1. Introduction: The Pacing Problem at the State Level
BY CAITLIN MCNALLY, ANNE MEEKER, AND AUBREY WILSON
1A — The Pacing Problem and GenAI’s Impact on Representative Bodies
American state legislatures have long been called the “laboratories of democracy,” a phrase coined by Supreme Court Justice Louis Brandeis in 1932 to describe the states as places where new approaches to governance could be tested.¹ Nearly a century later, the framing still holds: a substantial share of the procedural and policy innovations now shaping American government, from professionalized legislative research services to oversight tools, originate at the state level.
Yet the laboratories themselves have modernized unevenly, and the gaps between the innovators and the lagging are more pronounced as AI accelerates the speed of technological and societal change. Many governors’ offices have invested steadily in technology, staff capacity, and operational infrastructure over the past twenty years.² By contrast, legislatures have made uneven progress, and in many states they are less prepared for responding to and adopting a technology that is reshaping how governments work.
Over the last six years, POPVOX Foundation, a 501(c)(3) nonprofit organization founded by former legislative staffers, has focused on addressing this growing “pacing problem,” a term coined by legal scholar Gary Marchant to describe the growing gap between emerging technologies and the governance systems meant to keep pace with them.³ Representative Bodies in the AI Era: Volume I, a report published by POPVOX Foundation’s Executive Director Marci Harris and Managing Director Aubrey Wilson in January 2024, traced this gap as a structural challenge of representative bodies.⁴ Harris and Wilson documented how Congress and parliaments around the world had begun adapting to artificial intelligence, made recommendations for how legislative bodies can responsibly explore and deploy AI while preserving democratic governance, and noted subnational state and local AI uptake as a thread worth following.
In the period since the Representative Bodies report, generative AI has moved from an emerging technology to a routine input across legislative workflows. POPVOX Foundation’s Digital Parliaments Project now convenes partner parliaments across the Caribbean, the Balkans, and Africa around shared infrastructure and best practices.⁵ The New York City Council has built an internal digital architecture that uses AI for duplicate-detection across legislative proposals, retrieval-augmented research across institutional legal memos, and automated tracking of agency compliance with reporting requirements.⁶ Brazil’s Senate has connected its e-Cidadania citizen-input platform to legislative drafting through an AI tool that surfaces relevant public submissions to consultants drafting bills,⁷ and in Estonia, a former government CIO built a civic AI tool in a single day that scans every draft bill for inconsistencies — sparking a national conversation about whether legislative quality control should sit inside parliament, outside it, or both.⁸ The effects of the pacing problem have accelerated: it is now playing out simultaneously across national parliaments, subnational legislatures, and civil society, and the state-level layer of US lawmaking is no different.
A few years after ChatGPT’s general release and Representative Bodies’ publication, the laboratories of democracy are the logical next place to examine the pacing problem up close. This report returns to that thread. Drawing on interviews with state legislators and staff in twenty-two states, it asks where state-level AI adoption stands relative to the expectations Representative Bodies set — and what the capacity conditions of state legislatures explain about the answer.
1B — What “Capacity” Means in This Report
Legislative capacity, at its simplest, is a legislature’s ability to do its job: to draft and analyze legislation, oversee the executive, serve constituents, and absorb the information and demands flowing into the institution. POPVOX Foundation’s work on Congress and on parliaments globally approaches capacity as more than headcount and budget, treating the knowledge, relationships, and infrastructure that make legislative work possible as capacity in their own right. This expanded definition matters even more at the state level, where some of the most consistent capacity pressures documented in this research surface not in staff rosters but in the workarounds built to compensate for them: caucus structures that serve as force multipliers for thinly-staffed offices, informal mentorship that backfills the lack of formal onboarding, partnerships with academic institutions and civil society organizations that carry functions the legislature has not built internally. State legislative capacity is more constrained than aggregate staff counts and session-length comparisons suggest, and at the same time more elastic than those measures recognize.
When this report refers to legislative capacity, it refers to four dimensions:
People
The human side of legislative work: staffing levels and how they are distributed within and between chambers, office budgets and Member compensation, workload and its relationship to both, and the professional development infrastructure available to Members and staff. This dimension also includes the elastic capacity that formal rosters miss — campaign-supplemented staffing, intern programs, and caucus structures that concentrate expertise to serve many Members at once.
