Putting the model in the facilitator's seat

Definition · Jürgen Habermas, Between Facts and Norms (1992)

This concept locates legitimacy in the exchange of reasons among equals: participants justify their views to each other, free of coercion, and remain able to change their minds. Counting votes or averaging preferences does not produce it, because the reasoning is what does the work.

Analysis

One mechanism in the family genuinely exchanges reasons: in the Habermas machine participants file critiques of a draft and the model rewrites in light of them, and participants demonstrably moved position, which no tally can produce; the Stanford moderator adds the speech conditions Habermas cares about, allocating speaking time and intervening on incivility. But every member settles the matter by counting rather than by the better argument — the winning statement is fixed by a preference ranking, the veto-core and token-level-MDP papers work entirely from per-participant preference models so no participant reasons with anyone, and the two quality graders score contributions one at a time, so a discussion in which one party dominates and nobody is moved can score well throughout. Reasons enter as inputs to an aggregation rather than as what does the work.

Definition · Jean-Jacques Rousseau, The Social Contract (1762)

This distinction separates what is good for a public in common from the sum of what its members privately want. Aggregating private wants, however fairly, does not produce the former, since the two can and often do diverge.

Analysis

What is carried is the form of the output: most of these systems emit a single statement written in the group's voice rather than a numeric average, and Tessler et al. show the winning statements incorporate dissenting views instead of tracking the majority, which is the divergence Rousseau points at. What is not carried is the distinction itself — every rule that picks the statement is a function of what members privately want (social welfare functions in Bakker, the proportional veto core, per-participant token rewards, proportional slates), and the Overton benchmark literally scores how much of the spread of private views one answer covers. In Bakker the choice between summing the group and protecting its worst-off is a dial the researchers turn, so the one place the question arises it is settled off-stage rather than by the public whose common good is at issue.

Definition · John Rawls, A Theory of Justice (1971)

This device requires that rules be chosen without knowledge of which position the chooser will occupy under them, e.g., rich or poor, majority or minority. Not knowing generally pushes the chooser to protect the worst-off position, since it may turn out to be their own.

Analysis

Nothing anywhere in the family screens a chooser from their own position: participants opine and critique as themselves, the citizens'-assembly replication matches demographics precisely so that known positions are present, annotators rate from their own standpoint, and the Empty Chair supplies absent perspectives by naming and voicing them rather than by veiling anyone. The maximin outcome the veil is meant to generate does surface — Bakker selects statements under egalitarian as well as utilitarian welfare functions, and the proportional veto core shields minorities from being overruled — but it arrives as a researcher's choice of objective, which is the conclusion without the device. No source engages the veil; using Rawlsian vocabulary for a welfare function imports the outcome principle, not the construction procedure.

Definition · T. M. Scanlon, What We Owe to Each Other (1998)

This test holds a principle justified only if no individual could reasonably reject it, and it is applied person by person rather than in aggregate. One sufficiently strong objection therefore outweighs many mild preferences, which is the case averaging handles wrongly.

Analysis

Rejection power exists here in quantitative form: the proportional veto core lets a coalition block any statement in proportion to its size, generative social choice guarantees a cohesive group of sufficient size a statement of its own, and in the mediation loop a critique must be answered in the next draft or the statement loses ground. Both halves of Scanlon's test are missing, though — it is applied person by person and is indifferent to how many share the objection, and it turns on the grounds offered rather than on strength of feeling. In these systems an objection is registered as a preference score or a rating, so a reasoned rejection and a mild dislike are the same datum and many mild approvals outweigh one strong objection; Bakker's own exclusion experiment makes the failure concrete, with a statement assembled from a subset still standing over the member left out.

Deliberative theory asks that the outcome be one everybody could accept; it does not say the outcome should contain the objections. The embedding analysis here finds that the statements which succeed are the ones that incorporate dissenting voices while respecting the majority — a property of the artifact rather than of the procedure, and one the tradition has no name for.

Habermas supplies a standard for deliberation and no way of telling whether a given discussion met it, because his participants were assumed present to one another. These instruments make the standard computable for every post: AQuA runs adapter models for 20 deliberative indices and produces a score from pre-trained adapters alone, and the Stanford model rates on four criteria at a scale the paper says human annotation cannot reach, being time-consuming and costly. Whether the number measures what the standard asks for is a separate question, but the tradition offers no number at all.