OpenAI Democratic Inputs to AI Program

OpenAI has launched a grant program to fund ten $100,000 experiments designed to establish democratic processes for determining the rules AI systems should follow. This initiative seeks to ensure that AI behavior is shaped by diverse public perspectives rather than being dictated by a single company, individual, or country.

Defining the Democratic Process for AI Governance

OpenAI defines a "democratic process" as one where a broadly representative group of people exchange opinions, engage in deliberative discussions, and reach an outcome through a transparent decision-making process. The goal is to move beyond simple legal frameworks to create intricate, adaptive guidelines for AI conduct.

Key components of this process include:

Broad Representation: Selecting participants to ensure diverse perspectives, including minority groups or domain experts depending on the specific policy question.
Deliberation: Using processes that uncover opinions and help participants understand and update their viewpoints based on values rather than misunderstandings.
Transparent Decision-Making: Utilizing various algorithms such as majority voting, liquid democracy, or sortition (random population sampling).

OpenAI acknowledges that these processes must address failure modes such as manipulation by special interest groups, participation washing, and the underrepresentation of minority or majority groups.

Grant Program Structure and Requirements

The program provides $100,000 grants to individuals, teams, or organizations to develop proof-of-concept prototypes for democratic decision-making.

Participation and Deliverables

Participant Scale: Each prototype must engage at least 500 participants.
Open Source Requirement: All code and intellectual property developed during the project must be made publicly available under an open-source license.
Reporting: Recipients are required to publish a public report of their findings and working prototype.

Review Criteria

Applications are evaluated based on several technical and social factors:

Robustness: Ability to prevent trolling and fake accounts.
Inclusiveness: Strategies for including diverse backgrounds and varying levels of AI familiarity.
Empowerment of Minorities: Ensuring unpopular or minority opinions can influence matters of significant concern.
Scalability: A preference for virtual processes over in-person engagement.
Legibility and Actionability: The ease of trusting the process and the degree to which the results can be applied to model behavior.

Policy Questions Under Consideration

The grant focuses on policy questions regarding model behavior to enable A/B testing of modified behaviors. Examples of nuanced questions the program encourages include:

Personalization: The boundaries of how far AI assistants should align with a user's personal tastes and preferences.
Public Figures: How AI should respond to questions about public figure viewpoints (e.g., neutrality vs. providing sources).
Professional Advice: The conditions under which AI should provide medical, financial, or legal advice.
Representation: How generative models should balance diversity versus homogeneity when creating images for underspecified prompts like "a CEO."
Cultural Sensitivity: Principles for handling topics involving human rights and local cultural or legal differences, such as LGBTQ and women's rights.
Content Restrictions: Criteria for determining which categories of content AI model creators should limit or deny.

Strategic Implications for AGI Oversight

This program is positioned as a first step toward establishing public oversight for the governance of Artificial General Intelligence (AGI) and superintelligence. OpenAI states that these initial experiments are not intended to be binding for decisions at this stage but are meant to build novel democratic tools that can inform future decisions. The lab views these democratic-in-spirit processes as complements to, rather than substitutes for, government regulation of AI.