OpenAI Democratic Inputs to AI Grant Program Update

OpenAI has concluded the initial phase of its Democratic Inputs to AI grant program, funding 10 global teams to develop tools and methods for the collective governance of AI. This initiative aims to align AI model behavior with human values by creating scalable, democratic processes for collecting and implementing public input.

Innovations in Democratic AI Governance

OpenAI awarded $100,000 to 10 teams selected from nearly 1,000 applicants across 113 countries. These teams utilized diverse expertise in law, journalism, social science, and machine learning to create prototypes for participatory engagement. AI was frequently used within these processes to facilitate transcription, data synthesis, and customized chat interfaces.

Summary of Grant Projects

  • Case Law for AI Policy: Created a case repository of AI interaction scenarios to enable case-law-inspired judgments involving experts and stakeholders.
  • Collective Dialogues for Democratic Policy Development: Used collective dialogues to scale democratic deliberation and identify areas of consensus for policy development.
  • Deliberation at Scale: Developed AI-facilitated video calls to enable democratic deliberation in small group conversations.
  • Democratic Fine-Tuning: Elicited values via chat dialogues to create a "moral graph" for model fine-tuning.
  • Energize AI: Built a platform for alignment using live, large-scale participation and a "community notes" style algorithm.
  • Generative Social Choice: Applied social choice theory to distill free-text opinions into a concise, fairly representative slate of positions.
  • Inclusive.AI: Used decentralized governance mechanisms, such as DAOs, to empower underserved populations in AI decision-making.
  • Making AI Transparent and Accountable (Rappler): Combined offline and online processes to facilitate discussion on complex, polarizing topics.
  • Ubuntu-AI: Developed a platform to ensure African creative work is inclusively represented and that contributors are compensated.
  • vTaiwan and Chatham House: Adapted the vTaiwan methodology to create a recursive, connected participatory process for AI governance.

Key Technical and Social Learnings

The grant program revealed several critical challenges and insights regarding the collection of public input for AI alignment.

Volatility of Public Opinion

Public views on AI behavior can change frequently, sometimes on a day-to-day basis. This suggests that input-collection processes must be designed to capture both fundamental, stable values and be sensitive enough to detect meaningful shifts in opinion over time.

The Digital Divide and Participation Bias

Recruiting participants across digital and cultural divides remains a significant hurdle. Teams observed that participants recruited online tended to be more optimistic about AI, which correlated with higher support for AI model behavior in general. Additionally, technical limitations in speech recognition tools, such as Whisper, were noted in languages like Tagalog, Binisaya, and Hiligaynon, complicating transcription for the Rappler team.

Managing Polarization and Diversity

Finding compromise within polarized groups is difficult, particularly when small minorities hold very strong opinions. However, the Collective Dialogues team found that policy guidelines generated through their process—including on divisive issues like vaccine information—achieved over 72% support across Democrats, Independents, and Republicans.

There is a persistent tension between reaching a broad consensus and representing the diversity of viewpoints. The Generative Social Choice team used mathematical theory to highlight key positions and showcase a range of opinions while seeking common ground.

Public Trust and AI's Role in Governance

Participants expressed caution regarding the amount of power democratic institutions and AI developers should grant to AI systems. Trust is higher when AI is used for efficiency (e.g., generating policy clauses) but is paired with non-AI decision steps, such as expert curation or final human votes.

Implementation and Future Direction

OpenAI is establishing a "Collective Alignment" team composed of researchers and engineers. This team is tasked with two primary objectives:

  1. System Implementation: Designing and implementing a system to collect and encode public input on model behavior directly into AI systems.
  2. Pilot Integration: Continuing collaboration with external advisors and grant teams to run pilots that incorporate the prototypes developed during the grant program into the steering of OpenAI models.

Sources