Institutional knowledge
How the institution retains and transmits what its people know. Much of what makes a legislature function lives in human memory rather than documented systems, transferred through proximity, mentorship, and informal practice. Onboarding quality, turnover patterns, and the depth of the single-expert problem — bodies of specialized knowledge held by one person with no mechanism for what happens when they leave — all shape whether knowledge survives departures or walks out the door with them.
The broader network
The relationships that extend past the institution’s formal borders: with the Executive branch that implements its laws and supplies much of its data, with peer legislatures it looks to for models and precedent, with national networks like National Conference of State Legislatures (NCSL) and the Council of State Governments (CSG), with civil society organizations and academic partners that fill research and training gaps, and with lobbyists who carry procedural knowledge of their own. These networks can multiply institutional capacity — or create dependencies that raise their own capacity questions.
Technology adoption
Data architecture, AI policy clarity, IT capacity, vendor procurement infrastructure, and the presence or absence of domain-constrained tools that can serve as trust bridges for broader adoption. This is the dimension where state legislatures are simultaneously farthest behind comparable institutions on some measures and, in several cases, ahead of Congress on specific practices.
This framework builds on the comparative political science literature on legislative capacity, particularly the Squire professionalization index⁹ and Bowen and Greene’s two-dimensional scaling,¹⁰ by adding new dimensions that the research documented in this report shows are necessary to understand state legislatures today. Section 5 returns to the comparative-measurement question in detail.
Legislative bodies at every level face what the Foundation has come to think of as fractal challenges related to capacity: similar problems of information overload, constituent volume, staff retention, and institutional knowledge transfer recur across Legislative branches at multiple jurisdictions and scales. State legislatures share many of those challenges, but not just as smaller, less professionalized versions of Congress. They each operate with their own structural conditions, political dynamics, and institutional histories, and often with less resources than their federal counterparts.
The vast majority of state legislatures fulfill their responsibilities with fewer resources than their federal counterparts. In fact, the demands on individual legislators in several states meet or exceed those facing some Congressional offices. A Texas Senator represents roughly one million constituents, more than a US federal House district, and manages that constituency with a fraction of the staff. A constituent services staffer in a Florida state Senate office reported handling roughly 3,000 cases per year, a volume that exceeds what some US Congressional district offices manage.¹¹ Beyond raw numbers, state legislatures also navigate structural constraints rarely faced by Congress: balanced budget requirements that shape every fiscal decision, strict bill introduction limits in states like California,¹² and an emerging field of subnational diplomacy in which states maintain their own relationships with foreign governments on issues ranging from trade to climate.¹³
This resource constraint sometimes produces stagnation, but it also drives inventiveness. States have moved first before when adapting to their limits, as in the twentieth-century push to professionalize legislatures.¹⁴ The same dynamic is now visible in AI adoption: several state legislatures have outpaced Congress on internal AI integration, and some legislative services offices have built frameworks for staff AI use more developed than anything out of Washington.
1C — The Vol. 1 AI Adoption Approaches and Phases Framework Applied to State Implementation
This volume is structured as a progress report against Representative Bodies, and adopts its framework for categorizing AI adoption approaches into four categories. Section 3b applies these categories directly, presenting a use-case inventory of state legislature-based AI implementation.
Four adoption approaches
Representative Bodies organized the ways legislatures can deploy AI into an ascending ladder of control, customization, and institutional investment:
third-party COTS applications — commercial off-the-shelf software such as Microsoft 365 or CRM platforms that incorporate AI into tools staff already use;
third-party GenAI applications — general-purpose commercial tools such as ChatGPT or Claude, used under institutional policy for non-sensitive work;
custom GenAI tools — applications built by or for the institution that leverage commercial models through APIs against the legislature’s own data; and
institutional large language models — models the institution maintains itself, the most resource-intensive option and one Representative Bodies judged largely out of reach at the time.
Four adoption phases, with timeframes
Representative Bodies recommended a phased sequence of institutional action to proactively respond to the continuing emergence of GenAI on the horizon in January 2024:
An immediate phase (roughly three months) focused on enabling experimentation, clarifying tool-use guidelines, fostering inter-office information sharing, and — presciently, as Section 3c will show — tracking constituent engagement for inauthentic AI-generated campaigns
A short-term phase (three to six months) of hearings, expert studies, and prototype development;
A medium-term phase (six months to two years) of investment in technical infrastructure to embed AI in routine processes
A long-term phase (two to five years) in which legislatures with the preceding foundations in place could revolutionize constituent engagement, lawmaking, and oversight with custom-built systems.
