Generative Polis — Consensus-Building with AI, Beyond Polis

August 2, 2026 | Mizuki Oka, Yusuke Hayashi, Momoha Hirose

---

Session Report

From Voting Patterns to New Statements

At Digital Democracy Summit 2026, held at Keio University’s Mita Campus in Tokyo, our research team introduced Generative Polis, an ongoing project, in a session titled Generative Polis — Consensus-Building with AI, Beyond Polis.

The session focused on a concrete research question: How can statements that may be shareable across a group be constructed from the local predictions of individuals with different beliefs?

The project grows out of the broader idea of Symbiotic Alignment and is conceptually grounded in Collective Predictive Coding (CPC). CPC views shared symbols as emerging through distributed inference among agents with different local observations. In Generative Polis, this perspective motivates the use of participant-level predictions as local evidence when evaluating candidate statements as potentially shared representations. Rather than revisiting the full framework, the session concentrated on one practical design question: whether AI can help propose new expressions across different positions while leaving their evaluation to the participants themselves.

What Polis Makes Visible

Polis provides an important empirical starting point. Participants respond to statements with Agree, Disagree, or Pass, producing a voting matrix that reveals patterns across participants and statements. From these patterns, Polis can identify opinion groups as well as points of agreement that cross group boundaries.

In this sense, Polis is more than a polling tool. It offers a window into the collective structure of meaning within a discussion. Each vote records a local response from one participant to one statement; taken together, those responses reveal how different positions relate to one another.

Yet any deliberative process can evaluate only the statements that have entered the conversation. A formulation capable of bridging different groups may not have been articulated yet. The central question behind Generative Polis is whether AI can help propose such statements without replacing the participants’ own judgment.

Generative Polis: Grounding Generation in a Group

Generative Polis combines three components: voting data from Polis, participant-level response models, and candidate statements proposed by an LLM.

First, the system uses the existing voting history to estimate how each participant responds to different kinds of statements. The LLM then generates new candidate statements that have not yet appeared in the discussion. For each candidate, the model predicts how individual participants might respond. These local predictions are combined to evaluate and reweight the candidate set.

The distinction from ordinary LLM generation is not primarily the fluency of the resulting text. It is what the text is evaluated against. An LLM alone selects language according to linguistic plausibility and its learned preferences. Generative Polis instead uses the group’s distributed response structure as a constraint on generation.

Prediction is therefore not the final objective. It is a way of grounding generated statements in the people who make up the deliberating group. AI does not decide the consensus; it proposes new shared expressions that the group itself can evaluate.

From Common Ground to Uncommon Ground

The aim is also more demanding than finding a statement with high average support. A broadly agreeable statement may be too abstract to change the structure of a disagreement. Generative Polis is ultimately concerned with what Audrey Tang has called uncommon ground: a previously unrecognized formulation that crosses an existing divide by reframing the problem.

The same-sex marriage debate in Taiwan illustrates the difference. People on opposing sides both emphasized the importance of “family,” but they were referring to different relationships. One side sought legal protection for the partnership between two people; the other sought to preserve existing kinship relations between families. The breakthrough did not come from averaging the two positions. It came from a new legal formulation that protected the couple’s relationship without automatically creating kinship ties between their families.

Uncommon ground is therefore not merely a shared abstract value. It appears in a concrete reformulation that allows different values to coexist. It is not a compromise located halfway between two positions, but a restructuring of the disagreement itself.

An Ongoing Prototype

The session also introduced an early prototype using existing Polis voting data to explore whether newly generated statements could be evaluated in relation to participants’ prior response patterns. This work remains in progress and was presented as an initial exploration rather than a completed or validated result.

Toward a Vote–Generate–Vote Process

The next step is to close the loop: participants vote, the system proposes new statements, and participants then vote again. Testing this iterative vote → generate → vote process in practice will help us understand what role generated statements might play in an evolving discussion.

Digital democracy may not need AI that makes everyone agree. It may need AI that helps people remain different while developing the language required to think and act together. Generative Polis is an early attempt to build that possibility: not an AI that determines collective decisions, but an AI that helps a collective generate new expressions it can examine for itself.