information consumption habits, emotional states, private discussions, and changing life circumstances. Collecting such granular data continuously for every citizen presents immense technical, logistical, and, most critically, ethical and privacy challenges.²² Current projects, like ElectionSim, which use snapshots of social media activity ¹, fall far short of this continuous, holistic data stream. This practical constraint severely limits how “faithful” and “current” a voter twin can be in the truest sense of the term, and it raises privacy concerns that far exceed those associated with twinning inanimate objects or even biological organs.

B. “AI Voice” for Enhanced Representation

Another perspective on AI in citizen representation is offered by Eduardo Albrecht, whose work on “Political Automation” explores how AI agents can help incorporate marginalized communities into global governance frameworks.²⁵ This involves the concept of “AI Voice,” attributed to Scheuermann and Aristidou, where generative AI creates virtual representations that allow traditionally unheard stakeholders to “participate” in critical discussions on topics like climate action or sustainable development.²⁵

• Mechanism and Advantages: AI agents can engage in dialogue about hypothetical policy scenarios, offering insights into how marginalized communities might respond before policies are implemented. They can also support rapid needs assessments in crisis zones or facilitate virtual dialogues between AI representatives of conflicting groups to test potential policies.²⁵ The “Ask Amina/Ask Abdalla” project, representing refugee and combatant experiences, exemplifies this approach by combining 3D avatars, LLMs, and curated knowledge bases.²⁵
• Key Caveat: A crucial principle is that AI agents should augment rather than replace direct human participation. They should serve as bridges to inclusion where traditional participation faces barriers. Real citizens must maintain ownership over their digital counterparts and be able to adjust them to reflect their values accurately.²⁵ This approach focuses on enhancing presence and sharing perspectives, rather than direct voting substitution for all citizens.

C. Connection to Liquid Democracy

The idea of AI voter mimics intersects naturally with concepts of liquid democracy. Liquid democracy is a model of collective decision-making that combines elements of direct democracy (citizens voting on issues themselves) and representative democracy (citizens delegating their vote to a trusted proxy).²⁶

• AI’s Potential Role: Some libertarian perspectives on AI governance envision AI and Decentralized Autonomous Organizations (DAOs) potentially replacing traditional governance structures, prioritizing efficiency and individual freedom.⁴ In such a framework, a Digital Twin Citizen could be seen as a highly sophisticated, personalized delegate.
• Concerns: This vision raises significant concerns regarding fairness and accountability if AI decisions are driven by private interests, the potential for algorithms to operate without democratic oversight, and the risk of weakening democratic sovereignty.²⁷

The concept of the Digital Twin Citizen, while aiming to enhance informed decision-making by offloading cognitive burdens and providing AI-reasoned preliminary stances ⁸, faces a potential paradox. There is a risk that individuals might become overly reliant on their AI twin’s recommendations or effectively outsource their critical thinking.⁸ If the AI’s output is highly persuasive, or if citizens become complacent, the human “final decision” could devolve into a mere rubber-stamping exercise. This would undermine the very goal of genuine engagement and critical deliberation that the system intends to foster. The success of such AI-assisted democratic tools would therefore hinge on carefully designing them to encourage active human deliberation after receiving the AI’s input, fostering a collaborative human-AI process rather than passive delegation.

IV. The Feasibility of “Almost 100% Accuracy” in Voter Prediction

The aspiration to create AI agents that can predict a voter’s choice with “almost 100% accuracy” is a central element of the user’s query. Achieving such fidelity requires a profound understanding of human voter behavior and AI capabilities that can comprehensively model it. This section examines the nature of voter behavior, current AI predictive performance, and the fundamental challenges to attaining near-perfect individual-level prediction.

A. The Nature of Human Voter Behavior

Human political behavior, while exhibiting patterns, is extraordinarily complex and only partially predictable.

• Complexity and Probabilistic Prediction: While individuals often act in predictable ways given certain conditions, and political scientists use data like age, education, and ideology to forecast voting likelihood and party preference, these predictions are inherently probabilistic. They indicate likelihoods for groups with certain characteristics, not deterministic outcomes for specific individuals.²⁸
• Bounded Rationality and Cognitive Biases: Voters rarely engage in exhaustive information processing. Their decisions are often constrained by cognitive biases, emotional responses, and the immediate environment, a concept known as bounded rationality.⁹ While explicit political attitudes are predictive of choices, implicit preferences, subconscious influences, and heuristic shortcuts also play significant, sometimes unacknowledged, roles.⁹
• Dynamic and Contextual Influences: Voter preferences are not static. They are swayed by a multitude of dynamic factors, including prevailing economic conditions, the perceived popularity and persona of candidates, specific campaign events, influential media narratives ⁶, and the social influence of peers and significant others.⁹ Even the overall voter turnout in an election is a complex phenomenon influenced by numerous variables and is challenging to predict accurately.²⁹
• Data Quality Challenges: The data used to model voter behavior is often imperfect. Survey nonresponse, selection biases in samples, and the tendency for respondents to inaccurately disclose their true preferences due to factors like the “spiral of silence” or social desirability bias can significantly limit the quality and reliability of data available for prediction models.¹⁰

B. Current AI Capabilities in Prediction

Current AI systems have made strides in political analysis, but individual-level prediction with near-perfect accuracy remains elusive.

