Enterprise AI Analysis

Data Ethics

Data collection, curation, and annotation are fundamental yet problematic in Bangladesh. Concerns arise from the misuse of personal/public data, ignoring copyright, and inadequate compensation for training data providers, especially for foreign companies utilizing local data. Practices like providing "5 taka or food" to indigenous people for data collection were noted. While some academic researchers and engineers follow good practices voluntarily (e.g., balancing gender/ethnicity ratios for speech datasets, obtaining consent), this is not widespread. The rise of data annotation firms raises concerns about minimum wage and mental health impacts on annotators, highlighting the need for regulation to set industry standards and avoid exploitation.

AI Regulation and Audit Concerns

There is an urgent need for AI-specific policies and regulations in Bangladesh, as current legal frameworks are inadequate. The national AI strategy draft lacks comprehensive guidelines for critical areas like national security or cybercrime. Participants expressed unhappiness with previous drafts, emphasizing the need for participatory design involving academics, industry, and citizens. Concerns were raised about regulations being easily abused and potentially harming emerging AI innovation, particularly for local companies. The consensus suggests that AI regulation should be simple, culturally considerate, and focused on fostering the local AI ecosystem, especially for products coming from the global North. There is a strong call for independent oversight and auditing to ensure accountability and prevent misuse of regulatory power, with suggestions for government, authorized independent organizations, or international companies to maintain certification programs.

Stakeholder Perspectives on FATE

Fairness: Many stakeholders, especially non-technical ones, lacked adequate knowledge of fairness despite its importance in global South contexts. Discussions often centered on ensuring diversity and protecting marginalized users during data collection.

Accountability: With no current AI regulation, liability is largely user-centric. Participants demanded a separate constitutional policy section for AI risks and stressed that accountability structures must be culturally and geographically contextualized, not universally applied.

Transparency: While some researchers focused on transparent policy design and model-level transparency initiatives in other countries, general interest in AI transparency for basic software/hardware was low among most participants.

Explainability: A significant concern for Bangladesh, with XAI researchers noting a complete lack of awareness even among AI researchers. End-users do not understand the term, requiring contextualized explanations. Government sector participants acknowledged no specific expectations for model explainability.