In this Living Literature Review, I will guide you on an examination of how digital tools and advanced AI systems are already altering how democratic input is both generated and delivered to governments. We will begin with a brief history of voting methods in the US as an example of how technological evolution has historically shaped this particular type of democratic input. From here we explore two recent academic literature reviews from late 2024 on the topics of digital tools for citizen participation and the use of LLMs in politics and democracy. These literature reviews provide rich material from which we can explore more deeply on these topics. Then we explore, in detail, two specific empirical studies from 2024 that apply LLMs to the challenges of low voter turnout and low information on collective preferences. Both of these studies point to promising specific directions for LLMs to improve democratic input, but also highlight that LLMs are not ready to be “digital twins” or strong predict individual policy preferences. Finally, as something of a robustness check, we explore the recently released 2024 Federal AI Use Case Inventory to see how US Federal Agencies are developing and deploying LLMs to improve democratic input, with a specific focus on simulating collective preferences and behaviors as they relate to various policy domains.

Introduction

One of the key governing strengths of a democracy is democratic input. Democratic input provides the feedback needed for the governing system to respond, adapt, and evolve. Democratic input also sets the general direction of the goals of a functioning democracy. Democratic input captures the sense of democracy that Lincoln pointed at in his Gettysburg Address: “Government of the people, by the people, for the people”. However, as with bureaucracies1, democratic input has evolved alongside technological and cultural evolution.

One well documented example of this in the US context (the one with which I am most familiar) is voting. Voting is the mechanism that is most often associated with democratic input. In the US, voting methods are diverse, as they are run by individual states, but nonetheless one can see a broad trend of technological evolution from the late 1800’s until now.2 The broad strokes are that voting went from using your voice in a public setting, to submitting party-backed pre-filled ballots, to standardized paper ballots, to machines with a lever that you used to mark paper ballots, to touchscreen digital electronic screens with paper trails, to, finally, secure apps and web-portals. You can see representative pictures of several of these types here in Douglas Jones (2003) “A Brief Illustrated History of Voting.” I’ve included a few these pictures below:

While this evolution of voting is instructive, it is only one piece of the evolution of democratic participation towards more use of newer information and communication technologies (ICTs) and “electronic” participation (e-participation) tools.

In the following section, we begin with an overview of the standard “tools” of the “current paradigm” of democratic input from Shin et al’s (2024) “A systematic analysis of digital tools for citizen participation.” With the stage set for an understanding of how digital tools have already reshaped democratic input and citizen participation, we will then review the current evidence for the use of advanced AI, typically in the form of Large Language Models (LLMs) for democratic input. For this, we begin with an exploration of Aoki’s (2024) “Large Language Models in Politics and Democracy: A Comprehensive Survey.” This survey provides an overview from the academic literature of the ways in which LLMs are already being used as forms of democratic input. Then we will explore two specific studies that look at using LLMs for soliciting individual policy preferences, individual votes, and simulating collective preferences as captured by elections. Finally, we briefly explore the recently released 2024 US Federal AI Use Case Inventory for further examples of LLM use by US Federal Agencies to improve democratic input.

A Systematic Analysis of Digital Tools for Citizen Participation

The first of the two academic literature reviews from 2024 that we will explore is “A Systematic Analysis of Digital Tools for Citizen Participation.” Shin et. al. both give us a framework for thinking about digital participatory tools and analyze 116 tools.

This article aligns with the approach by systematically analysing digital tools in the context of e-participation. Digital participatory tools are defined as digital applications or systems specifically designed to enhance e-participation. Consequently, general tools such as Facebook and Zoom are excluded from the scope of this article. Instead of focusing on isolated instances, this article identifies prominent trends by analysing a comprehensive dataset of 116 digital participatory tools from three public repositories.

In the article, the authors raise three guiding research questions:

1. What are the main functions that digital tools provide for e-participation?

2. What are the prominent clusters of digital tools?

3. Do digital tools potentially enhance inclusiveness, deliberation, and empowerment?

The study goes on to identify specific “genes” across four dimensions that can be combined to create a “unique genome” for any specific digital tool. These “dimensions of genes” are:

1. What is being done? (goal): creation, decision

2. Who is doing it? (staffing): hierarchy, crowd

3. Why are they doing it? (incentive): money, love, and glory

4. How is it being done? (structure/process): collection, collaboration, group decision, individual decision

To explore these questions, the authors combined data from three public repositories Digital Participatory Platforms (containing 50 platforms); OECD Guidelines for Citizen Participation Processes (16 digital tools); and the Collective Intelligence through Digital Tools (62 tools). These combined repositories created a dataset of 116 unique digital tools that existed as of November of 2022. The dataset created has a row for each of the 116 tools and columns for each of the “genes” or features that are recorded for each tool. The genes correspond to various answers to the four questions above (what, who, why, when, and how). Then they perform a series of different types of cluster analyses to explore Research Questions 1 & 2.

The first analysis, corresponds to the first research question (What are the main functions that digital tools provide for e-participation?). This analysis then is about clustering genes based on their interaction, which in the analysis is the frequency of co-occurrences of genes.

For this analysis, three major clusters were identified, with the first cluster having significant strength. The first cluster primarily involves functions of problem and solution identification. The second cluster is primarily about collective decision making. The third cluster consists of only two genes: co-funding and technology regarding donation.

The authors describe the results to the first analysis like this:

This result indicates that digital tools are predominantly designed for crowdsourcing (e.g., information, opinions, and funding) to assist policymakers in making informed decisions.

The second analysis, corresponds to the second research questions (What are the prominent clusters of digital tools?). This analysis is about clustering similar tools based upon their attributes, their genes.

For the second analysis, the authors find 5 clusters of digital tools. The first cluster contains 56 of the 116 tools. So as with the first cluster of tool functions, the first cluster here is particularly strong. The first cluster is fundraising and problem and solution identification. Given the size and strength of cluster 1, the authors identified 5 sub-clusters within cluster 1. These 5 sub-clusters are:

1. Crowd-sourced mapping tools

2. Interactive and place-based survey tools for collecting citizens' voices

3. Facilitate idea creation and deliberations for consensus building or collective action

4. Analytical tools

5. Co-funding tools

In addition to this large Cluster 1, the authors also identified an additional 4 clusters of digital tools. Cluster 2 focuses on drafting and decision-making. Cluster 3 is a collection of specialized decision-making tools. Cluster 4 tools are provided by the public and nonprofit sectors, which facilitate a wide range of processes, from identifying problems to decision-making. And, finally, Cluster 5 includes various tools provided by small and medium-sized tech firms with dedicated support teams.

