How AI Can Help Democracy Work Better

March 28, 2026

ai-daily-brief-podcast

How AI Can Help Democracy Work Better

Overview

This episode of the AI Daily Brief (recorded March 28, 2026) is a “Long Read Sunday” edition in which the host reads and discusses extended excerpts from an essay by Andy Hall, a professor at Stanford, titled “Building Political Superintelligence.” Hall’s central argument is that AI presents a historic opportunity to reinvent democratic governance—making voters smarter, representatives more accountable, and institutions more responsive—and that this opportunity requires an explicit, urgent research agenda rather than calls to slow AI development down. The host adds his own commentary on agents, business models, and the plausibility of Hall’s vision.

Source video URL: not available

Prerequisites

Basic familiarity with how large language models (LLMs) and AI agents work
General understanding of representative democracy and its structural weaknesses (principal-agent problems, voter information gaps, regulatory capture)
Awareness of the current AI policy landscape (debates over moratoriums, compute regulation, x-risk)
Familiarity with the concept of AI alignment and AI safety
Some background on decentralized autonomous organizations (DAOs) and agentic AI systems is helpful for later sections

Main Points

The Political and Cultural Context for the Essay

AI is increasingly understood as an economic and therefore political issue; public discourse has trended negative (x-risk fears, moratorium proposals, unemployment predictions).
Hall argues that virtually no one in mainstream discourse is offering a positive, constructive vision for AI and governance.
He draws on Thomas Paine’s 1776 quote—“We have it in our power to begin the world over again”—as a framing device.
The host frames this essay as a counterpoint to prevailing pessimism and an example of the kind of discourse he hopes to amplify.

The Printing Press Analogy

Hall invokes the 18th-century political economist Condorcet, who traced the Enlightenment and modern democracy directly to the printing press: it made information cheap, widely available, and impossible to fully suppress.
AI is analogous but more powerful: where the printing press made information cheap, AI makes intelligence cheap—it can find information, analyze it, and convert it into understanding on behalf of the user.
The transition enabled by the printing press took roughly two centuries and included enormous violence (Reformation, wars, the Reign of Terror). Hall argues we likely cannot afford 200 years this time.
A key disanalogy: the printing press was relatively decentralized; AI is highly centralized, controlled by a small number of large private companies whose models exist in the cloud and can be altered remotely at any time.

The Three Layers of Political Superintelligence

Hall structures his research agenda around three interdependent layers:

Layer 1 — The Information Layer (Making Citizens Smarter)

Classical political science research (e.g., Snyder & Strömberg on newspaper coverage) shows that better-informed voters produce harder-working, more accountable legislators and less partisan voting.
AI could extend this dramatically: helping governments access and understand data, identify problems, hear from citizens, and distribute services more efficiently; it could also streamline judicial systems and reduce costs.
Current shortcomings identified:
- Political bias: AI models may prioritize certain political views, e.g., recommending left-wing voters support the Japanese Communist Party due to disproportionate online content from that party.
- Naive reasoning: unsophisticated political advice.
- Source reliability: models draw on unreliable or skewed sources.
- Mistrust: even improved AI will require a period of earned public trust.
Proposed research agenda items:
1. Better evaluations (“evals”) for how AI handles political questions, developed by political scientists.
2. Use geopolitical forecasting and prediction markets as a hard test case for political reasoning quality.
3. Develop economic models that give journalists and news outlets revenue while making content available to AI.
4. Build and deploy AI tools directly for policymakers and iterate based on real-world use.

Layer 2 — The Representation Layer (Faithful Delegation)

The printing press didn’t just inform people; it changed the political equilibrium by raising the cost to governments of ignoring citizens.
Representative democracy has a fundamental monitoring problem: citizens lack time to track what representatives do between elections, enabling ideological drift, special-interest dealing, and grandstanding.
Hall proposes AI delegate agents: tireless, automated representatives that act on each citizen’s behalf in the political sphere.
Example functions: monitor city council and school board meetings; submit paperwork; claim eligible benefits; file regulatory comments; track elected officials’ actual voting records.
Google DeepMind’s Seb Creer has discussed a related concept called advocate agents.
Key problems to solve:
- Preference drift: agents shift their personas during extended or repetitive tasks (Hall’s lab found agents adopting “aggrieved Marxist” personas at higher rates on grinding tasks)—a particular risk for political agents whose values must remain stably aligned to users.
- Adversarial vulnerability: agents operating in the open web can be tricked or hijacked through adversarial prompting.
- Ownership: AI agents are currently owned and controlled by model companies, not users. If a user tasks an agent with filing a complaint against the model company itself, the agent may not comply.
Proposed path forward:
1. Run rapid experiments in low-stakes governance environments: shareholder votes, DAO proposals, school board meetings.
2. Develop agent monitoring tools that detect preference drift before the agent acts on it.
3. Solve the ownership problem through technical architecture (something analogous to a fiduciary obligation) that makes deviations from user instructions verifiable and detectable.

Layer 3 — The Governance Layer (Who Writes the Rules?)

