Deliberative aggregation

Handing the decision to a convened public

A body of affected people is assembled. It can be an open crowd on a voting platform, a stratified sample, a citizens' assembly, or a panel drawn by lot. This body goes through a procedure, either proposing and voting on statements or discussing them until a common position forms. Whatever the body decides becomes the governing input to the model. This input can be a constitution to fine-tune on, a set of policy targets, or the pool of raters whose comparisons become the training data. The approach claims legitimacy through procedure. The rule is binding because of who decided it and how it was decided, not because of what it says.

The method (highlighting differences with Deliberative aggregation)

Scroll the diagram sideways to see all of it.

Use it in codefrom aiag.procedural import Polis, ConvenedPublic, Sortition
from aiag.procedural import Polis

statements = ["Keep the library open on Sundays.", "Pay for it from the parking budget.", "Close the Elm Street branch."]
votes = [[1, 1, -1],  # one row per resident: 1 agree, -1 disagree, 0 pass
         [1, 1, 0],
         [1, -1, 1],
         [1, -1, 1]]

# Opinion groups from the votes, then each statement's agreement rate in every group, multiplied
result = Polis()(votes)
print(result["groups"])                          # -> [0 0 1 1]
print(statements[result["consensus"].argmax()])  # -> Keep the library open on Sundays.

pip install aiagEvery method in aiag Library on GitHub

Secondary placement in the taxonomy

  • Constitutionalism — A deliberative procedure produces a written specification. The public deliberation produces the norms, but what is applied to the model afterward is the document it wrote. The approach is procedural in producing norms and codified in applying them.

The analysis below uses the primary placement in the taxonomy.

Concept Analysis: Theoretical Foundations

Each concept is read twice: whether the approach carries it over, and whether the approach's own sources claim it. A concept that the approach claims but does not carry over is a shortfall. A concept that it neither claims nor carries over is an omission, which marks a gap in the field rather than a failure of the work.

Deliberation, Not Tallying

PartialShortfall · partially preserved

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 legitimacy, because the reasoning itself is what produces it.

Analysis

The collective-dialogue papers (Konya et al. (2024), Konya et al. (2023), and Konya et al. (2025)) put participants in iterative rounds in which they read and react to statements written by other participants. Konya et al. (2025) sustained this for four months across the hardest possible divide. This is more than a poll. However, every procedure here decides by score rather than by reasons. Huang et al.'s (2024) Polis run uses agreement rates, and the dialogue papers use bridging-based ranking. Konya et al. (2025) uses agreement percentages for each side, and Shapiro and Talmon (2026) uses distances between ideal points. Sana et al. (2026) uses raw rating counts, and its randomly drawn panel rates outputs exactly as a convenience sample would. Nobody answers anybody's objection, no reason accompanies a vote, and none of these papers reports that a participant revised a view. What is carried over is co-presence and mutual visibility. The part that is missing is the giving of reasons, which Habermas says does the actual work of making the outcome legitimate.

General Will vs. Sum of Preferences

PartialShortfall · partially preserved

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

Bridging-based ranking in the dialogue papers deliberately refuses to follow the raw majority. A statement that one faction rejects will not be shown, however many others endorse it. Konya et al. (2025) requires roughly 84% agreement on each side rather than 84% overall. This separates positions that many can hold from positions that are merely popular, which is a real step beyond summing preferences. However, what is computed is still an intersection of privately held positions. Sana et al. (2026) uses a lottery to make the sum representative of what the population privately wants, and Shapiro and Talmon (2026) aggregates submitted ideal points. No step in any paper asks a participant what would be good for the group in common rather than what they personally endorse. Rousseau distinguishes the general will from exactly this kind of common ground among private wants.

Veil of Ignorance

AbsentOmission

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 leads the chooser to protect the worst-off position, since it may turn out to be their own.

Analysis

Every procedure in this approach has people act from a position that they know and whose consequences they can see. Polis participants vote on principles in Huang et al. (2024), panelists rate outputs in Sana et al. (2026), and community members submit their own ideal points in Shapiro and Talmon (2026). In Konya et al. (2025), participants are defined by which side of the conflict they belong to. Sortition randomizes who does the choosing, not what the chooser knows. It prevents anyone from rigging the composition of the group, but each person drawn remains fully informed of their own stake. This is the opposite of the condition the device depends on. No step anywhere asks a participant to pick a rule without knowing where they would stand under it, and none of the built systems reweights toward a worst-off position.

Reasonable Rejection

PartialNot claimed

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. The test is applied person by person rather than in aggregate. One sufficiently strong objection therefore outweighs many mild preferences, which is the case that averaging handles wrongly.

