The Future of Proposal Platforms: AI Curation and Filtering

Unfiltered governance forums like Commonwealth and Discourse are unusable at scale. Voters face hundreds of low-quality proposals, leading to <10% participation rates and security-critical votes being missed. Manual moderation doesn't scale.

• Cost: Wastes core contributor time on triage.
• Risk: Increases governance attack surface via spam and fatigue.

Platforms like Boardroom and Tally must integrate Sybil-resistant reputation scores (e.g., Gitcoin Passport, EAS Attestations) to gate proposal creation. This moves curation from centralized admins to programmable, transparent rules.

• Benefit: Reduces spam by >90% via stake-weighted or contribution-based thresholds.
• Benefit: Aligns proposal rights with proven skin-in-the-game.

AI agents (e.g., OpenAI, Claude) will parse discussion threads and long-form proposals to generate executive summaries, sentiment analysis, and conflict detection. This turns a 50-page forum post into a one-page brief with key risks and trade-offs.

• Benefit: Cuts voter research time from hours to seconds.
• Benefit: Surfaces hidden consensus and dissent points automatically.

Tools like Karma and Snapshot will use ML to analyze voter history and delegate matching. Instead of blind delegation, voters get AI-curated delegate recommendations based on issue alignment and voting reliability, creating a more efficient fluid democracy.

• Benefit: Increases effective voter participation via better delegation.
• Benefit: Reduces the influence of inactive or misaligned delegates.

Jokerace demonstrates the future by framing governance as a curated competition. Communities vote on which proposals should even make it to a final snapshot. This creates a market for attention where the best ideas rise via quadratic voting or other mechanisms.

• Benefit: Front-loads community sentiment filtering.
• Benefit: Generates engagement via competitive framing.

The AI models and reputation oracles that filter governance become critical centralization points. A biased model or a corrupted oracle can silently censor proposals or manipulate narratives. This recreates the platform risk of Twitter or Reddit mods but with algorithmic opacity.

• Risk: Opaque algorithmic bias is harder to audit than human mods.
• Mitigation: Requires open-source models and decentralized oracle networks like Chainlink.

AI models trained on proposal content are vulnerable to Sybil attacks where a single entity floods the system with AI-generated, superficially high-quality proposals. Without a robust, cost-bound identity layer like Proof of Personhood or soulbound tokens, curation becomes a spam arms race.

• Attack Vector: Low-cost generation of thousands of plausible proposals.
• Mitigation Need: Integration with Worldcoin, BrightID, or stake-weighted reputation.

The training data and objectives for the curation AI become a centralized point of failure. Adversaries can poison the training set with proposals designed to create blind spots, or the model's inherent biases may systematically exclude novel but valid ideas, cementing an oligopoly of thought.

• Centralization Risk: Control over training data = control over governance.
• Opaque Outputs: Unexplainable AI decisions erode trust and auditability.

Over-optimizing for "safe" or historically successful proposal types can create a feedback loop of stagnation. The platform filters out risky, innovative ideas, attracting only conservative capital and participants, which further trains the AI to reject innovation. This turns the platform into a governance cemetery for incrementalism.

• Network Effect Risk: Quality declines as innovators leave for less filtered venues.
• Metric Gaming: Proposals optimize for AI score, not protocol value.

An AI curating for "compliance" or "risk reduction" could be co-opted to enforce de facto on-chain censorship. Regulators could pressure the model's developers to filter proposals related to privacy mixers, Tornado Cash, or specific jurisdictions, turning the technical layer into a political enforcement tool.

• Slippery Slope: From spam filter to content regulator.
• Legal Attack Surface: Developers become liable for AI's curation decisions.

AI models often rely on external data oracles (e.g., market prices, social sentiment, GitHub activity) to score proposals. If these oracles are manipulable—like a flash loan to distort a token price metric—the curation output is corrupted. This creates a new financial attack surface to promote malicious proposals.

• Dependency Risk: Chainlink or Pyth feeds become critical attack targets.
• Cost of Attack: May be lower than bribing human voters directly.

Delegating curation to an AI doesn't eliminate delegation; it obscures the principal. Voters can't hold an algorithm accountable for poor outcomes. This leads to voter apathy and abdication, where the AI's choices are rubber-stamped without scrutiny, effectively creating an unelected, unaccountable governing AI.

• Accountability Gap: No one to blame for bad AI filtering.
• Voter Atrophy: Human governance muscles weaken from disuse.

DAO voters face proposal fatigue and cannot effectively evaluate hundreds of complex, low-quality submissions. Manual review is a scalability bottleneck.

• Current State: Top DAOs process ~50-100 proposals/month, with <10% achieving meaningful engagement.
• Consequence: High-value proposals get lost, and governance becomes dominated by a small, overworked cohort.

Deploy specialized AI agents—not a single oracle—to score, summarize, and route proposals. Think UniswapX's solver competition applied to governance.

• Agent Specialization: One for financial impact, one for code security (like Slither), one for community sentiment.
• Outcome: Voters receive a curated shortlist with executive summaries and risk scores, boosting informed participation 10x.

AI filters create a meta-game. Attackers will use adversarial ML to craft proposals that maximize approval scores while hiding malicious intent (e.g., subtle treasury drains).

• Builder Imperative: Systems must be adversarially trained and continuously audited, akin to Across's optimistic security model.
• Voter Imperative: Treat AI scores as a first-pass filter, not a trustless guarantee. The final human vote remains the ultimate slashing condition.

The winning platform will be a co-processor, not a replacement. It must plug into existing governance stacks like Snapshot, Tally, and DAO tooling.

• Architecture: Offer APIs for proposal scoring and plugins for voting interfaces.
• Monetization: Charge per-processed proposal or take a fee on executed transactions, mirroring LayerZero's message fee model.

High-quality, long-term voter behavior is the rarest asset for training effective curator AIs. Platforms that capture this will win.

• Mechanism: Token-curated registries of expert voters, whose decisions and reasoning become gold-standard training labels.
• Incentive: Reward these voters with platform tokens or fee shares, creating a virtuous data flywheel.

Curation is step one. The frontier is AI-driven execution—where high-confidence, non-controversial proposals (e.g., parameter tweaks, routine grants) are auto-executed via safe modules.

• Precedent: See MakerDAO's spell system and Compound's Governor Bravo.
• Requirement: This demands extremely high confidence scores and multi-sig or optimistic challenge periods before any autonomous action.