In analyzing these clusters together, the authors settle on 3 main clusters of digital participatory tools, provide types from each of these clusters, and the list example from the repositories. Their Table 7 below summarizes these findings.

This table helps to capture the answers to the three main research questions the authors began with. The main prominent clusters of tools are: 1) problem and solution identification, 2) co-funding, and 3) decision-making. Within Cluster 1, main types of functions are: crowd-source mapping, interactive survey, and analytical tools. Cluster 2 contains crowdfunding as a key type of co-funding. Cluster 3 contains several functions as well including: drafting, decision-making, and comprehensive.

With the main clusters and functions identified, the authors also can examine research question 3: “Do digital tools potentially enhance inclusiveness, deliberation, and empowerment?” Here the authors argue that the tools they observe do seem to provide enhanced opportunities for inclusiveness and public deliberation, but the authors are more skeptical of empowerment. This finding comes directly from a few of the key takeaways offered by the authors:

1. “a relative absence of feedback loops, highlighting a gap in mutual interactions between citizens and government”

2. “only a few cases in our dataset included real-time monitoring and assessment systems”

3. “while citizens are encouraged to provide their data (e.g., comments, proposals, and votes) under e-participation, there is insufficient focus on how their voices influence decision making and policy action”

4. “However, given the ever-increasing availability of digital tools, our dataset does not cover new tools or prototypes that might contain advanced AI technologies… nor did we cover generative AIs with significant potential for human-machine collaboration.”

5. “Lastly, this article placed less emphasis on democratic aspects, such as equity, equality, freedom of expression, representation, civic education, and empowerment — all crucial for measuring the quality of e-participation.”

And in closing, the authors make a plea for more work in this direction

Therefore, we call for future research with a dedicated research design and theoretical framework that concentrates on exploring the democratic aspects of digital tools. This holds significance due to the pivotal challenge of digital inclusion in e-participation, a crucial aspect for fostering equality and empowerment.

“A Systematic Analysis of Digital Tools for Citizen Participation” was published in September of 2024 in leading academic journal Government Information Quarterly. On December 1, 2024, Goshi Aoki provided a response to some of these takeaways in the ArXiv preprint titled “Large Language Models in Politics and Democracy: A Comprehensive Survey.”

Large Language Models in Politics and Democracy: A Comprehensive Survey

Ask and you shall receive, 3 months later, the first paragraph of the abstract of “Large Language Models in Politics and Democracy: A Comprehensive Survey” :

The advancement of generative AI, particularly large language models (LLMs), has a significant impact on politics and democracy, offering potential across various domains, including policymaking, political communication, analysis, and governance. This paper surveys the recent and potential applications of LLMs in politics, examining both their promises and the associated challenges. This paper examines the ways in which LLMs are being employed in legislative processes, political communication, and political analysis.

As the title and the abstract above suggest, this essay provides a comprehensive review of how LLMs in particular are being used to influence politics and democracy. For our purposes, we are most interested in how LLMs influence democratic input, so not all the sections covered are equally relevant. But, when one table can do so much to give you the high-level findings of each of the sections, it seems like this table should be the starting point.

So:

Before diving into the ways in which LLMs can be applied to politics and democracy as covered in this survey, I think it’s helpful to describe the “genetic makeup” of LLMs using the same sets of genes that the previous piece used to describe digital tools more generally. We run into a little difficulty as to what counts as a LLM, but I’ll briefly address this below.

The genetic makeup of LLMs looks something like the following:

1. Goals: co-production of knowledge; co-construction of policies and decisions

2. Providers: Private companies lead; Open Source plays a large role; some movement by Public sector

3. Rationales: problem identification, solution identification, drafting, decision-making, implementation, and assessment

4. Functionalities: Various human-human interaction, human-machine interactions, and Artificial Intelligence

What should make these particular genes noteworthy are the degree to which they capture the “general-purpose” nature of LLMs. That is LLMs have a wider set of capacities for use of digital tools than any of the digital tools explored in the previous study.

Now, we’re presented with one more challenge before we can turn to explore the impacts of LLMs on democratic input. And this one is pretty meta: what counts as democratic input? In particular, on what continuum of detailed and accurate must a simulation of the population be for it to count as a form of democratic input? When do simulations count as democratic input? For the purposes of understanding how LLMs might impact democratic input, I think simulations should be included. The idea is that a very accurate simulation of individuals or aggregated preferences would indeed be representative of the human preferences being simulated.

With all of this in mind, what is it that LLMs can do to influence democratic input?

This comprehensive survey lists 10 “application areas” which correspond somewhat loosely to the “function areas” from the digital tools paper. These application areas, along with a direct supporting quote from the survey itself, include:

1. Document Classification (“LLMs have shown promise in automating the analysis and classification of policy documents”)

2. Policymaking Drafting & Analysis (“real-time sentiment analysis and multilingual translation, which streamline decision-making in global governance contexts”)

3. Policy Emulation (“simulate institutional decision making” “LLM-powered agents could emulate realistic institutional behaviors such as incremental policy adjustments and stakeholder negotiation”)

4. Participatory Policy Design (“engaged stakeholders in iterative drafting and testing of LLM-generated policies”)

5. Text Analysis (“annotating political texts, classifying them by political relevance, negativity, sentiment, and ideology across multiple languages”)

6. Persuasive Messaging (“AI-generated messages can be as persuasive as human-generated content in influencing attitudes on political issues”)

7. Election Simulation (“LLMs can simulate and predict political dynamics”)

8. Public Deliberation (“Plurals, a system that leverages LLM-driven agents with personas to simulate focus groups.” “AI mediator that helped groups find common

   ground on divisive topics. The AI-mediated statements were rated higher for clarity and fairness, reducing divisions and promoting consensus.”

