Democratic Input Gathering for the AI Grant Program: Lessons and Action Plans

OpenAI ran the Democratic Inputs to AI grant program, which reflects public opinion to align AI models with human values. From about 1,000 applications, 10 teams were selected, each receiving $100,000 to design tools and ideas for AI governance.

The selected teams were composed of experts from diverse fields such as law, journalism, machine learning, and social science, and presented innovative approaches including:

• A new video deliberation interface
• A crowdsourced auditing platform for AI models
• Mathematical formalization for representation guarantees
• A belief-mapping method for model fine-tuning

During the projects, AI itself was usefully leveraged within the process, in forms such as custom chat interfaces, speech-to-text conversion, and data synthesis.

One key finding was that public opinion changes frequently. Based on the lessons learned from this program, OpenAI plans to build an end-to-end process for gathering input from external stakeholders and using it to shape and train actual model behavior.