Today, the immediate, short-term, and medium-term timelines outlined in Representative Bodies have fully elapsed and the long-term window is roughly at its midpoint. However, Representative Bodies’ adoption phases — paired with the four adoption approaches — can be used as an evaluation guide for whether or not a legislative institution has kept pace with technology’s rate of change over the last two years. For example, at the use-case level: a domain-constrained platform embedded in daily workflows is operating in medium-term territory regardless of when it was built, while a Member experimenting with a general-purpose chatbot is doing immediate-phase work. Second, at the institutional level: by Representative Bodies’ timeline, most legislatures should by now have issued guidelines, commissioned studies, and begun infrastructure investment. A handful of state implementations have reached the medium-term marker, showing that individual Members or offices are already ahead of schedule. But most state legislatures have not yet completed the immediate-phase items, meaning the institution has not caught up to its own early adopters. That gap between a few advanced implementations and the institutions that host them is the central pattern this volume documents, and the pattern the Section 3b table makes visible at a glance.
Section 3b presents state legislative AI implementation use cases, mapped to Volume I’s AI adoption and phases framework.
Recommendations
Finally, Representative Bodies issued recommendations for legislative bodies, from initiating early and treating data as a strategic resource through ensuring human oversight and engaging in global collaboration. Section 6 presents these recommendations and expands upon them in the context of existing state capacity and AI adoption challenges, regrouped under three state-level priorities.
1D — Roadmap
The report proceeds in six substantive sections. Section 2 describes the research methodology. Section 3 examines the state of AI adoption in state legislatures. Section 4 adds context to those adoption examples — the people, knowledge, network, and technology conditions that determine which legislatures can adopt AI and how, pairing the pressures documented in each dimension with the creative adaptations states have developed in response. Section 5 argues that the mismatch between conventional capacity measures and the adoption patterns this research observed calls for new ways of assessing legislative capacity in the AI era. Section 6 offers recommendations adapted from Representative Bodies’ framework to the state context. Section 7 concludes with the open questions this research surfaced.
¹ New State Ice Co. v. Liebmann, 285 U.S. 262, 311 (1932) (Brandeis, J., dissenting).
² “The 2025 State CIO Survey: Leading Change Through Uncertain Times,” National Association of State Chief Information Officers (NASCIO) (2025).
³ Gary E. Marchant, “The Growing Gap Between Emerging Technologies and the Law,” The Growing Gap Between Emerging Technologies and Legal-Ethical Oversight: The Pacing Problem (2011).
⁴ Marci Harris and Aubrey Wilson, “Representative Bodies in the AI Era: Insights for Legislatures,” POPVOX Foundation (January 2024).
⁵ “Departure Dialogues Project,” POPVOX Foundation (n.d.).
⁶ Beatriz Rey, “Inside the NYC Council’s Policy ’Primeval Soup’,” Modern Parliament (May 20, 2026).
⁷ Beatriz Rey, “From Citizen Ideas to Bills,” Modern Parliament (February 17, 2026).
⁸ Beatriz Rey, “From Slip-Up to Solution: How AI Can Help Fix Lawmaking in Estonia,” Modern Parliament (January 27, 2026).
⁹ Peverill Squire, “A Squire Index Update: Stability and Change in Legislative Professionalization, 1979–2021,” State Politics & Policy Quarterly (2024).
¹⁰ Daniel C. Bowen and Zachary Greene, “Should We Measure Professionalism with an Index? A Note on Theory and Practice in State Legislative Professionalism Research,” State Politics & Policy Quarterly (2014).
¹¹ Anne Meeker, “Casework Navigator: Taking Care of Yourselves in 2025,” POPVOX Foundation (December 4, 2024).
¹² California State Assembly, Standing Rules of the Assembly, 2025–26 Regular Session, Rule 49; California State Senate, Standing Rules of the Senate, 2025–26 Regular Session, Rule 22.5.
¹³ “Report of the Truman Center City & State Diplomacy Task Force,” Truman Center for National Policy (2022).
¹⁴ Peverill Squire, “American State Legislatures in Historical Perspective,” PS: Political Science & Politics (2019).
