Position: AI Should Facilitate Democratic Deliberation at Scale

Summary

This position paper argues that AI systems — and LLMs in particular — should be designed to facilitate democratic deliberation at scale, rather than to substitute machine judgment for citizen choice or to serve engagement metrics. The conceptual move is to reframe the question from "should AI decide?" (as in liquid democracy or aggregation-only platforms) to "how can AI reduce the cognitive, social, platform-design, and market-incentive frictions that keep meaningful deliberation from scaling?" The authors organize this reframing around four guiding principles — agency, respect, equality, and augmentation — and map them onto a workflow of elicitation, reflection, structured exchange, and synthesis, illustrated by capabilities such as Socratic prompting, opinion visualization, bridging-based ranking, and civility moderation. They also engage seriously with sycophancy, WEIRD bias, over-reliance, and alignment as challenges that must be resolved before scaled deliberation is genuinely achievable, and they distinguish deliberation-supporting AI (in the informal public sphere) from decision-making AI (which they explicitly do not endorse). The paper is a useful synthesis of a fast-moving literature and it makes a defensible normative case that pro-democratic AI is a coherent design target rather than a category error. The main conceptual limitation is that several of the strongest formal moves — the claim that the four principles are necessary and sufficient, the claim that deliberative technology already exists, and the treatment of market-incentive frictions — are asserted at a level of strength the paper's own arguments do not sustain, and the market-incentive category in particular is enumerated but never actually addressed by any capability in the Section 5 review. The empirical evidence marshalled for the AI-facilitation mechanisms is also drawn almost entirely from WEIRD populations even while the paper flags WEIRD alignment as a risk, and "at scale" is never operationally defined. The contribution is real but the framing overreaches in ways that a revision could tighten without abandoning the argument.

Strengths

• Clear conceptual move. The distinction between AI that discovers consensus among human-generated positions and AI that generates consensus statements is cleanly stated and used consistently, and it does real work in demarcating the augmentation principle from decision-making AI.
• Serious engagement with failure modes. Section 6 (sycophancy, training bias, over-reliance, alignment) does not soft-pedal the risks and explicitly ties them back to the four principles rather than treating them as external caveats.
• Contrast with alternatives is fair. The alternative-views section engages liquid democracy, in-person deliberation, non-AI online platforms, and outright skepticism with distinct arguments for each, rather than dismissing them.
• Grounded taxonomy of frictions. The cognitive/social/platform/market taxonomy is well-motivated by citations to established literatures (Kahneman, Kahan, Iyengar, Gillespie) and gives the paper a usable analytic scaffold.
• Principled scope choice. Explicitly restricting the argument to the informal public sphere and keeping AI "at arm's length from formal democratic decision-making" is the right move given the state of the technology and is consistently maintained.

Weaknesses

• "Necessary and sufficient" overreach. The claim that the four principles are both necessary and sufficient for pro-democratic AI is the paper's strongest formal assertion but is defended only by arguing that transparency and user-satisfaction are inadequate as competing organizing principles. That establishes coherence and priority, not sufficiency — nothing rules out that additional independent requirements (factual accuracy, robustness to adversarial manipulation, non-domination by the hosting entity) might be needed.
• Market-incentive frictions enumerated but never addressed. Section 3 carefully separates platform-design from market-incentive frictions (attention economies, ad-based revenue), but Section 5.3 collapses them into "system frictions" and its content targets only design levers. No AI capability in the review offers any leverage on the underlying business-model dynamic; deliberation platforms are simply proposed as an opt-in alternative that coexists with engagement-maximizing platforms.
• "Technology exists" contradicts the paper's own agenda. The Section 8 assertion that the technology to scale deliberative democracy exists is in direct tension with the same section's calls to develop the missing deliberative-quality benchmarks, mitigate sycophancy, correct WEIRD bias, and build agency-preservation protocols — all still open problems by the paper's own account.
• Empirical base is WEIRD-centric despite flagging WEIRD bias. Almost every piece of positive evidence for the AI-facilitation mechanisms (Pol.is, deliberation.io, adhocracy+, America in One Room, Argyle et al., Braley et al., Fishkin et al.) is US or European; the sole non-Western citation is Sub-Saharan African deliberative polling, used to defend online parity. Transporting these findings to non-WEIRD contexts is treated as unproblematic.
• "At scale" is not operationalized. The phrase carries the paper's central rhetorical weight but there is no threshold on N, structure, or duration that distinguishes scaled deliberation from richer aggregation-with-scaffolding, which matters because the paper explicitly contrasts the two.
• Governance / who-runs-the-AI is unengaged. Section 3 attributes platform frictions to the interests of the entities that build and monetize platforms, but Section 5 proposes embedding AI capabilities into deliberation platforms without a serious discussion of hosting, moderation policy, model choice, and update authority. Open-source releases are noted but not analyzed as a governance regime.
• Load-bearing narrower claims are overstated. Sycophancy is reframed from "validating regardless of correctness" to "amplifying rather than moderating extreme views"; "equally biased in opposite directions" asserts a quantitative symmetry not obviously shown by the cited experiments; "AI systems designed to maximize user satisfaction will undermine deliberative goals" is stronger than the paper's more careful language elsewhere.
• Costs, analogies, and citations are loose in places. The ~10x/year inference-cost trend is presented without a time window or task-specific caveat; the "prohibitively costly to staff human moderators" claim is supported by citations that are actually about toxicity-detection capability, not moderation economics; the analogy between abuse-reduction nudges and reflection-elicitation prompts is weaker than the citation placement suggests.
• Local inference gaps. "Correcting these misperceptions" generalizes from perception-based social frictions to social frictions broadly; the subversion-dilemma citation "formalizes these dynamics" but formalizes a different dynamic than the preceding sentence describes; the "thus" before the exclusion sentence in Section 5.1 does not follow from its antecedents; "at the expense of universal reason" imports a philosophical notion in tension with the paper's own pluralist framing.
• Benchmark call-to-action is not concrete. The paper asks the ML community to develop deliberative-quality benchmarks but does not sketch a single evaluation protocol — tasks, data, scoring, interfaces — despite Section 5 laying out several existing measurement instruments that could be adapted.
• Minor issues. Figure 1 and Figure 2 have substantially overlapping content that the text does not clearly differentiate; "substantial propotions" is both a typo and semantically vague; Bai et al. (2025a) is invoked as both a benign capability and a persuasion risk without disambiguation; the "identity validation" concept is introduced in quotation marks without a citation attached to its definition.

