Democratic governance through DAO-based deliberation and voting for inclusive decision making in AI models

Area of Science:

• Artificial Intelligence Ethics
• Human-Computer Interaction
• Social Computing

Background:

• AI development faces transparency issues, leading to discrimination and regulatory breaches, disproportionately affecting underserved populations.
• Traditional social science methods are insufficient for capturing user needs in the digital age due to limitations in deliberation and consensus-building.

Purpose of the Study:

• To develop and assess a democratic decision framework using Decentralized Autonomous Organizations (DAOs) for underserved groups to deliberate on AI issues.
• To reduce stereotypical biases, specifically gender bias, in text-to-image AI systems through stakeholder input.

Main Methods:

• A case study was conducted using a 2x2 experimental design with a randomized online experiment (n=177).
• Participants from the Global South and individuals with disabilities evaluated governance configurations, including ranked vs. quadratic aggregation schemes and equal vs. 20/80 differential decision power distribution.
• The study examined how governance mechanisms influenced perceptions of decision-making processes and AI model specifications.

Main Results:

• Participants converged on key aspects: user control in image generation, multiple output options, and the social appropriateness and accuracy of AI-generated images.
• The combination of quadratic preference aggregation (empowering minorities) and equal decision power distribution was perceived as the fairest and most democratic approach.
• Despite diverse backgrounds, participants reached consensus on crucial AI development considerations.

Conclusions:

• Appropriate governance mechanisms are vital for democratic decision-making in AI alignment.
• DAO-based frameworks can effectively facilitate consensus-building among diverse and underserved populations for AI development.
• The findings highlight a pathway for more inclusive and equitable AI systems by integrating user-centric governance.