9. Social Simulations (“GOVSIM platform simulates resource-sharing dilemmas, revealing that advanced LLMs, particularly when combined with communication and moral reasoning capabilities, can achieve partial sustainability in managing shared resources.”)

10. Economic Modeling (“LLMs as "Homo Silicus," computational analogs of humans for economic experiments. By replicating classic behavioral economics studies, the research showed that LLMs can exhibit context-sensitive behavior qualitatively similar to human data, offering a cost-effective and ethical alternative for piloting experiments and exploring economic theories.”)

Even just from these examples, we can see the ways in which LLMs are much more generally capable than any specific digital tool from the digital tool repositories explored in the previous study. One of the things about LLMs is that any one system can perform a wide array of functions, as demonstrated by the list of 10 applications above. However, as we shall we below, the impact and accuracy of LLMs in relation to democratic input is still limited. While LLMs show differential progress in capabilities across the application areas above, transformative cases like digital twins and voting reform still require improvements in LLM capabilities themselves.

With this in mind, let’s take a detailed look at two recent studies that empirically explore the ability of LLMs to simulate voting preferences at both the individual level and the societal level.

Two Empirical Cases: Digital Twins & Fair Voting

Large Language Models (LLMs) as Agents for Augmented Democracy

Gudiño, Grandi, and Hidalgo (2024) conducted a study exploring the use of LLMs as "digital twins" for an augmented democracy, where AI systems help represent citizens’ policy preferences. The researchers used data from Brazil’s 2022 presidential election, where 267 participants compared 67 policies from the platforms of candidates Lula da Silva and Jair Bolsonaro. Participants also provided demographic details, including political ideology, education, gender, and age. The study fine-tuned six LLMs, including GPT-3.5, LLaMA-2, and Falcon, using the Low-Ranking Adaptation (LoRA) method to train the models on this dataset, which is essentially a form of post-training fine tuning. The models were tested for their ability to predict individual preferences and aggregate population-level preferences, with results compared against a bundle rule—a heuristic assuming participants always favor proposals from the candidate aligned with their stated ideology.

Figure 1 from the study, visually displays this process:

The study found that fine-tuned LLMs consistently outperformed the bundle rule in predicting individual preferences, achieving accuracy rates of up to 76.7% with LLaMA-2 and 70.4% with GPT-3.5, compared to 71% for the bundle rule. When predicting preferences for pairs of proposals from the same candidate’s platform—a task where the bundle rule has no predictive power—the LLMs maintained accuracy levels between 65% and 77%. At the population level, probabilistic samples augmented with LLM predictions produced more accurate estimates of aggregate preferences than probabilistic samples alone.

This can be seen in Figure 2 from the study included below:

The study also highlighted important findings related to demographic factors influencing model accuracy. For example, the LLMs more accurately predicted the preferences of liberal participants compared to conservatives or centrists, and college-educated participants' preferences were better modeled than those of non-college-educated participants. Furthermore, while accuracy differences across gender were mixed (with some models favoring male participants and others favoring female participants), the models consistently performed slightly better for younger participants compared to older ones. These demographic variations suggest that while LLMs can capture nuanced preferences, their effectiveness may vary across different groups, which could present challenges for ensuring equitable representation in real-world implementations.

The findings suggest significant potential for using LLMs to enhance democratic participation by allowing citizens to express granular preferences beyond traditional party-line voting. By "unbundling" policy proposals from candidates’ platforms, LLMs can offer a more flexible approach to civic engagement, where individuals’ preferences on specific issues are represented in decision-making processes. However, the study also underscores the limitations of current LLM technology, particularly its struggles with nuanced or contradictory preferences and its dependence on the quality and diversity of training data. While the models showed promise in both individual and collective preference modeling, their performance varied based on participant demographics and specific tasks, such as predicting preferences across ideological divides or complex proposals.

Generative AI Voting: Fair Collective Choice is Resilient to LLM Biases and Inconsistencies

Majumdar, Elkind, and Pournaras (2024) conducted a study to evaluate how generative AI, particularly large language models (LLMs), can simulate voting behavior and support collective decision-making. The researchers employed a factorial design to simulate voting scenarios using over 50,000 AI-generated personas in 81 elections, encompassing cases like the 2012, 2016, and 2020 U.S. national elections and a participatory budgeting campaign in Aarau, Switzerland. They assessed three LLMs—GPT-3, GPT-3.5, and Llama2—alongside traditional voting aggregation methods, such as utilitarian greedy and equal shares, across various ballot formats ranging from simple binary choices to complex preference ranking. (Utilitarian greedy refers to a voting method that selects the most popular projects sequentially until the budget is exhausted, while equal shares aims to ensure proportional representation by distributing voting power equitably across all participants.) The study found that binary voting systems achieved higher consistency between human and AI choices (82.3%) compared to complex ballots (4.5%-27.2%), revealing the challenge of complex ballots for AI performance. By emulating human voters and their decision-making processes, the research explored the viability of LLMs in approximating both individual and collective preferences.

Figure 1 from the study is provided below. It graphically displays the research process and results.

A critical component of the study involved addressing the challenges posed by low voter turnout. The researchers simulated scenarios where abstaining voters were represented by AI personas, finding that AI could recover lost consistency—restoring the alignment of outcomes with what would have been achieved if more humans had participated—by up to 53.3% under low-turnout conditions. This recovery was especially significant in voting scenarios with fewer options but less effective in complex participatory budgeting contexts with many alternatives. The equal shares method was particularly robust, maintaining proportional representation and consistency even with as much as 50% of voters abstaining. In contrast, utilitarian greedy methods were less resilient in preserving equitable outcomes, especially as the proportion of AI representatives increased. These findings demonstrated the importance of aggregation methods in effectively incorporating AI into democratic participation.

