Generative AI Use and Academic Integrity Behaviour among Law Students in Bangladesh

Generative artificial intelligence presents legal education with two distinct integrity tasks: checking the reliability of generated legal material and making assistance visible when disclosure is required. This study examines their relationship through an exploratory analysis of the Group 4 Law-SDG database. Of 385 records, 245 identify the respondent as a law student and constitute the primary sample. Four conceptually matched items distinguish verification from disclosure and stated responsibility from reported practice. Analyses combine exact paired tests, Holm adjustment, respondent-level bootstrap intervals and population-averaged logistic models with respondent-clustered covariance estimation. Verification practice is endorsed by 204 students (83.3%), compared with 160 (65.3%) for disclosure practice: a paired difference of 18.0 percentage points (95% bootstrap interval 10.2–25.7). The corresponding normative difference is 15.5 points (8.2–22.9). Within each domain, duty-practice differences are small and statistically uncertain. Adjusted endorsement odds are higher for verification than disclosure (OR 2.67, 95% CI 1.71–4.18), while policy awareness, training and use frequency have uncertain conditional associations. Only 43 students (17.6%) endorse all five integrity-practice items. Results persist with stricter endorsement and in wider-sample checks. Low block internal consistency supports item-level interpretation. Numerical provenance remains unverified because the codebook labels identifiers and contextual comments synthetic; the latter are excluded. The study identifies a verification-disclosure asymmetry in recorded responses, without estimating misconduct prevalence or testing the full theory of planned behaviour. It proposes separate educational workflows for checking legal authority and documenting permissible AI assistance.