AI Assisted Case Analysis and Self Reported Evaluative Confidence among Law Students in Bangladesh

Abstract Generative artificial intelligence can support legal case analysis, but perceived assistance, reported verification and confidence in evaluating reasoning represent different educational outcomes. This study examines their alignment in a Bangladesh questionnaire database containing 385 consenting, AI-aware adults. The primary analysis concerns the 237 respondents identifying as law students; the full database provides descriptive context and sensitivity analysis. Low internal consistency supported item-level analysis rather than a composite critical-reasoning scale. Endorsements, paired contrasts, joint response patterns and exploratory regression were used to distinguish reported practices from self-assessed capability. Among students, 66.7% reported checking AI-generated legal information against authoritative sources, 41.4% reported comparing alternative interpretations, 48.9% expressed confidence in detecting weak or unsupported reasoning and 52.3% reported retaining independent judgement. Verification endorsement exceeded detection confidence by 17.7 percentage points (bootstrap 95% interval 9.3–26.2; Holm-adjusted p < 0.001). Only 18.6% endorsed verification, detection confidence and independent judgement together. None of four focal predictors-comparison, legal-source verification, awareness of plausible inaccuracies and prompting capability-was retained after multiplicity correction in the adjusted student model. Ordinal, full-sample and secondary judgement-outcome analyses likewise retained no focal coefficient. Self-regulated learning supplies a process-oriented interpretation, while technology acceptance helps delimit perceptions of assistance. The findings identify uneven alignment among self-reports rather than an observed deficit in legal reasoning or a causal effect of AI. Educational implications concern evaluating source checks, argument comparison and independent revision through performance-based tasks, with explicit attention to jurisdiction and authority. The study supports more discriminating assessment of AI-assisted legal learning rather than inferring competence from use or confidence alone.