• ElectionSim’s Performance: The ElectionSim project, which uses LLMs to simulate voters based on social media data, achieved a Micro-F1 score of 0.812 for vote-related tasks.¹ While this represents a strong performance for a complex social science prediction task, it is substantially below a 99-100% accuracy threshold.
• LLMs in Political Science: LLMs have demonstrated capabilities in simulating human behavior and processing large volumes of text to identify patterns and sentiments.² However, they are also prone to limitations such as generating incorrect or outdated information and can reflect biases present in their training data if not carefully augmented with external knowledge and validation.²
• Predictive Analytics in Campaigns: Companies like Resonate use AI to analyze voter data and claim to help campaigns predict what matters to voters, reporting metrics such as campaigns being “31% More likely to win when targeting with Resonate data”.⁷ These claims typically relate to the efficacy of targeted messaging for influencing segments of the electorate, rather than achieving perfect predictive accuracy for individual choices.
• Web Tracking Data for Attitude Prediction: Academic research investigating the use of web browsing data to predict political attitudes has found only a limited ability to do so. Predictions based on website visit histories tend to be modest in accuracy and dependent on the specific models and parameters used.³⁰ Lifestyle choices inferred from such data are generally not strong predictors of nuanced political views.

C. Fundamental Challenges to Achieving Near-Perfect Accuracy

Several deep-seated challenges hinder the quest for “almost 100%” predictive accuracy for individual voters:

• Data Representativeness, Quality, and Granularity:

• AI models are heavily reliant on the data they are trained on. Social media data, a common source for projects like ElectionSim, is not fully representative of the entire electorate and can be characterized by noise, performativity, and polarization.
• Official demographic or voter file data may lack the necessary nuance regarding an individual’s evolving beliefs, values, or immediate intent.
• Synthetic data, as used by Zhao et al. ²⁰, depends on the quality of the aggregate data used for its generation and the underlying assumptions of the generative model, potentially failing to capture true individual heterogeneity.
• There is often a mismatch between the granularity of available data and the complexity of the behavior being modeled. Current data sources (social media posts, demographic categories, survey responses) are often coarse proxies for the intricate, multi-layered cognitive and affective processes that drive a voter’s decision. An AI may model correlations expressed in this data but miss the underlying causal mechanisms if the data itself doesn’t contain those deeper signals.

• Algorithmic Bias: AI models can inherit and even amplify biases present in their training data (e.g., demographic, ideological, or behavioral biases), leading to systematically skewed or unfair predictions.³¹ This is a critical hurdle for creating truly “faithful” mimics that do not disadvantage certain groups.
• The “Black Box” Problem and Interpretability: Many advanced AI models, particularly deep learning systems like LLMs, operate as “black boxes.” Their internal decision-making processes can be opaque and difficult to interpret, making it challenging to understand why a specific prediction was made. This lack of explainability is a barrier to building trust, debugging errors, and ensuring accountability.³¹
• Dynamic and Evolving Opinions: Voter preferences are not fixed; they are subject to change in response to new information, campaign developments, major societal events, and shifts in personal circumstances.⁶ A model trained on historical data may struggle to accurately capture these future shifts. The continuous, real-time data feedback loop required for a true digital twin ²² is practically unachievable at a societal scale for all the factors that influence a voter’s evolving mindset.
• Emotional, Irrational, and Idiosyncratic Factors: Human decision-making is often influenced by emotions, subconscious biases, and elements of behavior that may appear irrational or are highly specific to an individual’s unique experiences and values.⁹ AI systems, particularly those relying on structured or textual data, face significant difficulties in accurately modeling these deeply human and often unpredictable aspects of choice.
• The “Heisenberg” Effect: The very act of polling or making predictions public can sometimes influence subsequent public opinion or behavior.⁶ If the predictions of digital voter twins were widely disseminated, they could potentially alter the very behaviors they are designed to predict, creating a feedback loop that complicates accuracy.

The pursuit of “almost 100%” accuracy for individual voter choices faces what might be termed a “last mile” problem. While AI can achieve reasonable levels of aggregate or probabilistic prediction (e.g., in the 70-85% accuracy range for certain tasks), closing the remaining gap to near-perfect individual prediction is exponentially more difficult. This is because the residual variance in behavior is often attributable to highly idiosyncratic, deeply personal, context-dependent, or genuinely stochastic (random) factors that are extremely challenging for AI to capture, model, or access data for. The closer one attempts to get to perfect prediction, the more comprehensive and potentially invasive data collection would need to become, and even then, fundamental aspects of human agency and unpredictability would likely remain.