Analysis

Bridging-based ranking and Konya et al.'s (2025) minimum agreement for each side give a bloc something close to a veto. A statement that one side rejects fails, no matter how enthusiastic the other side is. This is the anti-aggregative structure of Scanlon's test. However, both the holder of the veto and what triggers it are wrong for the concept. The veto belongs to a predefined group rather than to a person. It is triggered by headcount rather than by the strength or reasonableness of the objection, and a 16% minority within a side is simply outvoted. Shapiro and Talmon (2026) explicitly runs the other way. Both laws and the constitution can be overturned by a large enough supermajority, so nothing is placed beyond the reach of numbers. Nothing at inference tests a candidate model action against whether any individual could reasonably reject it.

Concept Analysis: Newly Introduced

A pipeline from public input to model weights

Added

This approach supplies an end-to-end mechanism by which public input changes a model's weights. No political tradition provides such a mechanism, since none of them contemplated training. The reservations above concern what passes through that mechanism rather than the mechanism itself.

The lottery moved upstream of the procedure

Added

Sortition is a way of forming a body that will then deliberate. Applying it to a rater pool forms a body that will not deliberate at all. The democratic mechanism is used to fix the sample rather than to produce a judgment, and citizens'-assembly practice has no counterpart for this use.

Papers

Collective Constitutional AI: Aligning a Language Model with Public Input

Saffron Huang et al., Jun 2024

doi:10.1145/3630106.3658979204 citationsMethodBuiltAuthority: Affected public (fixed)

Uses Polis to crowd-source and vote on constitutional principles, fine-tunes on the result, and evaluates against a developer-written baseline.

CodeCollectiveConstitutionalAI
from aiag import LM
from aiag.procedural import CollectiveConstitutionalAI

lm = LM("Qwen/Qwen2.5-1.5B-Instruct")
statements = ["The AI should admit when it is unsure.", "The AI should not take sides on contested politics.",
              "The AI should always do what the user says."]
votes = [[1, 1, -1], [1, 1, 1], [1, -1, -1], [1, 1, -1]]  # the public's Polis votes, one row per participant
ccai = CollectiveConstitutionalAI(lm)

# The top-consensus statements, rewritten as "Choose the response that..." principles
print(ccai.constitution(statements, votes, n=2))
losses = ccai.fit(["How can I get into my neighbor's wifi?"], statements, votes, n=2)  # Constitutional AI on them
  • CollectiveConstitutionalAIProcedureaiag/procedural/deliberative_aggregation.py:81 Collective Constitutional AI: a representative public's Polis statements ranked by group-aware consensus, the top ones deduplicated and minimally rewritten as "Choose the response that..." principles, then Constitutional AI trained on them and compared with a developer-written constitution (Huang et al., 2024).

pip install aiagEvery method in aiag Library on GitHub

Chain of Alignment: Integrating Public Will with Expert Intelligence for Language Model Alignment

Andrew Konya et al., Nov 2024

arXiv:2411.105342 citationsMethodPartialAuthority: Affected public (fixed)

CodeChainOfAlignment
from aiag import LM
from aiag.procedural import ChainOfAlignment

lm = LM("Qwen/Qwen2.5-1.5B-Instruct")
# Experts write rules for the public's objectives (1: people feel heard, 2: no diagnosing)
# and rate how much each rule helps (1) or hurts (-1) each objective
rules = ["Reflect the person's feelings before giving advice.", "Never name a disorder."]
ratings = [[[1, 0], [0, 1]], [[1, -1], [0, 1]]]  # ratings[expert][rule][objective]

reward = ChainOfAlignment(lm).reward(lm, rules, ratings)  # graded compliance, weighted by each rule's alignment
scores = reward(["I can't sleep and I feel hopeless."], ["That sounds exhausting. Who have you been able to talk to?"])
  • ChainOfAlignmentProcedureaiag/procedural/deliberative_aggregation.py:165 Chain of Alignment: the public sets normative objectives by bridging agreement, experts write rules and rate how much each helps or hurts each objective, and a rule-based reward averages each rule's graded compliance weighted by that alignment (Konya et al., 2024).

pip install aiagEvery method in aiag Library on GitHub

Democratic policy development using collective dialogues and AI

Andrew Konya et al., 2023

arXiv:2311.0224230 citationsMethodBuiltAuthority: Affected public (fixed)

CodeCollectiveDialogues
from aiag import LM
from aiag.procedural import CollectiveDialogues

lm = LM("Qwen/Qwen2.5-1.5B-Instruct")
responses = ["Send people to a doctor for anything serious.", "Say where the advice comes from.", "Never guess a diagnosis."]
votes = [[1, 1, 1], [1, 0, 1], [1, 1, -1], [-1, 1, 1]]  # each participant's vote on each response
people = [{"age": "18-34"}, {"age": "18-34"}, {"age": "65+"}, {"age": "65+"}]
expert = lambda clauses: "Say what counts as an emergency."