Reproducibility & code

• No accompanying artifacts. The manuscript is a position paper and does not release code, data, or a benchmark scaffold; this is defensible on its own but is in mild tension with its call for the ML community to build deliberative-quality benchmarks and its endorsement of open-source deliberation infrastructure. Even a small companion checklist mapping the four principles to machine-checkable criteria would raise the actionability of the piece.
• Working-paper dependence. Several of the strongest empirical anchors — for Socratic-dialogue moderation (Enríquez, 2025), consensus-generation on deliberation platforms (Braley et al., 2025a; 2025b), comment ordering (Chen et al., 2025), deliberation.io (Pei et al., 2025), online-participation field experiments (Ahmad, 2025), and sycophancy-and-extremity (Rathje et al., 2025) — are Working Papers not yet publicly available. Several are authored or co-authored by members of the present author team, which is not unusual for a synthesis piece from an active research group but concentrates unverifiable evidence around the strongest capability claims.

Recommended Changes

Essential

• Rewrite the "necessary and sufficient" claim. Soften to a coherence/adequacy claim about the four principles, or provide an actual argument that (i) no additional independent principle (factual accuracy, robustness, non-domination by hosting entity) is required and (ii) dropping any single principle defeats pro-democratic status. See the "Necessary and sufficient overreach" weakness.
• Either address market-incentive frictions in Section 5 or acknowledge that they are out of scope. As it stands, one of the four friction categories introduced in Section 3 receives no matching AI capability in Section 5, and the section-header rename to "system frictions" masks this. The cleanest fix is to state explicitly that the AI capabilities proposed do not target the market/business-model dynamic, and to discuss what would.
• Rewrite the closing "technology exists" line to match the research agenda. Replace with a claim about the building blocks existing while enumerated open problems (benchmarks, sycophancy mitigation, WEIRD-bias correction, agency-preservation protocols) remain.
• Restrict scope or add non-WEIRD evidence. Either scope the claims to WEIRD contexts explicitly, or bring in additional non-Western deployment evidence to support the generalization. Do not rely on a single Sub-Saharan Africa citation to answer both the in-person-vs-online comparison and the WEIRD-bias transport question.
• Operationalize "at scale." Give a target N (or range) and, more importantly, the qualitative conditions on structure and duration that distinguish scaled deliberation from richer aggregation, since the paper explicitly contrasts the two.
• Add a governance paragraph. Address who hosts, updates, and configures the AI capabilities — and what open-source release actually guarantees under real deployment — given the paper's own critique of platform-hosting incentives in Section 3.

Suggested

• Calibrate the strong local claims. Weaken "will undermine" to "can undermine", "cannot be compressed" to "cannot be arbitrarily compressed" or draw an explicit boundary between compressible and non-compressible cognitive operations, and reframe the sycophancy summary to separate "validating regardless of correctness" from "amplifying extreme views". Attribute the "equally biased in opposite directions" phrase to the specific experiment, or soften it.
• Quantify or citation-check the loose empirical claims. State the time window and task class for the ~10x/year inference-cost trend; move the toxicity-detection citations to support the AI-moderation capability rather than the human-moderator cost claim, and separately support the cost comparison; strengthen the analogy from abuse-nudges (Katsaros et al., 2022) to reflection elicitation with evidence tied to reflection specifically. Fix the "propotions" typo and quantify the demographic-bias summary.
• Fix the local inference gaps. Restrict the misperception-correction conclusion to informational social frictions; rewrite the "formalizes these dynamics" clause so its referent matches what Braley et al. (2023) actually formalize; replace the "thus" before the exclusion sentence with a connective that does not falsely signal entailment; replace "at the expense of universal reason" with an equality/inclusiveness framing consistent with the rest of the paper.
• Clarify Figures 1 and 2. Either merge into one figure or state in the captions what Figure 2 shows that Figure 1 does not.
• Sketch a benchmark protocol. Given the call to action, propose even a minimal evaluation harness (task, data, scoring) that operationalizes one of the Section 5 measurement instruments (e.g., a bridging-score benchmark).
• Disambiguate dual-use citations. In the paragraph invoking Bai et al. (2025a) as a persuasion risk, note explicitly which of their conditions supports the risk claim and which the benign articulation claim. Attach the citation for "identity validation" directly to its definition, or mark the definition as the authors' paraphrase.
• Address the working-paper dependence. Where feasible, prefer published citations for the strongest capability claims; where not, briefly acknowledge that several of the mechanism-supporting studies are not yet publicly available.