Furthermore, if it were indeed possible to predict individual voter choices with near-total certainty, it would raise profound philosophical questions about the nature of free will and determinism in political behavior. While political science inherently seeks to identify patterns and predictability ²⁸, it generally acknowledges a significant degree of individual agency and inherent unpredictability. The societal acceptance and trustworthiness of AI voter mimics might hinge on this philosophical point: do citizens wish to be perceived as perfectly predictable entities? This could significantly impact their willingness to be “represented” by such an AI.

Factors Influencing Voter Choice vs. AI’s Current Modeling Capacity

The following table outlines key factors known to influence voter decisions, alongside an assessment of current AI modeling approaches and their limitations concerning each factor.

This table underscores that while AI can model certain aspects of voter profiles and behaviors, particularly those explicitly expressed or correlated with available data, many deeper, more dynamic, and idiosyncratic drivers of human choice remain largely beyond the grasp of current technology, making the “almost 100% accuracy” target profoundly challenging.

V. Critical Considerations: Ethical, Societal, and Democratic Implications

The development and deployment of AI agents designed to mimic or represent voters, while technologically intriguing, are fraught with significant ethical, societal, and democratic implications. These considerations extend far beyond technical feasibility and predictive accuracy, touching upon the very foundations of fair representation, individual autonomy, and the integrity of democratic processes. The technologies enabling sophisticated voter mimicry for potentially benign understanding are virtually identical to those that could be used for manipulation. This inherent “dual-use” nature ²² is particularly acute in the political domain, as the “target” is the democratic will itself. Any project aiming to build AI voter mimics must, from its inception, confront the high probability of its findings or tools being co-opted for unethical political purposes.

A. Accuracy, Bias, and Fairness

• Algorithmic Bias: A primary concern is that AI systems can exhibit algorithmic bias, producing outputs that unfairly discriminate against individuals or groups based on various aspects of their social identity, including political orientation.³² Such biases can be introduced through unrepresentative or prejudiced training data, or through the design choices made during model development.²³ Political biases, in particular, can be more insidious and harder to detect and eradicate than other forms, such as gender or racial bias, because political leanings are often deeply intertwined with complex social narratives and identities.³² For example, AI models trained on historical data reflecting existing societal inequalities may inadvertently perpetuate these disparities in voter profiling or in any “representative” capacity they might assume.³³
• Representativeness and Disproportionate Impact: Ensuring that AI models accurately reflect the diversity of the population is paramount.³³ If “digital twins” of voters are flawed, biased, or based on incomplete data, any decisions or policies informed by them could disproportionately harm or disadvantage certain groups, particularly those already marginalized.²²
• Consequences of Flawed Models: The deployment of inaccurate or biased AI voter models carries severe risks. It could lead to the misrepresentation of voter segments, resulting in misinformed policy-making or the unfair allocation of resources if AI-generated “representations” are given credence. More broadly, consistently erroneous or biased outputs from such systems could significantly erode public trust in both the technology and the democratic institutions that might employ it.²³

If AI voter twins are utilized for any form of representation, even purely advisory, biases embedded within these AI systems can lead to systematically skewed policy recommendations or simulated public opinions. This, in turn, can influence actual policy outcomes in ways that create feedback loops, further marginalizing underrepresented or misrepresented groups. For instance, if a biased AI twin under-represents the concerns of a particular demographic, policies influenced by this AI’s input might neglect that demographic’s needs. The resulting negative outcomes for that group could then be reflected in future data (e.g., lower civic engagement, poorer socio-economic indicators), which, if fed back into AI models, could reinforce the original bias. This creates a potential downward spiral where the AI, initially intended to enhance democratic understanding or participation, actually undermines fairness and exacerbates inequality—a significant risk for any system aiming to “represent the citizen.”

B. Data Privacy and Security

• Collection of Sensitive Voter Information: The creation of effective AI voter models, especially “digital twins,” necessitates the collection and processing of extensive and often highly sensitive personal data. This includes demographic details, online behaviors, expressed political views, social connections, and potentially even inferred psychological traits.¹
• Data Protection Imperatives: The handling of such data must adhere to stringent data protection and privacy regulations.³³ Current practices by political parties regarding data use, such as through canvassing apps, have already raised significant privacy concerns, including issues around data sharing with commercial third parties (e.g., credit referencing agencies like Experian) and the presence of security vulnerabilities in these applications.²⁴ These concerns would be magnified exponentially with the far more detailed and comprehensive datasets required for sophisticated voter twins.
• Risk of Breaches and Misuse: Concentrated repositories of detailed voter information represent highly attractive targets for cyberattacks and unauthorized access.²² Breached data could be exploited for various malicious purposes, including identity theft, sophisticated social engineering, foreign interference, or commercial exploitation far beyond the originally stated purpose of voter simulation.
• Transparency in Data Utilization: A pervasive lack of transparency regarding how political entities currently collect, use, and share voter data is a well-documented problem.²⁴ This opacity would become even more problematic if citizens’ data were being used to construct digital representations intended to speak or act on their behalf, without clear understanding or consent regarding the full scope of that data’s application.