However, while the findings suggest that AI interventions could enhance democratic resilience, they also reveal significant limitations. The study showed that political ideology, race, and levels of political engagement introduced notable inconsistencies. For example, white participants exhibited a 33% higher rate of inconsistency between their choices and AI-generated votes compared to non-white participants, suggesting demographic biases in how well AI modeled voter preferences. Additionally, individuals who frequently discussed politics displayed 63.2% higher inconsistency in AI alignment compared to those who did not, highlighting the difficulty of modeling nuanced or strongly held political beliefs. While conformity biases—such as preferences for environmental projects—positively influenced consistency, these unconscious biases tied to demographic and ideological factors demonstrated how LLMs can struggle with equitable representation.

Finally, while the researchers advocate for the adoption of fair voting methods like equal shares, they do not fully address the practical and ethical challenges of integrating AI into real-world voting systems. Issues such as data privacy, trust, transparency, and the governance of AI systems remain critical gaps in the discussion. The results underscore the potential of AI to support democratic innovations, but they also highlight the need for caution and comprehensive safeguards to ensure that these systems do not inadvertently undermine the democratic principles they are designed to support. As such, the study provides valuable insights but invites further exploration of the complexities and risks associated with AI-driven voting systems.

Taken together these two studies suggest promising innovative methods for improving democratic input through the use of LLMs for augmenting and simulating voter preferences. However, while LLMs perform well in some cases, complexity of the ballot, complexity of the aggregation method, and demographic and political biases all work against the current use of LLMs for voting. Despite these limitations, LLMs can already play an augmenting role for additional democratic input for information concerning individual preferences and simulating aggregate voting outcomes.

Federal AI Use Case Inventory

Finally, we have the US Federal AI Use Case Inventory. We’ve explored how the comprehensive survey characterizes the academic literature, but what uses is the government pursuing? A comprehensive assessment is beyond the scope of our goals for this review, but let’s take a brief looks at some this inventory and the scope of LLM use in service of democratic input.

In 2023, the US Government began tracking AI use cases that are deployed throughout the US executive branch.3 In the first year of the database, released at the end of 2023, over 700 AI use cases were reported. On December 15th of 2024, the Office of Management and Budget released the updated inventory of Federal AI Use Cases.4 The 2024 inventory lists over 1,700 cases and is available for download on GitHub.5

An initial evaluation of the dataset, supports the general picture that has been drawn in this literature review. That is, LLMs and generative AI tools are beginning to see adoption, testing, and iteration by governments. In this inventory of over 1,700 cases, at least 200 of the use cases explicitly note that they used LLMs or generative AI as part of their use case. The data on the use case are not always complete nor contain all the relevant details, so one can assume 200 is a lower bound here. This is a surprisingly large number to me given the overall newness of LLMs and generative AI tools.

Among the programs that utilized LLMs or generative AI, there are several uses that are directly about democratic input, and there are also quite a few that use simulations of human behavior as a source of input to decision making. Below are 10 representative examples that I pulled from the 2024 inventory:

1. One example is the Department of Homeland Security’s (DHS) Planning Assistant for Resilient Communities (PARC). The purpose of this program is creating “AI assistants in generating actionable, efficient plans to improve community preparedness.” The AI output for this program is a “Beta Release GenAI Plan Generator (based on OpenAI Models) produces hazards and mitigation plans.”

2. Another example is of the Department of Health and Human Services' (HHS) Risk Assessment Model (RAM) for the National Risk Index. The purpose of this program is “in tracking program progress and engaging stakeholders in risk assessments.” The AI output for this program is “reporting tools outputs for monitoring and visualization of risk data.”

3. The DHS has another project called AI/LLM to Generate Testable Synthetic Data. The purpose of this program is to create “test data to simulate real-world scenarios” and enhance “the ability to validate and test AI models efficiently.” Its AI outputs include generating “synthetic test data that mimics human behavior or scenarios.”

4. HHS has another project called Using Generative AI for Stance Analysis of Public Inputs. The purpose of this project is to simulate “public sentiment and stances of policy issues based on input analysis.” The AI outputs include “summary insights and stance classifications.”

5. HHS has another project called Reddit Post Analysis for Sexual Health Using LLMs. The purpose of this project is to simulate “patterns in user behavior to identify trends in sexual health discussions” and to provide “actionable insights for public health outreach.” The AI outputs of this project include “analyzed trends and topic classification based on Reddit discussions.”

6. DHS has another project called Human Interaction Simulation in Emergency Response. The purpose of this project is to model “human decision-making in emergencies for training and scenario analysis.” The AI outputs are “simulated decision trees and response pathways.”

7. HHS has another project called Behavioral Simulation for Public Health Messaging. The purpose of this projects is to simulate “human behavior to evaluate the impact of public health campaigns” and refine “messaging strategies to improve outreach and effectiveness.” The AI outputs are “behavioral response models and recommendations for messaging improvements.”

8. The Department of Housing and Urban Development (HUD) has a project called Community Engagement Modeling Tool. The purpose of this project is to simulate “community feedback to predict engagement levels in housing projects.” The AI outputs are “engagement metrics and recommendations for community outreach.”

9. The Department of the Treasury has a program called Virtual Environment for Policy Simulation. The purpose of this project is to simulate “economic and policy scenarios to assess potential impacts.” The AI outputs include “policy outcome predictions and economic impact assessments.”

10. The United States Agency for International Development (USAID) has a program called AI-Driven Social Behavior Modeling for Diplomacy. The purpose of this program is to simulate “social behavior to improve diplomatic strategies and international collaboration.” The AI outputs include “behavior models and insights for diplomatic decision-making.”

While none of these examples involve voting, we see several examples of LLM systems being deployed in various of the application areas identified in the comprehensive survey examined earlier. These applications, from the list of previously identified 10 “application areas” from the comprehensive survey include: policy emulation, participatory design, text analysis, public deliberation, social simulations, and economic modeling. In future work, it would be useful to assess the effectiveness of some of these application areas as evidence that complements the effectiveness of voting tools.

Conclusion

This living literature review has explored how technology has altered democratic input. We began with the example of voting. Voting is particularly instructive because while it has changed significantly overtime, the basic mechanics have remained roughly the same for at least 150 years in the US. Essentially, citizens are given the opportunity to vote for a representative at somewhat regular integrals. Tallies are taken from paper and digital ballots and a winner is declared.

