Auditable Rules Outperform Opaque AI in Deliberative Polling

This paper introduces a novel approach to argument selection in deliberative polling, advocating for transparent, rule-based mechanisms over opaque, AI-driven rankers. The core idea is to ensure that voters can understand, recompute, and even contest the exposure of arguments that shape their opinions, treating legibility as a fundamental requirement rather than a trade-off against accuracy. The authors formalize a poll using bipolar justification sets and propose a specific rule based on reversed endorsement flow, parameterized by user-defined relation-weight functions. Through extensive simulations involving approximately 17,000 runs, the proposed rule-based system demonstrates performance remarkably close to the theoretical upper bound of any selection procedure, including unconstrained opaque rankers. While it performs comparably to random selection on coverage alone, it significantly outperforms on metrics like order of arrival and captured endorsement mass, especially under adversarial conditions. The research highlights that the choice of ranking mechanism is a policy decision, not merely a technical one, emphasizing that legible rules empower users with control over how information influences their vote. This framework is designed for an open-source, peer-to-peer platform, promoting transparency and civic engagement.