C. Manipulation and Democratic Integrity

• Microtargeting and Persuasion at Scale: AI technologies, particularly generative AI, enable the creation of highly scalable and personalized political advertising tailored to individuals’ inferred personality traits, psychological vulnerabilities, and specific concerns.³⁶ This allows for campaigns to “sway voters” by exploiting deep-seated psychological attributes deduced from online behavior and personal data.³⁶ While the persuasive effect of any single microtargeted ad might be small, the cumulative impact at scale, especially in closely contested elections, could be substantial.³⁶
• Proliferation of Deepfakes and Disinformation: AI tools can generate deceptively realistic false content, including “deepfake” videos or audio of political candidates saying or doing things they never did, or entirely fabricated narratives designed to mislead the public, suppress voter turnout, or incite social unrest.⁴ Incidents involving AI-generated robocalls impersonating political figures and fake audio recordings designed to influence elections have already been documented in various countries.⁴
• Erosion of Informed Public Discourse: AI-driven content curation algorithms, particularly on social media platforms, can reinforce echo chambers and exacerbate political polarization by selectively exposing users to information that aligns with their pre-existing beliefs.⁴ Concerns have been raised about “hypnocracy”—a form of digital governance that subtly manipulates public consciousness and emotional responses through AI-driven engagement algorithms.⁴
• Undermining Electoral Processes and Trust: Beyond influencing voters, AI can be weaponized to directly undermine the administration and perceived legitimacy of elections. This includes manufacturing fake evidence of electoral misconduct (e.g., ballot tampering), launching cyberattacks against election infrastructure, or spreading disinformation to disrupt vote counting and certification procedures.³⁸

D. Transparency, Accountability, and Control

• The Need for Explainability (XAI): The “black box” nature of many sophisticated AI models makes it difficult to understand their internal reasoning, which is a significant barrier to ensuring accountability, debugging errors, and building public trust.²³ Meaningful transparency in algorithmic decision-making is crucial, particularly when these decisions have political ramifications.²²
• Maintaining Human Oversight and Agency: A consensus among experts is that AI should be developed and deployed as a tool to assist human judgment, not to replace it.⁴² It is vital that real citizens retain ownership and control over their digital counterparts, including the ability to inspect, understand, and adjust them.²⁵ Ultimately, final decisions in democratic processes should remain with humans.⁸
• Establishing Accountability Mechanisms: Clear lines of responsibility and robust accountability mechanisms are necessary to address instances where AI systems cause harm, make erroneous “representative” decisions, or are misused.²² Identifying who is accountable—the developers, deployers, or overseers—is a complex but essential task.
• The Imperative of Informed Consent: The creation and use of digital twins of voters, especially if intended for any form of representation, must be predicated on clear, explicit, and ongoing informed consent from the individuals concerned. This consent must cover data sources, how the data will be used, the purpose of the model, and the extent of its potential agency.²²

E. Impact on Civic Engagement and Democratic Norms

• Risk of Exclusion and Silencing: The introduction of technology as an intermediary in political discourse can inadvertently create new barriers or distances between citizens and their governments. Opinions filtered or summarized by AI risk silencing nuanced voices or leading to self-censorship if individuals feel uncomfortable sharing sensitive political views with an AI system.⁴³
• Potential for Dehumanization of Politics: An over-reliance on AI in political processes could lead to a dehumanization of public decision-making.²⁷ Reducing complex political issues and diverse human perspectives to data points for algorithmic processing risks missing essential human values, emotional nuances, and the deliberative aspects of politics.⁴²
• Impact on Trust in Democratic Institutions: The misuse, malfunction, or perceived unfairness of AI voter systems could further erode already fragile public trust in democratic institutions and electoral processes.²³ Conversely, if AI tools are implemented transparently, fairly, and effectively to enhance citizen engagement or service delivery, they could potentially help to rebuild trust.⁵

The development of powerful AI voter mimics, particularly if spearheaded by private companies or entities with specific ideological agendas ²⁷, raises fundamental questions about democratic sovereignty. Who governs these potentially influential digital representations of the citizenry? If AI twins become significant in shaping public opinion or informing policy, as envisioned by concepts like the Digital Twin Citizen or projects like AI-GOV, then their underlying design, the data they use, and the algorithms that drive them effectively become new forms of governance mechanisms. This brings to the fore concerns about the “privatization of democracy” ²⁷ and the prospect of critical civic functions being guided by algorithms that operate without direct democratic oversight.²⁷ If a private entity or a partisan group develops the most “accurate” or persuasive voter twins, they could gain undue influence over the political landscape. Ensuring public accountability and democratic control over such potent technologies is a paramount, third-order challenge concerning the future power dynamics in an AI-mediated society.

