Generative AI Literacy and Research Self-Efficacy in Relation to Attitudes Toward Scientific Research Among Health Sciences Students: A Cross-Sectional Study

Abstract

Generative artificial intelligence (GenAI) has rapidly entered undergraduate research training, yet its relationship with research self-efficacy and attitudes toward scientific research remains underexplored among Health Sciences students. This cross-sectional study examined how self-reported GenAI literacy and research self-efficacy are associated with attitudes toward scientific research, separately and when modelled together. Of 101 submissions from undergraduate Health Sciences students enrolled in research courses at a private Peruvian university, 91 were analysed after 10 invariant response patterns were excluded. Students completed a 36-item self-report questionnaire whose items had been reviewed for content relevance and clarity by three experts. Analyses combined Spearman correlations, multiple regression with HC3 standard errors, a bootstrapped indirect-effect model and exploratory fuzzy-set qualitative comparative analysis (fsQCA). Reliability computed from the three dimension scores of each construct was α = 0.861 for GenAI literacy, 0.927 for research self-efficacy and 0.790 for attitudes. GenAI literacy correlated with attitudes (ρ = 0.475, Holm-adjusted p < 0.001), but in the female–male regression model (n = 90), research self-efficacy was the only one of the two capability scores that retained a statistically significant adjusted association (b = 0.353, 95% CI 0.087 to 0.619, p = 0.010; GenAI literacy b = 0.088, 95% CI −0.214 to 0.389, p = 0.564; R2 = 0.405). The indirect term through research self-efficacy (0.313, 95% CI 0.092 to 0.536) is a statistical decomposition of overlapping, contemporaneous scores rather than evidence of a mechanism. Discriminant validity between GenAI literacy and research self-efficacy was not established (HTMT = 0.912), and parallel analysis favoured two factors rather than three. Under the primary calibration (theoretical anchors 2-3-4 with crossover memberships assigned to 0.501), functional knowledge and the critical–ethical evaluation of GenAI appeared in both terms of the intermediate fsQCA solution (consistency = 0.913, coverage = 0.741), but the minimised solutions changed with alternative anchors and with the handling of cases at the crossover. The findings describe associations among self-reported measures; whether developing both capabilities jointly improves research attitudes remains a question for future evaluation.