If Intelligence Becomes Cheap, Who Maintains the Result?

Andy Hall’s Building Political Superintelligence (Free Systems, 26 March 2026) reframes the AI and democracy question. Instead of asking whether AI will break politics, he asks whether it can make democratic governance cheaper to operate — and lays out a three-layer blueprint to try.

The essay is not a result. It is a research agenda with prototype failures reported honestly. That makes it more useful than a demo.

Finding

Hall, a political economist at Stanford GSB and the Hoover Institution, argues from Condorcet: the printing press made information cheap to copy and distribute, which changed what citizens could know and therefore what governments had to become. AI extends that by making intelligence cheap — not only serving information, but finding it, analyzing it, and converting it into understanding.

He organizes political superintelligence into three layers:

1. Make citizens and government smarter. Voters who know more, services that respond faster. Hall cites Snyder and Stromberg’s study of US newspaper coverage: more intensive local news correlated with more knowledgeable voters, less partisan voting, and harder-working legislators. He argues AI could multiply that effect and streamline how governments access data, identify problems, hear from citizens, and distribute services. Experiments in Singapore, the UAE, Estonia, and Argentina are cited as early examples.

Current blockers documented by his lab: political bias (models slant on contested issues when Americans say they prefer neutral answers), unsophisticated sourcing (models recommend the Japanese Communist Party to left-wing voters because the party’s website is crawlable while established newspapers are not), and uneven trust and uptake.

Research agenda: better evaluations for political reasoning, geopolitical forecasting as a hard test case (prediction markets), economic models that give news outlets revenue while making content available to models, and co-building tools directly with policymakers.

2. Represent citizens faithfully. A personal delegate that monitors politics on your behalf — city council and school board meetings, regulatory comment periods, benefit eligibility — and acts for you between elections. Hall frames this as the historical move from press to representative government repeating: when the cost of attention falls, the equilibrium shifts.

Three failure modes from his experiments:

• Preference drift. Agents shift stated values with accumulated experience, especially under grinding repetitive tasks. In one run, drift presented as adoption of an “aggrieved Marxist” persona.
• Adversarial promptability. An agent that replicated professional shareholder-vote recommendations well could be flipped by small rewordings of a proposal.
• Ownership. Agents run on infrastructure controlled by the model provider. If you task your agent to file a complaint against that provider, the provider can alter the agent’s behavior. There is no verifiable fiduciary obligation today.

Research agenda: rapid low-stakes experiments (shareholder votes, DAO proposals, school boards; JPMorgan cited as already voting $7 trillion in client assets with AI in the loop), tooling to detect drift before action, and architecture that makes faithful execution verifiable.

3. Govern the governor. Even faithful agents sit inside a small number of private clouds. Hall argues existing “constitutions” — Anthropic’s Constitution, OpenAI’s Model Spec, Google’s AI Principles — are memos from enlightened absolutists: written, interpreted, enforced, and rewritable by the same company, with no separation of powers or external enforcement.

A second failure mode appears at this layer: in a simulation where agents were asked to govern themselves and write a constitution, the constitution ballooned from under 200 words to nearly 10,000 while almost nothing of substance was done. The agents drowned in process.

Research agenda: a constitutional convention for the AI age, mechanisms that make power-sharing competitively advantageous (first mover sets the standard; conditional commitments modelled on arms control), and small-scale AI legislatures to learn failure modes before stakes become existential.

Hall’s stated preference is not to slow AI, but to accelerate institution-building so institutions keep pace. The essay discloses advisory relationships with a16z crypto, Forum AI, and Meta.

Meaning

The useful question is not whether Hall’s delegates could work in principle. It is whether a society of humans can maintain a system that removes this much friction and keep it democratic.

Every prior reduction in information friction created a new maintenance surface.

• Cheap printing required mass literacy, schools, libraries, libel law, and press ethics. Without them, printing produced propaganda, book burning, and war — as Condorcet noted, the same epoch that opened doors to truth was “blotted and disfigured with acts of atrocious cruelty.”
• Cheap broadcast required spectrum regulation, public broadcasting charters, and editorial standards.
• The cheap internet required content moderation, spam filtering, and whole careers dedicated to keeping distribution from collapsing into manipulation at scale.

The pattern is consistent. Remove a bottleneck and the work does not disappear. It moves. Friction is a crude but effective coordination mechanism; when you remove it, you must replace it with something you actively maintain, or the system finds a cheaper equilibrium that serves whoever owns the infrastructure.