VI. Strategic Insights for Developing AI Voter Agents

Developing AI agents that aim to mimic or represent voters is an endeavor laden with both transformative potential and significant peril. A responsible approach requires a clear-eyed assessment of these dual aspects and a steadfast commitment to embedding democratic values at every stage of the project lifecycle.

A. Synthesizing Potential and Peril

The potential benefits of AI in understanding and engaging with voters are compelling. These technologies could offer deeper insights into voter preferences and behavior, potentially leading to more responsive policy-making. They might open new avenues for citizen engagement, particularly for marginalized or underrepresented groups, by providing novel platforms for their voices to be heard and considered.⁵ AI tools could also facilitate more structured and informed deliberation on complex issues.

However, these potential upsides are counterbalanced by formidable challenges and risks. As detailed in previous sections, achieving truly accurate and faithful mimicry of individual human choices faces immense technical hurdles related to data limitations, the complexity of human cognition, and the dynamic nature of political opinions (Section IV). Furthermore, the ethical risks are pervasive and profound, encompassing algorithmic bias, threats to data privacy and security, the potential for sophisticated manipulation and disinformation, and the overarching danger of undermining democratic norms and institutions (Section V).

B. Recommendations for a Responsible Project Approach

Any project venturing into the development of AI voter agents must adopt a strategy grounded in caution, ethical foresight, and a citizen-centric philosophy. The following recommendations provide a framework for such an approach:

1. Embed Ethical Frameworks from Inception (Ethics-by-Design)

Ethical considerations cannot be an afterthought or a superficial compliance exercise. Principles of fairness, accountability, transparency, privacy, and safety must be woven into the fabric of the project from its very conception and guide every stage, including data collection, model design and training, deployment, and ongoing oversight.23 This requires the active involvement of a diverse range of stakeholders—including ethicists, social scientists, legal experts, representatives from citizen groups, and policymakers—in defining and enforcing these ethical guardrails.33 Given the high stakes involved in political representation, a “minimum viable ethics” should be established, outlining non-negotiable principles such as a commitment to non-manipulation, ensuring the political neutrality of the core AI model, and building robust protections against the model’s use for voter suppression or disinformation.

2. Prioritize Transparency, Explainability, and Contestability

Efforts should be made to develop AI models whose decision-making processes are as transparent and explainable as possible, not only to developers but also, to a reasonable extent, to end-users or independent overseers.22 This is crucial for building trust and enabling accountability. Furthermore, mechanisms should be established that allow citizens to understand how their “twin” or profile is constructed, to scrutinize its assumptions, and to contest or correct its outputs if they are perceived as inaccurate or misrepresentative. This aligns with the principle that citizens should maintain ownership and the ability to adjust their digital counterparts.25

3. Implement Robust Data Governance and Privacy-Preserving Techniques

Strict adherence to all applicable data protection laws and regulations is fundamental.35 Projects must employ robust data governance frameworks and utilize privacy-preserving techniques wherever feasible. These may include data minimization (collecting only essential data), anonymization or pseudonymization, differential privacy, federated learning, and secure data storage and transmission protocols.33 Full transparency regarding data sources, how data will be used, and any data-sharing policies is essential. Crucially, explicit, informed, and ongoing consent must be obtained from individuals for the collection and use of their data in creating and operating these AI agents.

This concept of consent must be dynamic. If a digital twin is intended to “faithfully mimic” a voter over time, reflecting their evolving views and values ⁶, then both the AI twin itself and the citizen’s consent for its operation must also be dynamic. A one-time consent for a static profile is insufficient for a continuously learning and potentially acting agent. This implies a need for mechanisms allowing citizens to regularly review their AI twin’s profile, re-consent to its operation under clearly defined terms, and make necessary adjustments to ensure it accurately reflects their current perspectives.²⁵

4. Proactively Address and Mitigate Biases

Bias in AI systems is a pervasive challenge. Developers must make concerted efforts to use diverse, representative, and carefully vetted datasets for training models.33 Regular audits of the models for various forms of bias (e.g., demographic, ideological, regional) are necessary, along with the implementation of appropriate debiasing techniques.31 It must be acknowledged that complete elimination of bias is often an elusive goal; therefore, a commitment to continuous monitoring, iterative improvement, and transparent reporting on bias metrics is crucial.23

5. Set Realistic Goals for Accuracy and Define the Scope of “Representation”

It is vital to set realistic expectations regarding the predictive accuracy of AI voter agents. The ambition of “almost 100% accuracy” in predicting individual human choices is currently beyond technical reach and may even be undesirable if it leads to an oversimplification of human agency or an erosion of privacy (Section IV). The project must clearly and narrowly define what “representation” means in its specific context. Is the AI agent intended to provide insights for researchers, serve as an advisory tool for policymakers, facilitate dialogue among citizens, or potentially engage in some form of delegated decision-making? The technical requirements, ethical safeguards, and societal risks vary dramatically depending on this definition (Section III). If any form of delegation is contemplated, it should begin with extremely low-stakes applications and always ensure robust human oversight and the ability for human override or final decision-making.8

6. Foster Multi-Stakeholder Collaboration and Public Deliberation

Building trust and ensuring that these powerful technologies serve genuinely democratic values requires broad engagement with academia, civil society organizations, government bodies, and the public at large.22 Initiatives such as “AI Elections Labs” or regulatory “sandboxes” could provide controlled environments for piloting new tools, assessing their impacts, and developing best practices before any consideration of wider deployment.44 Public deliberation on the ethical and societal implications is essential to shape norms and guide responsible innovation.