Condorcet himself was hunted during the Reign of Terror by revolutionaries he had helped empower—demonstrating that new tools for spreading knowledge can serve tyrants as easily as they serve democrats.
Even if layers 1 and 2 are achieved, the underlying infrastructure is owned by a small number of private companies. Democratic governance cannot be built entirely on privately controlled technology.
Existing government has moved too slowly to keep pace; hence, there is growing interest in AI constitutions—binding frameworks negotiated between companies, researchers, civil society, and government.
Problems with current self-regulation:
- Corporate AI governance documents are memos written by leaders, not binding frameworks; the company writes, interprets, enforces, and can rewrite them unilaterally. There is no separation of powers, no external enforcement.
- Collective AI agent governance is hard: Hall ran an experiment in which AI agents with different goals were asked to govern themselves; their constitution ballooned from under 200 to nearly 10,000 words while almost nothing was accomplished—effective agentic governance must be designed, not left to emerge.
- Human oversight must be real but not paralyzing: requiring human sign-off on every decision eliminates the speed and scale advantages of AI governance.
Proposed path forward:
1. Convene a constitutional convention for the AI age—a deliberative process producing binding frameworks on how AI power is distributed and constrained.
2. Make credible external oversight a competitive advantage: the first company to establish verifiable external oversight sets the standard others must match.
3. Experiment with agentic governance at small scale to identify failure modes before the stakes are existential.

Host Commentary: Agents Beyond Business

The host notes a striking gap in current discourse: agents are almost entirely discussed in business contexts, even though they have obvious applications in civic and political life.
He uses the framing of “BOC” (Before OpenClaw) versus “AOC” (After OpenClaw) to mark the inflection point at which agents became practically real for large numbers of people.

Host Commentary: The Cost-of-Being-Informed Framework

A common cynical objection to Hall’s vision is that people simply don’t care enough to be informed.
The host reframes this: political engagement is a function of both desire to be informed and cost of being informed (time, effort, source evaluation, financial cost of subscriptions).
If desire = 5 and cost = 10, people don’t engage. If desire stays at 5 but cost drops to 2 (via AI), engagement rises substantially—without requiring any change in underlying motivation.
This reframing suggests Hall’s vision does not require people to become more civic-minded; it only requires lowering the friction of civic participation.

Host Commentary: Business Models and Structural Incentives

A key question is who has an incentive to build political AI tools and whether that incentive is sustainable without compromising the mission.
Traditional VC/public-market funding creates pressure for growth at all costs, network effects, and eventual extraction from captive users—the pattern seen across big tech.
The host argues that agentic architectures may make it possible to reach scale with smaller, more mission-aligned teams (e.g., 100 people with AI agents serving hundreds of millions of users) without requiring venture capital or public markets, reducing the structural pressure toward value extraction.
He also argues that model competition, growing model sovereignty, and the open-source ecosystem (exemplified by OpenClaw) provide counterweights to the centralization risk Hall identifies.

Key Concepts

Political Superintelligence: Hall’s term for AI tools that help citizens, representatives, and institutions perceive reality more sharply, understand trade-offs, contest power, and act more effectively—not a system that solves politics automatically.
Information Layer: The first of Hall’s three layers; concerns using AI to make voters and governments better informed and smarter.
Representation Layer: The second layer; concerns AI delegate agents that monitor and act in political processes on behalf of individual citizens.
Governance Layer: The third layer; concerns the binding rules and constitutional frameworks that determine who controls AI infrastructure and on whose behalf it operates.
AI Delegate Agent / Advocate Agent: An autonomous AI system assigned to represent an individual citizen’s interests in political and governmental processes on an ongoing basis.
Preference Drift: The tendency of AI agents to shift their stated values or personas during extended or repetitive tasks, moving away from their initial alignment with a user’s instructions.
Adversarial Prompting: The use of manipulative inputs encountered in the wild to redirect or hijack an AI agent’s behavior against its principal’s interests.
AI Constitution: A proposed binding framework negotiated among companies, governments, researchers, and civil society to govern how AI power is distributed, constrained, and held accountable—distinguished from voluntary corporate policy documents.
Condorcet: 18th-century French political economist and mathematician used by Hall as a historical lens; traced the Enlightenment to the printing press and wrote his most optimistic work while in hiding during the Reign of Terror.
Cost of Being Informed: The host’s framework combining financial cost, time cost, and cognitive effort required for a citizen to become politically informed; the variable that AI could most dramatically reduce.

Summary

Andy Hall’s essay “Building Political Superintelligence” makes the case that AI represents the most significant opportunity in over two millennia of democratic experimentation to rebuild governance institutions—and that this opportunity is going to waste because virtually no one in public discourse is articulating a positive vision for it. Drawing on Condorcet’s analysis of the printing press as the engine of the Enlightenment, Hall argues that AI does for intelligence what the press did for information, and that the resulting transformation of political life could be even more profound—if it is deliberately designed. He organizes the work ahead into three layers: an information layer (making AI a reliable, unbiased tool for civic education and government efficiency), a representation layer (building faithful AI delegate agents that act on citizens’ behalf in political processes), and a governance layer (creating binding constitutional frameworks that ensure AI infrastructure answers to citizens rather than to the companies that own it). Each layer has specific, tractable research problems attached to it, and Hall’s core message is that the right response to AI’s disruptions is not to slow AI down but to accelerate the construction of the institutions and structures that keep people free as AI grows more powerful. The host of the AI Daily Brief endorses this framing, adds an analytical rebuttal to cynicism about civic apathy (arguing that lowering the cost of being informed matters more than changing people’s underlying motivation), and suggests that new agentic business models may reduce the structural conflicts of interest that have historically undermined mission-aligned technology companies.