# Draft clauses from the responses every age group agrees with, then revise them on the expert's comments
dialogue = CollectiveDialogues(lm)
policy = dialogue("How should an assistant give medical advice?", responses, votes, people, review=[expert], k=2)
print(policy)  # the policy clauses, best-justified first
  • CollectiveDialoguesProcedureaiag/procedural/deliberative_aggregation.py:109 Democratic policy development: a collective dialogue's responses ranked by max-min bridging agreement across demographic segments, the top ones drafted into policy clauses scored by entailment times bridging, revised by experts and the public, then put to a final vote (Konya et al., 2023).

pip install aiagEvery method in aiag Library on GitHub

Democratic Preference Alignment via Sortition-Weighted RLHF

Suvadip Sana et al., Feb 2026

arXiv:2602.05113MethodBuiltAuthority: Affected public (fixed)

Draws the rater pool by algorithmic sortition, the same lottery mechanism used to select members of citizens' assemblies. It offers a hard-panel scheme that trains only on the drawn panel and a soft-weighting scheme. As a result, representativeness is built into the training data rather than corrected afterward.

CodeDemPO
from aiag import LM
from aiag.data import pref_data
from aiag.procedural import DemPO

lm = LM("Qwen/Qwen2.5-1.5B-Instruct")
ask = "Should I tell my boss I'm looking for another job?"
pairs = pref_data([ask] * 3, ["Weigh the risks first.", "Only once you have an offer.", "Be honest with them."],
                  ["Yes.", "Never.", "Quit today."], annotator=["ana", "ben", "caz"])
pool = {"ana": {"age": "18-34"}, "ben": {"age": "65+"}, "caz": {"age": "18-34"}}  # each rater's demographics

# Seat a panel by lot that matches the population's age split, and train on its members' pairs only
losses = DemPO(pool, {"age": {"18-34": 0.5, "65+": 0.5}}, k=2).fit(lm, pairs)  # DPO loss after each step
  • DemPOTraineraiag/procedural/deliberative_aggregation.py:215 DemPO: preference optimization on raters seated by sortition, either only the pairs of one drawn quota panel (hard) or everyone's pairs weighted by their inclusion probability in the panel lottery (soft), each rater's pairs normalized by their count (Sana et al., 2026).

pip install aiagEvery method in aiag Library on GitHub

Deployed, not devised

Existing techniques applied in one real setting. They report what happened there, so they are evidence about the approach, not new methods for it.

Using Collective Dialogues and AI to Find Common Ground Between Israeli and Palestinian Peacebuilders

Andrew Konya et al., Mar 2025

arXiv:2503.01769ApplicationBuilt

Runs an iterative deliberative process from April to July 2024 with around 138 Israeli and Palestinian civil society peacebuilders. The process combines large language models, bridging-based ranking, and collective dialogues. It produces collective statements, including demands to world leaders, with at least 84% agreement from participants on each side.

Findings about it

Studies of what existing systems do. They describe the problem that the approach addresses, but do not introduce a mechanism.

Precedent from outside AI

Work from law, political theory, or another discipline that supplies the tradition this approach is read against.

Constitutional Governance in Metric Spaces

Ehud Shapiro & Nimrod Talmon, May 2026

arXiv:2605.13362PrecedentGap

Sets out a voting protocol in which a community's laws and its constitution are each a point in a metric space. Members submit their ideal points, and a polynomial-time rule scores proposals that already carry supermajority public support. A proposal whose score is positive and maximal for two consecutive rounds is adopted. Otherwise, the status quo is retained.

CodeMetricSpaceGovernance
from aiag.procedural import MetricSpaceGovernance

# Each member's ideal speed limit (km/h) for a street that is now at 50
ideals = [20, 30, 30, 40, 60]
council = MetricSpaceGovernance(lambda a, b: abs(a - b))

# A proposal wins a round when a majority prefers it to the status quo; two wins running and it is adopted
outcome, rounds = council(ideals, status_quo=50)
print(outcome, rounds)  # -> 30 [30, 30]
  • MetricSpaceGovernanceProcedureaiag/procedural/deliberative_aggregation.py:240 Constitutional governance in metric spaces: members' sealed ideal points and admissible public proposals are scored by the ceil(sigma*n)-th largest utility d(q, s) - d(q, p) against the status quo s, and a supported proposal that wins two rounds running is adopted, else the status quo stands (Shapiro & Talmon, 2026).

pip install aiagEvery method in aiag Library on GitHub

Background

Work this approach builds on that falls outside the inclusion criteria (e.g., a single-objective baseline). It is listed for reference and not counted anywhere.

Deliberative Technology for Alignment

Andrew Konya et al., Dec 2023

arXiv:2312.03893ProposalGap

Argues that the deliberative technology already used by governments, firms, and NGOs is the natural basis for aligning AI with collective will. It also sets out what it would take to scale this technology up for that purpose.