Abstract: This paper examines the governance of multimodal large language models (MM-LLMs) through individual and collective deliberation, focusing on analyses of politically sensitive videos. We conducted a two-step study: first, interviews with 10 journalists established a baseline understanding of expert video interpretation; second, 114 individuals from the general public engaged in deliberation using Inclusive.AI, a platform that facilitates democratic decision-making through decentralized autonomous organization (DAO) mechanisms. Our findings show that while experts emphasized emotion and narrative, the general public prioritized factual clarity, objectivity of the situation, and emotional neutrality. Additionally, we explored the impact of different governance mechanisms: quadratic vs. weighted ranking voting and equal vs. 20-80 power distributions on users decision-making on how AI should behave. Specifically, quadratic voting enhanced perceptions of liberal democracy and political equality, and participants who were more optimistic about AI perceived the voting process to have a higher level of participatory democracy. Our results suggest the potential of applying DAO mechanisms to help democratize AI governance.
Introduction
As multimodal large language models (MM-LLMs) are increasingly deployed to interpret complex social and political events, the question of "alignment"—deciding whose values these models should reflect—becomes critical. Current governance often relies on top-down corporate decisions or narrow expert panels, creating a "participation gap" that excludes the broader public. This research introduces Inclusive.AI, a platform designed to bridge this gap by utilizing Decentralized Autonomous Organization (DAO) mechanisms to facilitate public deliberation and voting on AI behavior.
Methodology
The study followed a two-step approach to understand the differences in how experts and the public value AI interpretations of politically sensitive content.
- Expert Baseline: Qualitative interviews were conducted with 10 journalists and media scholars to identify professional standards for interpreting political videos.
- Public Deliberation: 114 US-based participants used the Inclusive.AI platform. After watching a clip from the 2020 US Presidential Debate, they engaged in human-AI deliberation (clarifying values with a GPT-4 agent) and group deliberation (discussing outputs with other users).
To test governance efficacy, the researchers employed a 2x2 experimental design:
- Voting Method: Quadratic Voting (QV), where the cost of a vote is the square of the number of votes cast, versus traditional Weighted Voting.
- Power Distribution: Equal voting power versus a "Pareto" distribution (20% of users holding 80% of tokens) to simulate real-world governance dynamics in decentralized systems.
System Architecture
Inclusive.AI integrates several modules to support democratic AI governance. The platform handles user chat history via HTTP and GPT-4 prompting for individual deliberation. For collective deliberation, it uses websockets to facilitate live group chats. The governance layer utilizes blockchain technology to issue voting tokens, retrieve addresses, and record votes via a snapshot server, ensuring a transparent and verifiable decision-making process.
Key Findings
The Alignment Gap
The study revealed a stark contrast between expert and public priorities. Journalists criticized LLMs for lacking "storytelling," "tone," and "contextual depth," seeking a nuanced analyst. In contrast, the general public prioritized "factual clarity," "objectivity," and "emotional neutrality," essentially demanding a "just the facts" approach.
Democratic Governance and Quadratic Voting
Quadratic Voting emerged as a superior mechanism for expressing preference intensity. Participants found QV to be more meaningful and fair than weighted voting, as it prevented a small number of high-token holders ("whales") from completely dominating the outcome. Statistical analysis showed that QV, particularly when combined with equal power distribution, significantly enhanced participants' trust in the decision-making process ($P = 0.013$). Furthermore, using the Variety of Democracy (V-Dem) scale, the researchers found that decentralized governance reinforced perceptions of liberal democracy and political equality.
Partisan Differences
Voting patterns also reflected partisan nuances. Republicans were more likely to favor "integrating a user feedback loop," whereas Democrats showed a stronger preference for "providing more specific facts."
Conclusion
This research demonstrates that DAO-based mechanisms like Quadratic Voting can effectively democratize AI governance by allowing diverse stakeholders to deliberate and vote on model behaviors. While the study highlights the potential for increased trust and perceived fairness, future work must address the scalability of such systems and the technical challenges of preventing voter fatigue in large-scale implementations.
Summary updated