Perceived Authorship Responsibility and Reported Practices in AI Assisted Legal Writing in Bangladesh

Abstract Generative artificial intelligence can assist legal writing while leaving unresolved how writers understand responsibility for accuracy, verification, disclosure and explanation. These responsibilities should not be assumed to form a single construct or to correspond directly with reported practice. This study analyses 385 adult, AI-aware survey records from Bangladesh, including 245 law students, using separate authorship statements and content-matched reports of verification and disclosure. Analyses combine response distributions, paired endorsement contrasts, joint patterns, ordinal correlations and an exploratory proportional-odds model of perceived disclosure responsibility. Responsibility for accuracy was endorsed by 89.6%, the duty to verify authorities by 85.2%, required disclosure by 66.8%, explaining and defending submitted text by 67.8%, and avoiding reliance on ununderstood writing by 60.3%. Verification duty and reported verification had similar endorsement proportions, yet 95 records differed in binary endorsement. Required disclosure and reported disclosure differed in 170 records. These response combinations do not establish misconduct or an intention–behaviour gap. Seven selected ordinal associations were small and uncertain after Holm adjustment. In the adjusted disclosure-responsibility model, policy awareness, reported disclosure and accuracy responsibility had odds ratios of 1.20, 1.16 and 1.22 respectively, with intervals spanning one. Student-restricted estimates were similarly uncertain. The findings distinguish widely endorsed principles from their cross-item alignment rather than validate a causal policy-to-integrity mechanism. They support evaluating task-specific disclosure, source verification and explanation through observable assessment activities. Undocumented recruitment and fieldwork dates, mixed respondent status, single-item measures and source-documentation limitations constrain inference. The article provides a measurement-based foundation for legal-writing education without equating favourable self-reports with demonstrated compliance.