The Knowledge Barrier – Methodology

A.1 Primary Data: The ICMA 2024 Survey

Survey Design

The empirical foundation for this book is the ICMA 2024 Survey on Artificial Intelligence in Local Government, administered by the International City/County Management Association in collaboration with the Urban Institute. The instrument included:

Closed-ended questions measuring AI priority levels (q1, four-point scale), current utilization by service area (q2, seven categories), perceived potential (q3, matching categories), governance actions taken (q4, six action types), barriers to adoption (q5, seven categories), and concerns about AI implications (q6, seven categories).
Open-ended questions inviting respondents to describe implementation experiences, organizational challenges, and practitioner perspectives in their own words.

The survey employed checkbox (select-all-that-apply) format for barrier, concern, governance, and service-area questions. This format allows respondents to indicate multiple responses but does not capture intensity, ranking, or relative importance among selected items. When this book reports that “77 percent cite lack of awareness/understanding,” this means 77 percent selected that option—not that they ranked it as their most important barrier.

Sample Characteristics

Characteristic	Value
Total respondents	635
Primary respondent type	Chief administrative officers and assistant CAOs (78%)
Population distribution	41% from communities between 5,000 and 25,000
Government form	Council-manager well represented (67%)
Geographic coverage	All four Census regions

Respondent Demographics

The survey captures leadership perspectives. The dominant respondent profile—CAOs in council-manager governments serving mid-sized communities—shapes every finding. When the survey reports that 12 percent cite workforce displacement as a concern, this reflects what leaders think about worker impacts, not what workers experience. The council-manager overrepresentation is a feature of ICMA’s membership composition, not a survey design flaw.

A.2 Limitations

Small Community Underrepresentation

Only 10 percent of respondents came from jurisdictions under 2,500 population. The United States has approximately 19,500 municipalities below 5,000 residents. Documented disparities may understate actual disparities.

Self-Selection Bias

Respondents chose to complete the survey, potentially overrepresenting those with greater interest in or awareness of AI. This bias likely produces conservative estimates of the knowledge barrier and potentially inflated estimates of AI adoption and governance activity.

Social Desirability in Barrier Reporting

“Lack of knowledge” is a blame-free response—it does not implicate leadership failures, budget mismanagement, or political dysfunction. Bergen and Labonte (2020) documented how survey respondents systematically gravitate toward socially safe answers.

Overlapping Barrier Categories

The barrier categories are not mutually exclusive. Knowledge, personnel, policy, and funding barriers are interrelated. The checkbox format does not capture these interdependencies.

Point-in-Time Snapshot

The survey captures attitudes and practices at a specific moment in a rapidly evolving field. The data represents a slice of a trajectory—valuable as a baseline, but subject to rapid obsolescence in its specific statistics.

Currency of Secondary Sources

Several key sources carry 2025 publication dates. Because AI-related research is evolving rapidly, readers should verify current availability and check whether updated versions have been published.

A.3 Data Analysis

Quantitative analysis was conducted using Python (pandas, numpy) with visualizations generated using matplotlib and seaborn. Analysis included:

Frequency distributions for all survey variables
Cross-tabulations by population size, region, government type, government form, and respondent role
Opportunity gap calculations comparing current utilization to perceived potential by service area
Population size disparity analysis across community size categories

Statistical Approach

The analysis is primarily descriptive. The ICMA survey was designed as a practitioner survey, not a probability sample, and the self-selected response pool does not support inferential statistics with known confidence intervals. This book does not report p-values, confidence intervals, or tests of statistical significance for the ICMA data—not because the data lacks value, but because the sampling frame does not support such inferences.

Qualitative Analysis

Open-ended survey responses were analyzed thematically—identifying recurring patterns, distinctive perspectives, and outlier viewpoints—rather than through formal qualitative coding.

A.4 Secondary Sources

Peer-Reviewed Research

Hatz et al. (2025) — Two-wave panel survey of 1,028 local elected officials.
Noy & Zhang (2023) — Controlled experiment documenting productivity and quality effects of ChatGPT.
Dell’Acqua et al. (2023) — Field experiment with 758 management consultants.
Becker et al. (2025) — Controlled study finding experienced developers 19% slower with AI.
Alon-Barkat & Busuioc (2023) — Study of automation bias among public sector decision-makers.
Feiock (2013) — Theoretical framework for cooperation challenges in fragmented governance systems.

Research Institution Reports

Urban Institute (Prinvil, Stern, & Ramos, 2025) — Three-tier framework for AI applications.
Deloitte (Chaurasia et al., 2025) — State and local government AI adoption patterns.
Roosevelt Institute (Shorey, 2025) — Worker experiences with AI in public sector.
MissionSquare (2024, 2025) — State and local government workforce surveys.
Tony Blair Institute (2024) — AI potential for local government, PEARS framework.

Historical and Comparative Sources

Moon (2002) — E-government adoption among municipalities.
Norris & Reddick (2013) — Gap between technology rhetoric and adoption reality.
Nedovic-Budic & Godschalk (1996) — GIS adoption documenting knowledge-barrier pattern.
Rogers (2003) — Diffusion of innovations framework.

A.5 Replication

The ICMA survey dataset is available through ICMA’s survey research program. The Python code used for quantitative analysis is available upon request from the author. Specific analytical decisions (variable coding, cross-tabulation categories, population size breakpoints) are documented in the analysis code.