Beyond the governance of individual AI twins, consideration must be given to the governance of the “meta-twin”—the overarching system or platform that creates, manages, and updates these AI agents. The design choices embedded in this meta-architecture (e.g., what data sources are prioritized, which learning algorithms are employed, how “representation” is defined programmatically) will have widespread and uniform effects on all AI twins produced by it. Therefore, the transparency, ethical oversight, and democratic accountability of this twin-creation platform itself are as critical, if not more so, than the governance of individual AI agents. This raises complex questions about who should control such foundational infrastructure—public institutions, private consortia, or open-source communities—to ensure its alignment with democratic principles.

7. Focus on Augmentation, Not Replacement

The development philosophy should frame AI voter agents as tools designed to augment human understanding, enhance civic engagement, and support democratic processes, rather than aiming to replace human judgment, political participation, or democratic representation.25 The ultimate goal should be to empower citizens with better information and new avenues for participation, not to disempower them by outsourcing their political agency to algorithms.

VII. Conclusion: Charting a Responsible Path for AI in Democratic Processes

A. Recapitulation of AI Voter Simulation’s Dual Potential

The exploration of AI-driven voter simulation and the concept of “digital twin citizens” reveals a technology of profound dual potential. There is a genuine and understandable interest in leveraging AI to gain deeper insights into the complex tapestry of voter behavior, to analyze public opinion with greater nuance, and perhaps to explore novel forms of citizen engagement and representation. The projects and conceptual frameworks examined in this report, from ElectionSim’s LLM-driven voter pools ¹ to the theoretical underpinnings of Digital Twin Citizens ⁸ and AI Voice ²⁵, illustrate active efforts to harness this potential.

However, the vision of creating AI agents that can “faithfully mimic” individual voters with near-perfect predictive accuracy, and subsequently act as their representatives, confronts immense technical, practical, and, most critically, ethical hurdles. The complexities of human decision-making, the limitations of current data sources and AI modeling capabilities, and the irreducible element of human agency make the “almost 100% accuracy” aspiration an elusive, and perhaps philosophically problematic, goal.

B. The Primacy of Democratic Values

As societies navigate the integration of AI into various aspects of life, its application within the sensitive domain of democratic processes demands unwavering adherence to fundamental democratic principles. Values such as fairness, transparency, accountability, citizen agency, privacy, and the overall integrity of electoral processes must be the paramount guiding forces.²³ The pursuit of technological sophistication or predictive power in “mimicking” voters must not come at the expense of these core values. The allure of efficiency or novel forms of interaction cannot overshadow the potential for AI to exacerbate existing inequalities, undermine informed discourse, or erode trust in democratic institutions.

It is crucial to avoid the fallacy of what might be termed a “representational singularity”—the idea that an AI, by achieving near-perfect prediction of a citizen’s choices, thereby becomes a legitimate representative. Democratic representation is a rich, normative concept that encompasses far more than mere predictive accuracy. It involves elements of trust, legitimacy, accountability, the capacity for reasoned deliberation, and responsiveness to new arguments and evolving circumstances—qualities that current AI systems cannot fully replicate. Equating predictive power with representational legitimacy is a category error that could lead to dangerous oversimplifications of democratic practice.

C. The Path Forward: Caution, Collaboration, and Citizen-Centricity

The responsible path forward in exploring AI for voter understanding and potential representation necessitates a posture of profound caution, a commitment to broad collaboration, and an unyielding focus on citizen-centricity.

• Cautious Incrementalism: Development and deployment of such technologies should proceed incrementally, prioritizing fundamental research and small-scale, rigorously monitored pilot programs conducted in low-risk, transparent environments.⁴⁴ The potential for unintended consequences is too great for rapid, untested rollouts.
• Interdisciplinary Research and Societal Dialogue: There is an ongoing and urgent need for sustained interdisciplinary research involving computer scientists, political scientists, ethicists, legal scholars, sociologists, and psychologists to fully understand the multifaceted impacts of AI on democracy.²² This research must be coupled with broad and inclusive public deliberation to shape societal norms and inform regulatory frameworks.
• Empowering Citizens: Ultimately, the goal of any AI system intended to engage with or “represent” citizens must be to empower those citizens, not to diminish their agency or outsource their critical faculties. The focus should be squarely on developing AI for democracy, in ways that are guided by democratic principles.²⁷ This means ensuring that real citizens remain informed, in control of their democratic participation, and sovereign over their digital selves.

