Human-centred mechanism design with Democratic AI

Area of Science:

• Artificial Intelligence
• Social Science
• Economics

Background:

• Aligning artificial intelligence (AI) with human values remains a significant challenge.
• Developing AI systems that can understand and incorporate human preferences is crucial for societal adoption.

Purpose of the Study:

• To develop a human-in-the-loop research pipeline, termed Democratic AI, for creating AI systems aligned with human values.
• To use reinforcement learning to design a social mechanism that achieves majority human preference.

Main Methods:

• A human-in-the-loop research pipeline called Democratic AI was developed.
• Reinforcement learning was employed to design a social redistribution mechanism.
• A large-scale online investment game was conducted with human participants to evaluate the AI-designed mechanism against a human-designed one.

Main Results:

• The AI-designed mechanism successfully redressed initial wealth imbalances among players.
• The AI-designed mechanism incorporated a system for sanctioning free riders.
• The AI-designed mechanism won the majority vote of human participants, indicating preference.

Conclusions:

• Democratic AI provides a proof of concept for value-aligned policy innovation.
• Optimizing AI for human preferences is a viable strategy for developing trustworthy AI systems.
• This approach demonstrates the potential for AI to design effective and preferred social mechanisms.