However, voting is not the only method of democratic input. We explored a systematic account of the digital tools that enable additional forms of these inputs to democracy. These digital tools were categorized by their “genes” (goals, provider, and function). These tools provide new pathways to provide feedback to the government. However, one of the key findings was a relative absence of feedback loops that allow for a genuine process of integration of feedback by the government itself. This account also is clear in its lack of attention to advanced AI systems.

Following this, we explored a comprehensive survey of the ways in which advanced AI (in particular LLMs) is already being deployed towards politics and democratic input. This comprehensive survey identified at least 10 application areas for LLMs including: document classification, policymaking drafting and analysis, policy emulation, participatory policy design, text analysis, persuasive messaging, election simulation, public deliberation, social simulations, and economic modeling.

From here, we turned to two specific studies from the comprehensive survey that explore voting and policy preference elicitation in particular. Here we found attempts at using LLMs as digital twins at the individual level and election simulations at the societal level to assist in providing additional forms of democratic input. While both of these methods show promise in improving the quality and amount of democratic input, the low accuracy of digital twins capturing individual policy preferences and LLM biases in quality of prediction present major obstacles to the adoption of LLMs as voting mechanisms.

Finally, we explored the 2024 US Federal AI Use Case Inventory to examine what forms of democratic input and citizen participation were being created, trialed, and implemented by US Federal Agencies. Here we found over 1,700 use cases with at least 200 of them explicitly working with advanced AI systems such as LLMs. From within this set of 200 cases, we identified 10 specific use cases that focused on democratic input and citizen participation which included: planning assistants, risk assessment models, generate testable synthetic data, stance analysis, emergency simulation, behavioral simulation, and policy simulation.

Taken together, we see a clear co-evolutionary relationship between technological innovation and democratic input. In particular the internet and various digital tools have enabled a myriad of new forms of democratic input, however much use of these early digital tools has not been fully integrated into a continuous feedback process. LLMs in particular seem poised to apply general purpose tools to democratic input. While not quite ready to provide individual digital twins, the comprehensive survey identified at least 10 general application areas where LLMs are beginning to be used by governments as a source of democratic input. Additionally, all sorts of AI-tools are being created and trialed by the US Federal Government to make further use of AI to identify individual and collective preferences.

Finally, as AI capabilities continue to dramatically increase, one should expect that the ability to accurately ascertain individual policy preferences and accurately aggregate them will continue to improve. As these capabilities arise, they will present important questions as to whether democratic representation is best instantiated in the form of elected human representatives, or if improved forms of direct democracy can become more effective, efficient, and equitable at delivering the benefits from democratic forms of governance.

References

Aoki, G. (2024). Large Language Models in Politics and Democracy: A Comprehensive Survey (Version 1). arXiv. https://doi.org/10.48550/ARXIV.2412.04498

Gudiño-Rosero, J., Grandi, U., & Hidalgo, C. A. (2024). Large Language Models (LLMs) as Agents for Augmented Democracy. https://doi.org/10.48550/ARXIV.2405.03452

Majumdar, S., Elkind, E., & Pournaras, E. (2024). Generative AI Voting: Fair Collective Choice is Resilient to LLM Biases and Inconsistencies (No. arXiv:2406.11871). arXiv. https://doi.org/10.48550/arXiv.2406.11871

Shin, B., Floch, J., Rask, M., Bæck, P., Edgar, C., Berditchevskaia, A., Mesure, P., & Branlat, M. (2024). A systematic analysis of digital tools for citizen participation. Government Information Quarterly, 41(3), 101954. https://doi.org/10.1016/j.giq.2024.101954

1

Governing with AI

From Bureaucracy to Bytes and Back

Introduction…

Read more

a year ago · 4 likes · Justin Bullock

Introduction

This is the first entry in the Governing with AI Living Literature Review. In this entry I trace the evolutionary history of the bureaucracy, with a specific eye towards the bureaucracy’s relationship with automation and digitalization. As you will see, bureaucracies have begun evolving from the classic “street-level” bureaucracy to become “screen-level” bureaucracies — organizations infiltrated by digital computation — and on to “system-level” bureaucracies where algorithms have begun to play a larger role in both government decision making and in government action.

I begin with a brief account of the “ideal bureaucracy” type as described by Max Weber. From here, I trace our evolving understanding of bureaucracies from Herbert Simon’s “administrative behavior” lens, and on to the more recent discussions around digital discretion, artificial discretion, and artificial bureaucrats. Finally, I close with some reflections on what this literature suggests for the future of bureaucracies in a world where advanced AI continues to proliferate throughout the governance ecosystem.

From Bureaucracy

To govern, to steer, to control, one must have an administrative apparatus, something that administers power. Pharaohs, Emperors, Kings and Queens understood this. And now, Presidents, Prime Ministers, and yes, Kings and Queens still understand this. Throughout history, and until present day, it has been bureaucracies, wielded by governments to conduct public administration (and war), that have shaped the countless lives of the people living under the rule of those governments.

For many, “BUREAUCRACY” is a dirty word. And, fair enough! Bureaucracies have done, and continue to do, some dirty things. They even sometimes spread ADMINISTRATIVE EVIL. But this is just one side of the coin that is bureaucracy. Just as with markets, bureaucracies do bring about harm, but, as with markets, bureaucracies are necessary for modern life, and so many of the social and physical goods that are found across the globe. The bureaucracy is one serious piece of the backbone of both modern and historical civilization.

While the basic concept of a bureaucracy has been practiced at least as long as we’ve had writing, record keeping, and governments, it was not considered a formal object of modern study until Max Weber’s Economy and Society was first published in 1921. Weber had this notion of “ideal type” which he applied to describing the structure of bureaucracy. For Weber, the ideal type of bureaucracy included:

1. Well-defined and organized competencies.

2. A hierarchical organizational structure, with codified channels for making and

   transmitting information and decisions.

3. Modern administration that is based on:

   1. documents preserved as original copies or concepts, and

   2. a bureau that includes a staff of subordinated bureaucrats and writers of all kinds, working in the office, together with relevant resources including material goods and documents.

