The problem Katz targets is what he calls the accidental scale. When a survey module asks respondents whether they are nervous or excited about AI, whether AI threatens their job, whether AI companies should be regulated, and whether the federal government is equipped to regulate the technology, analysts and journalists often average these answers into one index of public sentiment. The practice is tempting because the items are topically connected, but topical connection is not the same as measuring the same thing. A person can feel uneasy about AI, fear for their job, still want strict regulation, and doubt that Washington has the technical competence to deliver it. Those are four distinct judgments about four distinct objects.

To test whether the items really belonged together, Katz used principal-component analysis on polychoric correlations, a technique suited to ordinal survey responses, on the complete-case subset of 460 respondents who answered all four attitude and policy questions. The results were unambiguous. The Kaiser-Meyer-Olkin statistic, a standard measure of sampling adequacy for common-factor analysis, came in at 0.48, below the 0.50 threshold that Kaiser himself classified as unacceptable and well below the 0.60 conventionally required for adequate common structure. Two components had eigenvalues above one, and the government-capacity item loaded almost entirely on its own component, near 0.96, with negligible cross-loading on the sentiment component.

Reliability was equally poor. The ordinal alpha for the three sentiment-related items was just 0.45, far below the 0.60 floor typically considered inadequate for scale scoring. Oblique rotations using promax and direct-oblimin methods confirmed that the separation was not an artifact of forcing components to be orthogonal; inter-component correlations stayed below 0.07. A companion sensitivity analysis on the three non-employment items, estimated on a larger complete-case sample, reproduced the same two-component pattern, showing that the negative finding was not driven by who happened to be asked the job question. The weak first component that did emerge linked excitement, lower job concern, and opposition to regulation, but with reliability this low it cannot support meaningful scale scoring.

The government-capacity item deserves particular scrutiny. When respondents were asked whether the U.S. federal government is equipped to regulate AI if it decides to do so, their answers correlated more strongly with presidential approval and with a battery assessing administration performance than with the AI affect item. The correlation with the administration-performance battery was 0.28, and with presidential approval 0.25, compared with much weaker links to AI sentiment. This suggests the question is functioning partly as a measure of confidence in the incumbent government rather than as a pure indicator of attitudes toward artificial intelligence. In a partisan environment, asking whether a named institution can handle a named technology inevitably blends two objects of evaluation.

Having established that no composite index was defensible, Katz turned to the single item with the broadest theoretical reach: the nervous-to-excited question, which was asked of the full sample and captures a global affective appraisal of the kind risk-perception research treats as the broadest summary of orientation toward an unfamiliar hazard. Weighted regression with HC3 heteroskedasticity-robust standard errors, corroborated by multiple imputation across 20 datasets, revealed a clear hierarchy of predictors. Personal AI use was the strongest correlate, followed by prospective economic optimism, age, and education. The ordering held in both the listwise specification, where AI use carried a standardized coefficient of 0.33, and the multiply imputed specification, where it was 0.31.

The age result is especially instructive. At the zero-order level, age showed no significant association with AI excitement. But once AI use and other predictors were controlled, age became positively associated with excitement, a suppression pattern that appeared in both estimation strategies. Younger respondents use AI more frequently, and use is positively associated with excitement. Holding use constant, older respondents actually score somewhat higher on the affect item. The popular narrative that older people are uniformly more fearful of AI does not survive this multivariate scrutiny, and apparent generational differences may partly reflect unequal exposure rather than deep dispositional gaps.

The descriptive picture from the survey is itself revealing. About 70 percent of respondents were nervous rather than excited about recent AI advances, while roughly one in five sat on the excited side. Roughly two thirds reported using AI at least occasionally, but only 18 percent used it every day or most days. Support for closer federal regulation of AI companies was broad at near 78 percent, yet confidence that the federal government is equipped to regulate AI was low at near 19 percent. Among employed respondents, about one third worried that AI could threaten their job. These are not facets of a single mood; they are distinct evaluations that happen to share a topic.

Katz is careful to bound his claims. The analysis covers one survey module in one state, drawn from registered voters who skew older and more racially homogeneous than many national populations, and the cross-sectional design cannot establish whether use causes favorability or favorable respondents simply select into use. The affect outcome is a single bipolar item without a neutral category, which may compress ambivalence. The dimensionality result rests on complete cases, and while the three-item companion analysis addresses the most consequential selection concern, the finding should be read as a bounded negative result about this item set, not as proof that public attitudes toward AI are intrinsically fragmented. Purpose-built instruments, including validated multidimensional scales and reliable brief unidimensional measures, exist in the psychometric literature and continue to be refined.

The practical implications, however, are hard to escape. Survey organizations should test dimensionality before building an AI-attitude index, separate behavioral use from evaluative attitude unless theory explicitly joins them, and distinguish policy preference from institutional trust in governance batteries. Support for regulating AI companies and confidence that a particular government can regulate them are conceptually different judgments, and combining them blurs the line between wanting oversight and trusting the available overseer. As AI moves from specialist concern to routine political battleground, the temptation to report a single public mood will only grow. Katz’s warning is that even a small module of four topically related items can behave as an accidental scale, and that headline numbers built from such scales can be numerically precise while being conceptually incoherent. The accidental scale should be diagnosed and avoided, not treated as an implicit default for applied polling on AI.

Subject of Research: Testing the dimensionality of public attitudes toward artificial intelligence

Article Title: Testing the dimensionality of public attitudes toward artificial intelligence

Article References: Katz, J. E. (2026). Testing the dimensionality of public attitudes toward artificial intelligence. Discover Artificial Intelligence, 6(1), Article 1518. https://doi.org/10.1007/s44163-026-02338-1

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02338-1

Keywords: artificial intelligence, public opinion, survey methodology, psychometrics, AI governance, risk perception, digital divide, attitude measurement, Testing, dimensionality, public, attitudes