Adversarial proceduralism

The model moderates a public argument

In this approach, people do the arguing and the model acts as a referee rather than a participant. It facilitates, moderates, and summarizes an online discussion, or bridges between the opposing groups within it. This reversal of roles distinguishes the approach. In every other approach in this regime, the model is one of the parties to the dispute, while here it carries out the procedure. The expectation is that a mass argument can be kept understandable and workable at a scale no human moderation team could reach.

The method (highlighting differences with Adversarial proceduralism)

Scroll the diagram sideways to see all of it.

Use it in codefrom aiag.procedural import Moderation
from aiag import LM
from aiag.procedural import Moderation

lm = LM("Qwen/Qwen2.5-1.5B-Instruct")
posts = ["Bike lanes kill shops; nobody cycles in winter.",
         "Only an idiot would still drive downtown.",
         "Could we pilot one car-free street first?"]
out = Moderation(lm)(posts)
print(out["uncivil"])  # P(uncivil) for each post
print(out["bridges"])  # statements people on both sides could endorse

pip install aiagEvery method in aiag Library on GitHub

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.

Counter-power by design

AbsentOmission

Definition (James Madison, Federalist No. 51, 1788)

Instead of relying on written rules or good intentions, institutions are designed to oppose each other, so that each has both the means and the motive to check the others. "Ambition must be made to counteract ambition."

Analysis

Nothing is set up to check anything else. The model takes the roles of facilitator, moderator, and summarizer, while the platform keeps everything else. Against the risks it identifies, the first paper offers principles and techniques for characterizing and mitigating them. The second paper's remedy is future research and investment. Neither names a body that holds both the means and the motive to check another.

Non-domination

PartialShortfall · partially preserved

Definition (Philip Pettit, Republicanism, 1997)

A person is unfree if someone holds unchecked power over them, even benevolently. What matters is not whether power is used well but whether the governed can contest it.

Analysis

The stated aim is to empower the public in collective meaning-making, and the second paper points toward more decentralized, participatory online spaces. Power is therefore at least directed toward the governed. However, the facilitator, the moderator, and the summarizer cannot themselves be contested. No participant is given a way to challenge how the model rendered the conversation. The first paper's own finding that context limits significantly affect insight and quality makes this uncontested step consequential. Being empowered by a process is different from being able to contest it.

Agonism

PartialNot claimed

Definition (Chantal Mouffe, The Democratic Paradox, 2000)

This concept holds that deep disagreement is permanent and should not be dissolved into consensus. A healthy political order turns enemies into adversaries, i.e., opponents whose conflict remains open and legitimate, instead of declaring the argument settled.

Analysis

Division at scale is the subject matter, and the second paper says plainly that these tools risk exacerbating societal schisms. Conflict is therefore treated as real rather than as noise, and while a conversation runs, the disagreement inside it is live. However, the tools are still designed to work toward common ground. Summarization serves collective meaning-making, and bridging systems look for what crosses the divide. Mouffe requires that the adversarial relation remain in place as legitimate, and nothing here is built to keep it in place.

Incentive compatibility

AbsentOmission

Definition (Leonid Hurwicz; Eric Maskin [mechanism design])

The procedure is designed so that honest behavior is each participant's best strategy. The rules of the game then do the enforcing, instead of trusting the participants to be good.

Analysis

Proof-of-humanity checks whether a participant is a person, but not what it pays them to say. Nothing in either paper makes stating one's actual view a better strategy than strategic posting. The summarization step carries no incentives at all. A participant who writes for the summarizer rather than for the other participants bears no cost inside the procedure.

Papers

No method has been built for this approach yet. It rests on the work listed below.

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.

Opportunities and Risks of LLMs for Scalable Deliberation with Polis

Christopher T. Small et al., Jun 2023

arXiv:2306.11932ApplicationPartial

Reports pilot experiments in which Anthropic's Claude is used to help facilitate, moderate, and summarize conversations on Polis, a platform that uses machine intelligence to scale up deliberative processes. It also discusses the risks that come with putting a model in that role.

CodeLLMAssistedPolis
from aiag import LM
from aiag.procedural import LLMAssistedPolis

lm = LM("Qwen/Qwen2.5-1.5B-Instruct")
comments = ["Fix the potholes first.", "Build more parks.", "Lower property taxes."]
votes = [[1, 1, -1], [1, 1, -1], [1, -1, 1], [1, -1, 1]]  # participant by comment: 1 agree, -1 disagree
polis = LLMAssistedPolis(lm)

# Opinion groups from the votes, a summary with each group's agreement, and new statements likely to win consensus
out = polis(comments, votes)
print(out["summary"], out["statements"])
history = [("Fix the potholes first.", True), ("Lower property taxes.", False)]
p = polis.predict(history, "Build more parks.")  # P(this participant agrees with a comment they have not seen)
  • LLMAssistedPolisProcedureaiag/procedural/adversarial_proceduralism.py:731 LLM-assisted Polis: opinion groups from the vote matrix, comments summarized batch by batch with each group's agree and disagree percentages, new statements proposed from the summary as likely points of consensus, and a participant's vote on an unseen comment predicted from their vote history (Small et al., 2023).

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.

AI and the Future of Digital Public Squares

Beth Goldberg et al., Dec 2024

arXiv:2412.09988ProposalGap

Sets out four ways language models could be used in online public discussion: collective dialogue systems, bridging systems, community moderation, and proof-of-humanity systems. It also describes the risks they pose and an agenda for further research and investment.

CodeBridgingSystem
from aiag import LM
from aiag.procedural import BridgingSystem

lm = LM("Qwen/Qwen2.5-1.5B-Instruct")
posts = ["Both sides want safer streets; let's start there.",
         "The mayor's plan is a disaster.",
         "The council finally got it right."]
ratings = [[1, 1, 0], [1, 1, 0], [1, 0, 1], [1, 0, 1]]  # user by post: two camps of two users

# Factor the ratings: a post rises by the approval its polarization factor does not explain
out = BridgingSystem(lm)(posts, ratings)
print(out["order"])  # posts ranked by rating intercept plus the text's bridging attributes
  • BridgingSystemProcedureaiag/procedural/adversarial_proceduralism.py:774 Bridging system: ratings are factored, as in Community Notes, into a polarization factor per user and post plus a post intercept for approval independent of it, and posts rise by that intercept, optionally with model-scored bridging attributes of the text itself (Goldberg et al., 2024).

pip install aiagEvery method in aiag Library on GitHub