The introduction of sophisticated AI agents into the political sphere could, over time, subtly alter the baseline of what is considered “normal” political interaction. Engagement that is currently direct and unmediated could become increasingly AI-facilitated. While this might foster new forms of civic literacy or efficiency, it also carries the risk of fostering new forms of civic apathy or dependence if not carefully managed. The long-term societal impact on political culture and the quality of democratic life is a critical, albeit speculative, consideration that warrants ongoing vigilance, research, and adaptive governance. The challenge lies in harnessing the analytical power of AI without surrendering the human core of democratic self-governance.

Works cited

1. ElectionSim: Massive Population Election Simulation Powered by Large Language Model Driven Agents – arXiv, accessed June 8, 2025, https://arxiv.org/html/2410.20746v1
2. Agent-Enhanced Large Language Models for Researching Political Institutions – arXiv, accessed June 8, 2025, https://arxiv.org/pdf/2503.13524
3. Political-LLM: Large Language Models in Political Science, accessed June 8, 2025, https://political-llm.org/
4. AI and Democracy: The Coming Civilization – HEC Paris, accessed June 8, 2025, https://www.hec.edu/sites/default/files/documents/AI-and-Democracy_The-Coming-Civilization_Yann-ALGAN-HEC-Paris_January-2025.pdf
5. Driving Civic Engagement: How AI is Revolutionizing Municipalities – CommBox, accessed June 8, 2025, https://www.commbox.io/driving-civic-engagement-with-ai-is-revolutionizing-municipalities/
6. The Role of Artificial Intelligence in Political Polling – Lerner – University of Delaware, accessed June 8, 2025, https://lerner.udel.edu/seeing-opportunity/the-role-of-artificial-intelligence-in-political-polling/
7. Proven, AI-Powered Voter Insights for Targeted Ad Campaigns – Resonate, accessed June 8, 2025, https://www.resonate.com/solutions/politics/
8. Digital citizens could shake up democracy in Switzerland and …, accessed June 8, 2025, https://www.swissinfo.ch/eng/digital-democracy/digital-citizens-could-shake-up-democracy-in-switzerland-and-beyond/88817898
9. Voting Intention and Choices: Are Voters Always Rational and …, accessed June 8, 2025, https://pmc.ncbi.nlm.nih.gov/articles/PMC4757036/
10. Predicting Popular-vote Shares in US Presidential Elections: A Model-based Strategy Relying on Anes Data – Cambridge University Press, accessed June 8, 2025, https://www.cambridge.org/core/journals/ps-political-science-and-politics/article/predicting-popularvote-shares-in-us-presidential-elections-a-modelbased-strategy-relying-on-anes-data/CC6B59BAC456CAC43C96F27FBAAF2290
11. Information, incentives, and goals in election forecasts – Columbia University, accessed June 8, 2025, https://sites.stat.columbia.edu/gelman/research/published/jdm200907b.pdf
12. Election Night Forecasting With DDHQ: A Real-Time Predictive Framework, accessed June 8, 2025, https://hdsr.mitpress.mit.edu/pub/zr6hjsfl
13. Method to Forecast the Presidential Election Results Based on Simulation and Machine Learning – MDPI, accessed June 8, 2025, https://www.mdpi.com/2079-3197/12/3/38
14. Agent-Based Modeling in Political Decision Making – Oxford Research Encyclopedias, accessed June 8, 2025, https://oxfordre.com/politics/display/10.1093/acrefore/9780190228637.001.0001/acrefore-9780190228637-e-913?d=%2F10.1093%2Facrefore%2F9780190228637.001.0001%2Facrefore-9780190228637-e-913&p=emailAIpBYSg7UI9pI
15. ElectionSim: Massive Population Election Simulation Powered by Large Language Model Driven Agents – arXiv, accessed June 8, 2025, https://arxiv.org/html/2410.20746v2
16. AI Platform Transforms X’s Political Debates Into Solutions for Global …, accessed June 8, 2025, https://www.morningstar.com/news/accesswire/1003907msn/ai-platform-transforms-xs-political-debates-into-solutions-for-global-crises
17. ElectionSim: Massive Population Election Simulation Powered by Large Language Model Driven Agents | Request PDF – ResearchGate, accessed June 8, 2025, https://www.researchgate.net/publication/385317618_ElectionSim_Massive_Population_Election_Simulation_Powered_by_Large_Language_Model_Driven_Agents
18. A Large-Scale Simulation on Large Language Models for Decision-Making in Political Science – arXiv, accessed June 8, 2025, https://arxiv.org/html/2412.15291v3
19. CV – Yue Zhao – University of Southern California, accessed June 8, 2025, https://viterbi-web.usc.edu/~yzhao010/files/ZHAO_YUE_CV.pdf
20. arXiv:2412.15291v4 [cs.CL] 10 Apr 2025, accessed June 8, 2025, https://arxiv.org/abs/2412.15291
21. Agent-Based Models of Quadratic Voting, accessed June 8, 2025, https://ccl.northwestern.edu/2021/Kelter2021.pdf
