Abstract

Democratic and computational decision systems are increasingly using artificial intelligence to collect opinions, summarize disagreements, and combine different viewpoints into decisions. However, many common aggregation methods, such as majority voting, averaging preferences, or training one reward model from diverse feedback, often simplify disagreement too much. They may ignore the deeper structure behind why people disagree, including differences in preferences, uncertainty, and reasoning.
This thesis studies common-ground mechanisms that use structure within disagreement instead of relying only on averages or majority counts. Here, "structure'' refers to measured differences in group response, stated concerns, uncertainty, reasoning relations, and answer-space composition. Whether those measurements are valid is treated as an empirical question rather than assumed from the mechanism.
The thesis develops this idea in two connected settings. The first setting focuses on consensus among humans. I present a framework for policy search in which large language models use survey-conditioned profiles to screen how different groups may respond to policy propositions. The system searches over constrained rewrites intended to increase agreement and reduce between-group disagreement while preserving the original policy intent. In the refined run, the selected rewrite raises simulated agreement but the baseline calibration is weak, so this is screening evidence rather than a forecast of public response. A separately fielded Wave 2 proposition has a reported aggregate agreement share of 58.1\%; without a concurrent original-proposition control, respondent-level data, and confirmed item-level denominators, that distribution does not estimate the rewrite's causal effect.
The second setting focuses on consensus among AI systems. I develop an exact condition for when confidence-adjusted delegation overturns majority and examine it through two independent multi-sample LLM studies. In both, groups of sampled answers act as delegates: they retain support according to answer uncertainty and route the remainder through reasoning similarity. The first study finds mixed, model-dependent changes relative to majority. The second finds that voter-level propagation mostly matches fair majority on one-of-many multiple choice, but per-option propagation improves micro-F1 by 1.9 to 6.3 points on a select-all-that-apply benchmark. In the reported studies, entropy-derived retention is a more consistent correctness signal than the current embedding-based routing measure.
Together, these studies support a narrower principle: AI-mediated consensus mechanisms should represent and test meaningful structure in disagreement rather than rely only on persuasive language, averages, or vote counts. For humans, the relevant structure includes subgroup responses, stated concerns, and legitimacy constraints. For AI systems, it includes uncertainty, reasoning relations, delegation networks, and answer-space composition. In both cases, a coherent mechanism is not sufficient; its measurements must also be shown to support the interpretation assigned to them.