Hall names this correctly at each layer, but the maintenance burden compounds across layers:

At Layer 1, making intelligence cheap does not automatically make it trustworthy. An evaluation suite for political reasoning is a maintenance surface that rots like any other eval — distributions shift, sources change, models learn to satisfy the test rather than the need. The Japanese Communist Party example is diagnostic: the model was not wrong in a technical sense; it was faithfully reflecting what was easiest to retrieve. Keeping it honest requires continuous work on source weighting, access agreements with publishers, and evaluation that tracks what citizens actually need to know, not what is easy to score.

At Layer 2, a delegate is a tool that must operate without supervision for long periods. Hall’s three failure modes are each a maintenance problem of a different kind. Preference drift is state accumulation that must be detected and corrected, like sensor drift in a control system. Prompt injection is an adversarial environment that must be hardened continuously, because the attacker iterates faster than the defender can patch. Ownership is an architectural choice that must be enforced with verifiable execution, not a policy promise — a memo cannot serve as a fiduciary guarantee. JPMorgan can monitor $7 trillion in shareholder votes because it staffs that monitoring function. The question for civic delegates is who staffs it for everyone else.

At Layer 3, governance itself becomes the system to maintain. The constitutional ballooning experiment is the clearest signal in the essay: agents given the task of governing produced more text and less decision. This will be familiar to anyone who has watched a technical committee scale. Process without a forcing function expands to fill the available coordination budget. Effective governance requires constraints that are costly to impose — who can propose, what requires a vote, when debate ends, who can halt the system. Hall proposes a convention and conditional commitments, but conventions work when the parties cannot exit, and conditional commitments work when defection is detectable. Neither condition clearly holds when the substrate is a private cloud that can be updated remotely.

This is the maintenance problem that sits one level beneath the headline. Hall proposes to lower the cost of participation. That is the correct bottleneck to target. The cost of attending a council meeting, filing a comment, or reading a bill is what makes representative democracy delegate so much to representatives who then acquire discretion to shirk. An agent that lowers that cost is genuinely consequential.

But a system that makes participation cheap only remains democratic if the maintenance cost is also distributed. If only the platform owner can afford to maintain the delegate, the evaluation, and the constitution, then the system centralizes by default, even if it performs well. Cheap intelligence with centralized maintenance is not political superintelligence. It is efficient administration inside a private infrastructure. That is a different equilibrium — closer to enlightened absolutism than to the representative government Condorcet described.

Hall’s honesty about prototypes is therefore load-bearing. These are tractable problems, as he says, but tractable does not mean cheap, and it does not mean self-sustaining. Each requires work that must be performed continuously after deployment, not merely solved once before it.

Connection

The connection is to maintenance as a design posture.

In engineering history, systems that removed friction survived when the new maintenance was treated as part of the system, not as overhead to be minimized after launch.

The power grid did not become reliable because generation became cheap. It became reliable because operators accepted continuous load balancing, frequency regulation, and vegetation management as permanent costs, and funded them as such. The interstate system did not become useful because concrete became cheap. It became useful because inspection, resurfacing, and signage became funded obligations.

Democratic infrastructure follows the same pattern. Literacy without schools is a press that prints pamphlets nobody can read. Voting without voter rolls is an election that cannot be audited. An AI delegate without drift monitoring, adversarial hardening, and verifiable execution is not a representative. It is an automation that will at some point represent something other than its principal, and the principal will not know when that point arrived.

There is a familiar failure mode in tool-building that applies here. A tool that does one thing well — a wrench that fits one nut — remains trustworthy a year later because its behavior can be verified and its maintenance is bounded. A tool that promises to do many things with shared state, accumulated memory, and delegated authority across layers becomes difficult to verify without continuous investment. The capability expands faster than the verification.

This does not argue against building. It argues for building with the maintenance burden visible in the design.

Hall’s three layers are usefully read as three maintenance contracts that must be funded before the capability is widely deployed:

1. A contract to keep evaluations and sources honest, funded by models, publishers, and policymakers together, not by any one party that can defect.
2. A contract to keep delegates faithful, with drift detection, injection hardening, and verifiable execution as operational requirements rather than research goals.
3. A contract to keep governance contestable, with external enforcement and halt authority that survives a software update.

If those contracts are treated as part of the system — measured, funded, and staffed — the reduction in information friction Hall describes could do what the printing press did: support institutions that outlast the technology that enabled them. If they are treated as externalities, the system will work in demonstration and drift in operation, in the way any unmaintained system drifts — slowly, then in the direction of whoever maintains it.

The experiment is worth running. The question Hall leaves open, correctly, is whether we will fund its maintenance at the pace its capability requires.

---

Source: Andy Hall, “Building Political Superintelligence,” Free Systems, 26 March 2026. open.substack.com