4. The work of the bureaucrat usually requires significant training for the specific tasks to be accomplished.

5. A full-fledged position occupies all the professional energy of the bureaucrat to process its tasks, regardless of limits to his mandatory working hours.

6. The duties of the position undertaken by the bureaucrat are based on general learnable rules and regulations, which are more or less firm and more or less comprehensible. The knowledge of these rules and regulations thus constitutes a special kind of ‘applied science,’ which the bureaucrat possesses.

While Weber had a particular view of bureaucracy, it was Herbert Simon (1947;1997) who offered an account of bureaucracy through the lens of information processing rather than describing the desired ideal bureaucratic structure that Weber offered.

For anyone who knows a little bit about the rest of Simon’s career, it should not be surprising that this was his take. Simon also pioneered the field of AI, which, at its core, is intimately concerned with information processing. Later in his career, in the final edition of his classic work “Administrative Behavior” (1997), Simon even more directly tied the rise of computational power (as a source of information processing) to the workings of administrative behavior of administrative organizations.

Enter the Bytes.

Simon noted that, even in 1997, digital computation had already begun to play a major role in how organizations processed information. Simon astutely noted that we were moving from a world where information itself was the bottleneck to high-quality organizational decision making to one where the attention of administrators themselves would become the bottleneck. The trick, it turned out, was still to identify the relevant information that needed to be applied to a particular decision. This was the beginning of a deep encroachment of bytes into bureaucracy.

A mere 5 years later, bureaucracy scholars Bovens & Zouridis, laid out, in detail, what they saw as digital computation’s increasing presence within public administration and what this signaled for the evolving shape of bureaucracies themselves. Bovens & Zouridis (2002) argued that bureaucracies were on an evolutionary pathway from “street-level” bureaucracies (those that might resemble Weber’s classic conceptualization) on to “screen-level” bureaucracies, and were rocketing towards a new form of bureaucracy called “system-level” bureaucracy. Screen-level bureaucracies described an evolutionary type where human bureaucrats began to see much of their work mediated through the interface screen of their personal computer monitors. These authors observed that much administrative work was beginning to be completed while staring at these monitors and entering digital data into a variety of digital databases.

However, Bovens & Zouridis could see that these screen-level bureaucracies were not the final evolutionary type. A simple extrapolation of the trend in utility and usage of both digital computation and algorithmic, automated analysis would show that these tools were likely to increase their presence within bureaucracies. They called the next evolutionary type of bureaucracy the “system-level” bureaucracy. In these bureaucracies, the role of automation, computation, and AI would further supplant the role of the individual discretion of the human bureaucrat. That is, while screens had augmented the capabilities of individual bureaucrats, in a system-level bureaucracy AI would automate much of the decision-making process, the information processing, within bureaucracies. Human discretion would be replaced by digital discretion for completing countless organizational tasks, and from our vantage point in 2024, we can see this was quite prescient.

But what does the proliferation of digital discretion mean for public administration?

This has been an area of inquiry within public administration for the last decade or so. The most well known piece on the topic is Busch and Henriksen’s (2018) “Digital Discretion: : A systematic literature review of ICT and street-level discretion” In this piece, Busch and Henriksen survey the evidence and the commentary on how digital computation (see: information and communication technology of ICT) had begun restructuring the way in which decisions were being made within bureaucracies. In their conclusion, Busch and Henriksen state:

In this study, we report from 44 scholarly articles on ICT and street-level discretion. Societal problems such as increasing and more complex demands on public service provision, and errors and corruption are creating pressures on politicians and government officials to provide services of higher quality and in a more efficient manner. The review shows that the environment in which ICT is implemented and used is vital for understanding why digital discretion diffuses and how the impacts of it are. For certain types of street-level work such as mass transactional tasks, ICT has reduced or even eliminated the use of human judgment. Examples of mass transactional tasks are the handling of student grant loans and tax reports where the data is numerical and readily available for government agencies, and where decisions are made based on schematic rule sets. In other types of street-level work such as social work, the discretionary practices of street-level bureaucrats are influenced by ICT to a lesser degree or not influenced at all. Contextual explanations for the prevalence of digital discretion can be attributed to factors such as the degree of professionalization, formulation of rules, computer literacy, and the level of information richness required. The impact of digital discretion is less explored in types of street-level work in between these extremes which opens avenues for future research. Another promising area for future research seems to be the increasing use of advanced technology such as artificial intelligence. This technology is now to a considerable extent able to deal with tasks of high complexity, and can thus address many of the shortcomings that the critics of digital discretion put forward…….

We conclude the literature review with claiming that the scope of street-level bureaucracy is decreasing. While certain types of street-level work seem to avoid extensive changes due ICT, it makes more and more sense to talk about digital bureaucracy and digital discretion since an increasing number of street-level bureaucracies are characterized by digital bureaucrats who operate computers instead of interacting face-to-face with their clients.

Busch and Henriksen were observing a change in the literature. Human discretion was beginning to be replaced by digital discretion. For more “standardized” tasks, those identified as “mass transactional tasks” by these authors, digital discretion was beginning to replace human discretion. However, for more complex tasks, less routine tasks, tasks where professional norms and judgements are needed digital discretion still fell short of human discretion. However, Busch and Henriksen, even in 2018, saw that AI may eventually be able to handle these more complex cases and overcome the limitations of digital discretion.

One additional study from 2018 points at a particular case study worth briefly exploring, Hansen, Lundberg, and Syltevik’s “Digitalization, street‐level bureaucracy and welfare users' experiences.” They explore the case of NAV:

Our case concerns the Norwegian Welfare and Labour Organization (NAV). In addition to the need to seek a more empirically grounded understanding of ICT in welfare bureaucracies, we focus on NAV for three reasons. First, NAV is one of several welfare bureaucracies around the world that has implemented ICT on a broad basis. The NAV organization is at the forefront of these developments, and it faces the same problems as the others. Second, as is the case for welfare agencies in other Scandinavian states, NAV is a large and important part of Norwegian society. Approximately one-third of the Norwegian state budget is spent by NAV on services and benefits provided to citizens. In any given year, NAV serves half of the Norwegian population and manages over three million cases. Some receive benefits on the basis of easily verifiable criteria without face-to-face interaction, whereas others struggle with problems that require frequent interaction with NAV over a longer time. How individuals manage their situation and their experiences may vary depending on factors such as their education, age, language and other skills. Thus, NAV is an excellent site for exploring both the limitations and possibilities of new technology for service users. Third, NAV is an interesting case because the Norwegian population is at the forefront, worldwide, of the use of new web technology. Sixty-seven per cent of the population are daily users of Facebook (TNS Gallup 2013), 97 per cent have internet access, and 80 per cent use the web daily (Vaage 2013). Thus, when NAV adopts new technological devices, it does so under what could be described as the best possible circumstances.

