Deliberative aggregation

A body of affected people is constituted — an open crowd on a voting platform, a stratified sample, a citizens' assembly, or a panel drawn by lot — and put through a procedure: proposing and voting on statements, or talking them through until a common position forms. Whatever that body settles on becomes the governing input to the model: a constitution to fine-tune on, a set of policy targets, or the pool of raters whose comparisons become the training data. The claim is legitimacy by procedure — the rule binds because of who decided it and how it was decided, not because of what it says.

A model takes over jobs a human facilitator, chair or annotator would do around a group's disagreement. It drafts a candidate group statement from everyone's opinions and critiques and redrafts it as participants rank the versions, picks slates of statements that represent the whole spread of opinion under formal fairness guarantees, chairs a live session by managing the speaking queue and intervening on incivility, voices stakeholders who are not in the room, and rates each contribution for justification, novelty and openness in place of the human coders who used to. Some of this runs live inside a session and some works offline from a collected set of opinions; what the family shares is that the model occupies a role in the deliberation rather than consuming its output afterward.

The standard pipeline pools everyone's comparisons into one reward model, which quietly averages disagreeing people into a single stand-in judge. This family makes an explicit collective decision rule the thing the system optimizes instead: maximize the worst-off group's reward, find the policy that beats every rival head to head, draw from a maximal lottery, aggregate whole policies rather than reward numbers, match the population's actual distribution of views, or — in the limiting case, where the object designed is an economic redistribution rule rather than a language policy — optimize directly to win an election among the people the rule governs. The same machinery is turned back on itself to say what any such rule can deliver: the axioms it satisfies, how far it can fall short of the ideal, whether one labeler who misreports can wreck it, and impossibility results showing that no protocol satisfies everyone.

Rather than building the aggregation into training, these systems gather many evaluative signals after the fact and combine them with a voting rule instead of an average. Several models from different families, or one compact reward model fitted per annotator, each score the candidate outputs and vote; the same move treats existing benchmark and tournament results as ballots and searches for the ranking of agents that contradicts the fewest of them. Because the voters live at evaluation time, who votes is a knob that can be turned without retraining anything — which is what separates this from rewriting the objective.

Instead of aggregating whatever judgments people happen to volunteer, the system authors the question each person faces, because the question is where the normative work happens. A model runs an open-ended interview, invents the edge case that would be most informative, walks someone through critiquing an agent's behavior until their own preferences take shape, or asks what value they actually brought to a hard case and then which value is wiser here. The designs differ in how the question is chosen, and the difference matters: some pick the next comparison adaptively for information gain, while one applies a single fixed framing — choosing principles without knowing which position you would occupy — that pushes answers toward protecting the worst-off. What separates all of them from voting platforms is that the agenda is written by the system, not by the participants.

Here the collective answer is not a tally at all. Concrete judgments about particular cases and candidate general principles are adjusted against each other until the two cohere — modeled formally under measurable account, systematicity and faithfulness criteria, read as a description of what constitutional AI already does, and shown to be the same search as coherence-driven self-improvement. Attached to the machinery are the arguments for why the target should be justified reasons rather than counts: that preferences do not represent values, and that alignment should be settled by the claims people can fairly make on each other.