Predicting Perceived Barriers to Wider AI Use in Six Survey Samples

We assessed whether demographic characteristics, reported AI use and attitudes predict the greatest perceived barrier to wider artificial intelligence (AI) use. The analysis included 3528 adults in six provider-labelled survey samples whose geographic membership could not be independently verified. All ten response alternatives were retained. Multinomial logistic regression, random forests, and histogram gradient boosting were evaluated using nested leave-one-site-out and random five-fold validation; CatBoost was added as an exploratory comparison. Random forests improved log loss over the corresponding training priors by 0.083 under sample holdout and 0.102 under random folds. Most of the raw loss difference between designs reflected the difference between their priors. Matching training sizes and the specified sensitivity analyses retained modest probability gains and weak classification. Under sample holdout, forest accuracy was 22.0% and macro F1 was 0.093, with zero recall for four categories. CatBoost showed similarly limited performance. These findings describe prediction with the evaluated models and available questionnaire variables. They do not establish geographic mechanisms or operational usefulness, and support directly asking respondents about their selected barrier.