22. Digital twins – dynamic models that respond to real-time data – POST Parliament, accessed June 8, 2025, https://post.parliament.uk/digital-twins-dynamic-models-that-respond-to-real-time-data/
23. The ethical conundrum of electoral AI #3 – International IDEA, accessed June 8, 2025, https://www.idea.int/news/ethical-conundrum-electoral-ai-3
24. New report: Voter data privacy concerns over apps used by political parties, accessed June 8, 2025, https://www.openrightsgroup.org/press-releases/new-report-voter-data-privacy-concerns-over-apps-used-by-political-parties/
25. AI Agents in Global Governance: Digital Representation for Unheard …, accessed June 8, 2025, https://multilateralism.sipa.columbia.edu/news/ai-agents-global-governance-digital-representation-unheard-voices
26. Liquid Democracy – Hoover Institution, accessed June 8, 2025, https://www.hoover.org/research/liquid-democracy
27. AI Must Be Governed Democratically to Preserve Our Future – HEC Paris, accessed June 8, 2025, https://www.hec.edu/en/knowledge/articles/ai-must-be-governed-democratically-preserve-our-future
28. 2.3 Human Behavior Is Partially Predictable – Introduction to Political Science | OpenStax, accessed June 8, 2025, https://openstax.org/books/introduction-political-science/pages/2-3-human-behavior-is-partially-predictable
29. Forecasting Turnout – Harvard Data Science Review – MIT, accessed June 8, 2025, https://hdsr.mitpress.mit.edu/pub/0c5ylgo1
30. Full article: Predicting political attitudes from web tracking data: a …, accessed June 8, 2025, https://www.tandfonline.com/doi/full/10.1080/19331681.2024.2316679
31. The Future of Politics: Algorithmic Insights – Number Analytics, accessed June 8, 2025, https://www.numberanalytics.com/blog/future-of-politics-algorithmic-insights
32. Algorithmic Political Bias in Artificial Intelligence Systems – PMC – PubMed Central, accessed June 8, 2025, https://pmc.ncbi.nlm.nih.gov/articles/PMC8967082/
33. Ethics in Artificial Intelligence and Government – Institute for Citizen-Centred Service -, accessed June 8, 2025, https://iccs-isac.org/assets/uploads/research-repository/JC-Research-Report-Ethics-in-AI-and-Government-April-2023.pdf
34. Full article: Artificial intelligence and democracy: pathway to progress or decline?, accessed June 8, 2025, https://www.tandfonline.com/doi/full/10.1080/19331681.2025.2473994
35. Data protection in political campaigns – Michalsons, accessed June 8, 2025, https://www.michalsons.com/blog/data-protection-in-political-campaigns/71021
36. persuasive effects of political microtargeting in the age of generative …, accessed June 8, 2025, https://academic.oup.com/pnasnexus/article/3/2/pgae035/7591134
37. Summary Artificial Intelligence (AI) in Elections and Campaigns, accessed June 8, 2025, https://www.ncsl.org/elections-and-campaigns/artificial-intelligence-ai-in-elections-and-campaigns
38. How Artificial Intelligence Influences Elections and What We Can Do About It, accessed June 8, 2025, https://campaignlegal.org/update/how-artificial-intelligence-influences-elections-and-what-we-can-do-about-it
39. Election Interference in An Age of AI-Enabled Cyberattacks and Information Manipulation Campaigns – Freeman Spogli Institute for International Studies, accessed June 8, 2025, https://fsi.stanford.edu/sipr/content/election-interference-age-ai-enabled-cyberattacks-and-information-manipulation-campaigns
40. The Role of Social Media Algorithms in Shaping Public Opinion During Political Campaigns, accessed June 8, 2025, https://www.researchgate.net/publication/386347360_The_Role_of_Social_Media_Algorithms_in_Shaping_Public_Opinion_During_Political_Campaigns
41. Addressing the Voter Information Deficit With AI – Karsh Institute of Democracy, accessed June 8, 2025, https://karshinstitute.virginia.edu/news/addressing-voter-information-deficit-ai
42. AI on the Ballot: How Artificial Intelligence Is Already Changing Politics – Ash Center, accessed June 8, 2025, https://ash.harvard.edu/articles/ai-on-the-ballot-how-artificial-intelligence-is-already-changing-politics/
43. AI for public participation: hope or hype? – ECNL, accessed June 8, 2025, https://ecnl.org/news/ai-public-participation-hope-or-hype
44. Secretary Fontes Releases Final Report on AI and Election Security: Proposes Groundbreaking AI Elections Lab | Arizona Secretary of State, accessed June 8, 2025, https://azsos.gov/news/903
45. AI and Democracy: Scholars Unpack the Intersection of Technology and Governance, accessed June 8, 2025, https://isps.yale.edu/news/blog/2025/04/ai-and-democracy-scholars-unpack-the-intersection-of-technology-and-governance