Testing the dimensionality of public attitudes toward artificial intelligence

Abstract

Public opinion about artificial intelligence (AI) is increasingly measured and invoked in debates about AI governance, yet heterogeneous survey items are often interpreted as indicators of one underlying attitude. Using a weighted survey of registered voters in the U.S. state of Pennsylvania ( N = 929), this article tests whether four AI attitude and policy items form a coherent dimension and examines the correlates of a nervous-to-excited affect item. A principal-component analysis of polychoric correlations in the complete-case subset ( N = 460) does not support a strong unidimensional scale. A companion sensitivity analysis of the three non-employment items on the larger available sample reproduces the same two-component pattern, so the negative scaling finding is not an artifact of the employment-contingent job item. A weak first component links excitement, lower job concern, and opposition to regulation, but reliability is low (ordinal alpha = 0.45; KMO = 0.48). An item asking whether the U.S. federal government is equipped to regulate AI loads separately and aligns more closely with evaluations of the named institution than with AI affect. Weighted regression with HC3 heteroskedasticity-robust standard errors, corroborated by multiple imputation with 20 datasets, shows that AI excitement is associated most strongly with personal AI use, followed by prospective economic optimism, age, and education. This ordering holds in both the listwise and multiply imputed specifications. The findings are bounded to this survey module, but they show why AI affect, behavioral exposure, policy preference, and institution-specific governance judgments should not be combined into an accidental scale without first establishing that they measure a common construct.