State officials who actually use AI see clear utility while ideology shapes risk worries: frequent users report greater benefits, whereas conservative respondents report lower perceived AI risks — yet individual attitudes do not reliably map onto whether agencies adopt AI.

Main Finding

There is a clear asymmetry in what shapes state-government professionals’ attitudes toward AI: hands‑on frequency of AI use strongly predicts perceived administrative benefits, while political ideology (conservatism vs. liberalism) is the dominant predictor of perceived governance risks. These two dimensions—perceived utility and perceived risk—are distinct and have different correlates. Individual attitudes, however, do not reliably predict reported organizational AI implementation (analysis limited by sample size).

Key Points

• Attitudes toward governmental AI are multidimensional:
  • Perceived benefits (utility) and perceived risks (bias, misinformation, accountability, privacy) are separate outcomes and can coexist in the same respondent.
• Predictors of perceived benefits:
  • Frequency of AI use is the strongest predictor: more frequent users report substantially higher confidence in AI’s governmental utility.
  • General familiarity with AI also predicts perceived benefits, but direct use adds explanatory power beyond familiarity.
• Predictors of perceived risks:
  • Political ideology is the dominant predictor: more conservative respondents report lower perceived risks from governmental AI.
  • Once ideology is included, the association between AI use and risk perception becomes nonsignificant.
• Robustness:
  • Results are robust to HC3 standard errors and influential‑observation sensitivity analyses.
• Organizational adoption:
  • Individual attitudes (utility or risk perceptions) do not significantly predict self‑reported organizational AI implementation in this sample; organizational and institutional constraints likely matter.
• Causality caveat:
  • Directionality between use and perceived benefit is ambiguous (pre-existing positive attitudes might increase use, yet use may also increase perceived utility).

Data & Methods

• Data:
  • Original survey of individuals with experience in U.S. state government (sample details such as N and sampling frame are not provided in the excerpt).
  • Key measures: frequency of AI use, self‑reported familiarity with AI, political ideology, perceived benefits of governmental AI, perceived risks of governmental AI, government experience, office size, and reported organizational AI implementation.
• Empirical approach:
  • Regression models predicting two distinct dependent variables: perceived benefits and perceived risks.
  • Nested model strategy: assessing incremental explanatory power of AI familiarity/use vs. ideology.
  • Controls include AI familiarity, political ideology, government experience, and office size.
  • Robustness checks: HC3 standard errors and influential‑observation sensitivity analyses.
• Limitations noted:
  • Cross‑sectional, observational data limit causal inference.
  • Adoption analysis constrained by sample size.
  • Reliance on self‑reports (attitudes and organizational implementation).

Implications for AI Economics

• Diffusion and adoption dynamics:
  • Frequency of use raising perceived benefits suggests path dependence and positive feedback: early users and pilots can change beliefs and stimulate further uptake. Economists modeling technology diffusion in the public sector should include agent‑level experiential feedback loops, not only informational or cost variables.
  • Training, exposure, and pilot programs are likely high‑leverage policy instruments to increase perceived productivity gains of public‑sector AI and accelerate beneficial adoption.
• Political economy and regulation:
  • Ideology driving risk perceptions implies that jurisdictions will differ systematically in appetite for regulation, oversight investment, and allowable uses of AI. This heterogeneity affects cross‑state policy competition, regulatory fragmentation, and the national political economy of AI governance.
  • Cost–benefit assessments and social-welfare models must account for politically heterogeneous valuations of risk (i.e., different discounting of governance harms), which may lead to differing optimal policies across jurisdictions.
• Welfare analysis and externalities:
  • Because perceived utility and perceived risk are separable, welfare evaluations should separately quantify efficiency gains (e.g., administrative productivity, service speed) and governance externalities (e.g., bias, misinformation, trust erosion). Aggregating into a single “preference” or scalar attitude risks mismeasuring tradeoffs.
• Institutional constraints and policy design:
  • Individual-level persuasion (changing beliefs via information/training) may be insufficient to change organizational adoption; institutional incentives, budgets, procurement rules, legal mandates, and professional norms matter. Economists advising public investment in AI should consider organizational frictions and principal–agent issues when estimating adoption elasticities.
• Research and measurement implications:
  • Empirical models of public‑sector AI deployment should incorporate:
  • Heterogeneous agent experience (use vs familiarity).
  • Political/ideological heterogeneity in risk valuation.
  • Organizational/institutional variables (size, budget, legal constraints).
  • Future economic work should aim for longitudinal or experimental designs to disentangle directionality between use and perceived benefits, and to quantify how experience‑induced belief changes affect subsequent adoption and performance.

Overall, the study suggests that to understand and predict AI uptake and its economic effects in the public sector, models must jointly address experiential learning effects (which raise perceived productivity) and political valuation of risk (which shapes governance choices), while accounting for organizational constraints that moderate the translation of individual attitudes into institutional action.