In this case, while exploring the evolutionary case of NAV from ~2008-2016, the authors found that the transformation to screen- and system-level bureaucracies had not (yet) taken place at NAV.

The effect on the relationship between users and the organization is complex. Bovens and Zouridis’ (2002) vision of screen- and system-level bureaucracy has not been a characteristic of the development of NAV so far. This is consistent with research in other countries, which shows only marginal shifts towards system-level bureaucracy (cf. Buffat 2015; Reddick 2005). NAV is a street-level bureaucracy that has some features of a screen-level bureaucracy for some groups. Currently, new technology has not replaced face-to-face encounters. This may be explained by the recentness of these ICT developments and the difficulty of developing well-functioning ICT systems, and because NAV still offers other contact channels. If this is altered, the effect of the emerging digital divide on access to welfare services and benefits could be strengthened. Until now, what we have seen for users with more complex dealings with NAV is more of a hybrid nature, combining digital and traditional communication. In some users’ dealings with NAV, ICT may reduce face-to-face contact, while for other users there is supposed to be more face-to-face contact. One example of this is the process of making individual activity plans that require personal contact. Such plans are an important tool in activation programmes in Norway.

The NAV case highlights that even under somewhat ideal circumstances (population at the forefront of technological adoption), NAV had not evolved into a full screen- or system-level bureaucracy. That, instead it was better still thought of as a street-level bureaucracy that had begun combining both digital and traditional (face-to-face) communication.

Additionally, In 2022, Henriksen coauthoring with a different colleague, Ranerup offers another case study to explore in the paper titled “Digital discretion: Unpacking human and technological agency in automated decision making in Sweden’s social services”. In this article Ranerup and Henriksen describe the case of the Trelleborg Model.

Trelleborg, the first municipality in Sweden to use automated decision making for social assistance decisions. This innovation project, the Trelleborg Model, is a management model now used in many other municipalities in Sweden. Trelleborg, Sweden’s southernmost town, is an industrial town of around 45,000 people.

Below is Figure 1 taken directly from Ranerup & Henriksen’s (2022) “Digital discretion: Unpacking human and technological agency in automated decision making in Sweden’s social services.” This Figure 1, from their article, highlights the way in which the citizen interacts both with caseworkers (human discretion) and a digital application and algorithmic decision (digital discretion). Here we also see something of a hybrid approach where human and digital discretion are applied at different places in the decision making process.

In this hybrid process described above, the authors find that the elements of digital discretion plausibly lead to improvements in the achievement of ethical values, democratic values, and professional values. However, the limitation here is that digital discretion is only being applied to standardized, and routine decisions. Despite this the article finds that the Trelleborg Model still provides for improvements on : 1) avoiding unethical actions and corruption, 2) adherence to rules and procedures that can be codified, 3) empowering citizens by giving them a more active role in decisions affecting their lives and rights, 4) increasing governments’ political legitimacy, 5) increasing decision-making accountability, 6) improvements in the quality of decisions, 7) increased efficiency, 8) preventing errors, and 9) reducing the cost of case management.

Taken together, this suggests that while digital discretion has not transformed Trelleborg’s approach to welfare service delivery, the deployment of digital discretion has provided many improvements to service delivery. The organizations covered in these cases had not been radically transformed by the introduction of digital discretion, but, while human discretion may be waning for some task sets, digital discretion remained restricted to making improvements on tasks where little discretion is required and the decisions to be made are routine and standardized, as is often the case with social insurance benefit programs.

With digital discretion being deployed, technological advancement carries on. AI, in particular, has seen a dramatic leap in capabilities. As AI systems did begin to show capability improvements, they began to receive more attention in the literature.

The Rise of Artificial Discretion and Artificial Bureaucrats

Concurrent with the discussion on digital discretion, there also developed a line of inquiry into artificial discretion. While digital discretion is a slightly more broad concept, artificial discretion explores how AI in particular is deployed to make decisions and pursue actions within the bureaucracy. This concept was introduced into the literature in 2019 in a paper titled “Artificial discretion as a tool of governance: a framework for understanding the impact of artificial intelligence on public administration.” In this paper, Young, Bullock, and Lecy extensively argue that while AI may be applied to a wide array of tasks within governments, there are a number of important issues within public administration that bureaucrats need to consider before widely deploying these tools.

In this article, the authors identify their different uses of AI based upon the level of discretion required for completing that task. The authors argue that for tasks that require low discretion, complete automation may be the most appropriate use of AI. For medium discretion tasks, AI can be used as a decision-support tool and for predictive analytics. Finally for tasks that require high discretion, AI can be used for new data generation, reduction of data complexity, and relationship discovery. Table 1 from this paper, which summarizes this relationship between AI use and discretion required, is provided below.

In addition to this general relationship between the use of AI and discretion required, the authors also examined the use of AI at different levels of organizational analysis. These levels are the micro-level (individual decisions), meso-level (organizational decisions), and the macro-level (institutional decisions). The authors give examples of each of these tasks in the Table 2 from their article, which is included below.

The following year two of these colleagues Young and Bullock with another colleague Wang explored two cases of how AI was reshaping not only how individual decisions were made but also the evolution of bureaucracies as well. In “Artificial intelligence, Bureaucratic Form, and Discretion in Public Service” (2020) these colleagues looked at the two cases of policing and insurance processing. Their argument is that different policy domains contain different bundles of tasks that require relatively more or relatively less discretion. In this case policing is designated as often requiring medium to high levels of discretion while insurance processing tasks generally require lower levels of discretion. The authors go on to use this reasoning to explore changes in both the ratio of human to artificial discretion and the magnitude of change in the bureaucratic form. The findings are summarized in the authors’ Table 2, which is included below.

