Gender Bias Concerns and Support for Algorithmic Accountability in Bangladesh

Governance of gender bias in artificial intelligence (AI) involves distinct commitments: testing for unequal outcomes, assigning responsibility, explaining safeguards and providing remedies. An overall measure of support can obscure differences among these commitments. This cross-sectional study analysed 385 adult responses to a common questionnaire on AI, gender, equality and law in Bangladesh. We examined individual response distributions, compared endorsements within respondents and estimated exploratory adjusted associations with two separate outcomes: support for effective remedy and anticipated confidence from transparency about bias testing. Effective remedy and organisational responsibility were each endorsed by 302 respondents (78.4%); testing for gender-related discriminatory outcomes was endorsed by 215 (55.8%). The paired remedy-testing difference was 22.6 percentage points (95% CI 16.1 to 29.1; Holm-adjusted p < 0.001). Both were endorsed by 167 respondents (43.4%). Bias concern was associated with anticipated confidence from transparency (b = 0.204, 95% CI 0.096 to 0.312; adjusted p < 0.001), but the corresponding association with remedy endorsement on the original rating scale was imprecise and near zero. These findings distinguish endorsement of redress from endorsement of a preventive organisational practice. They also distinguish anticipated reassurance from observed changes in trust. The sample is concentrated in legal study and practice and does not support population prevalence estimates. The findings motivate separate evaluation of testing, communication and access to remedy.