Back to Bureaucracy

As this entry in the living literature review has highlighted, what has traditionally been considered as AI is already at work within and throughout bureaucracies, placing evolutionary pressures upon them. But several key considerations are still missing from the current story of governing with AI. For example, currently, we are in the midst of a gigantic leap in the capabilities and generality of AI systems. The progress over the past 10 years has been well documented. This progress was punctuated on November, 30th of 2022 with the release of ChatGPT.

The question is: what does this mean for bureaucracies? What does it mean for “governing with AI”?

This will be a consistent focus for following entries in this living literature review.

So, when we turn “back to bureaucracy” itself, in large part what we’re considering here is, in an age of rapidly improving AI capabilities, how will the government administrative apparatus itself administer power?

Will bureaucracies remain as the relevant organizational type?

Conclusion

This first entry in the Governing with AI Living Literature Review has explored the evolution of AI within Public Administration. An understanding of this history is needed to understand how frontier AI systems will result in a further evolution of bureaucracies and public administration.

Bureaucracies have a long history in administering government power. Max Weber described the “ideal type” bureaucracy in 1921. In 1947, Herbert Simon added to this description by putting a focus on information processing by bureaucracies. By 1997, Simon saw how digital computation was beginning to infiltrate the bureaucracy. By 2002, Mark Bovens and Stavros Zouridis had begun to see a deep evolution of bureaucracies from the classic street-level bureaucracies on to screen-level and system-level bureaucracies. In 2018, Peter Busch and Helle Henriksen observed that as part of this evolution, human discretion was being replaced and augmented by digital discretion. By 2019, Matthew Young, Justin Bullock, and Jesse Lecy were exploring under what circumstances artificial discretion should be used by bureaucracies.

But, in 2022, ChatGPT was released and generative pre-trained transformer AI models burst onto the public scene. These “foundation models” heralded a new age of AI, characterized by increasing intelligence and generality. And while this entry in the living literature review has set the historical stage for the entrance of these frontier AI systems, the next several entries will explore the direct impact of these systems on how governments administer power.

References

Bovens, M., & Zouridis, S. (2002). From Street‐Level to System‐Level Bureaucracies: How Information and Communication Technology is Transforming Administrative Discretion and Constitutional Control. Public Administration Review, 62(2), 174–184. https://doi.org/10.1111/0033-3352.00168

Buffat, A. (2015). Street-Level Bureaucracy and E-Government. Public Management Review, 17(1), 149–161. https://doi.org/10.1080/14719037.2013.771699

Bullock, J., Young, M. M., & Wang, Y.-F. (2020). Artificial intelligence, bureaucratic form, and discretion in public service. Information Polity, 25(4), 491–506. https://doi.org/10.3233/IP-200223

Busch, P. A., & Henriksen, H. Z. (2018). Digital discretion: A systematic literature review of ICT and street-level discretion. Information Polity, 23(1), 3–28. https://doi.org/10.3233/IP-170050

Hansen, H., Lundberg, K., & Syltevik, L. J. (2018). Digitalization, Street‐Level Bureaucracy and Welfare Users’ Experiences. Social Policy & Administration, 52(1), 67–90. https://doi.org/10.1111/spol.12283

Ranerup, A., & Henriksen, H. Z. (2022). Digital Discretion: Unpacking Human and Technological Agency in Automated Decision Making in Sweden’s Social Services. Social Science Computer Review, 40(2), 445–461. https://doi.org/10.1177/0894439320980434

Reddick, C. G. (2005). Citizen interaction with e-government: From the streets to servers? Government Information Quarterly, 22(1), 38–57.

Simon, H. A. (1997). Administrative Behavior (3rd ed.). Free Press.

Simon, H. A. (1947). Administrative Behavior (1st ed.). Macmillan.

Weber, M. (1978). Economy and society: An outline of interpretive sociology (Vol. 1). Univ of California Press.

Young, M. M., Bullock, J. B., & Lecy, J. D. (2019). Artificial Discretion as a Tool of Governance: A Framework for Understanding the Impact of Artificial Intelligence on Public Administration. Perspectives on Public Management and Governance, gvz014. https://doi.org/10.1093/ppmgov/gvz014

I’m excited to announce the launch of Governing with AI: A Living Literature Review.

There is a growing conversation about how Artificial Intelligence (AI) is impacting how governments are held accountable, provide goods and services, coordinate to solve problems, and receive democratic input. At the same time, AI capabilities themselves are quickly improving, changing what is possible for both AI and governments to do. Additional advancements in capabilities will further impact not only how governments decide, act, and implement things, but also provide for new forms of governance. These fast-paced changes present immense opportunities and challenges for humanity. Keeping up with these changes, and their consequences, is a challenge all by itself.

Enter: Governing with AI. The Living Literature Review.

Once a month I will write a concise (and hopefully engaging!) literature review on some aspect of how AI is impacting governance. Here are a three topics, I’ve already begun researching (with thanks to Claude for the creative titles):

1. From Bureaucracy to Bytes and Back: AI's March Through Public Administration

2. AI vs. Global Crises: Computational Solutions for Planetary Problems

3. Digital Democracy 2.0: AI as the New Frontier of Civic Engagement

Each of these reviews will be ~3,000 words and integrate findings from peer-reviewed articles, pre-prints from ArXiv and SSRN, official government reports, and other high-quality sources. Each review will be posted to the Governing with AI website and will be emailed directly to subscribers in newsletter format. Additionally, several times a year I will update a review and post it as a new entry (while keeping the original review archived on the website).

I hope you’ll subscribe, follow along, share with others, and comment.

If you haven’t already you can subscribe now by clicking the button below.

Subscribe now

You can expect the first literature review to be published mid-November!

P.S. Shout outs to Seb Krier for creating several engaging post banners, and to Richard Ngo for placing the idea of the “Governing with AI” title into my brain in an excellent Twitter/X thread on AI safety where Richard actually uses the phrase “governance with AI”.