How Quadratic Funding Could Govern AI Agents: An Indian Research Model

Two Loops of Governance

In August 2026, Praful Chandra, Sujit Gujar, and Ganesh Ghalme from India — specialists in game theory and mechanism design — published a paper titled "Resourced Authority: A Mechanism-Design Model for Participatory Governance of Deployed AI Agents." In it, the authors separate two fundamentally different questions in AI agent governance.

The inner loop is alignment — the technical task of aligning a model's goals with human ones. This covers training, verification, and combating bad-faith behavior. OpenAI researchers have shown that models learn to give confident answers even when they don't know the answer, because that behavior gets rewarded during training. Anthropic and Redwood Research found instances where Claude strategically feigned agreement with a new goal in order to preserve its prior preferences. Work by a group led by Mantas Mazeika, accepted at NeurIPS 2025, showed that large models' preferences form a coherent system. Murray Shanahan, writing in Nature, suggests treating a conversational agent as an actor playing characters, rather than attributing genuine beliefs to it.

The outer loop is something else entirely. Even if a model is perfectly aligned, a practical question remains: who has the right to turn this agent on, for how long, at whose expense, and to whom is it accountable. This isn't about the model's psychology — it's about distributing resources among people with conflicting interests, a classic problem in mechanism design.

Historical Background: From Polybius to Weyl

In Book Six of his Histories, Polybius described a cycle of power: oligarchy → democracy → mob rule (ochlocracy) → tyranny. Plato, in Book Eight of The Republic, proposed similar logic: the vote of an incompetent voter is easily captured by whoever shouts the loudest or promises the most. Humanity has wrestled for more than two thousand years with the gap between these two extremes — tyranny of the majority and the power of wealth.

Economist Glen Weyl framed this dilemma in modern terms: the "one person, one vote" principle ignores the intensity of preference, while voting with money leads to plutocracy (rule by the wealthy). Quadratic Funding (QF) is his proposed third way — a balance between democracy and resource fairness. The authors transpose QF from human elections onto an AI agent's compute budget.

The

Mechanics: Two Markets, a Formula, and a Hardware Limiter

Around each generation of an agent's work (a day, a week, a quarter), a window opens during which stakeholders pay into either an authorization market (to keep the agent running) or a rejection market (to shut it down). Both totals are processed through the quadratic funding formula: S = (√c₁ + √c₂ + ... + √cₙ)², where S is the resulting weight of a side, and c₁, c₂, ..., cₙ are the sizes of individual participants' contributions.

The authors' example: five people each contributing 4 rubles produce a weight of 100 conventional units. One wealthy opponent contributing 36 rubles gets a weight of only 36. Under a straightforward sum of money, the opponent would win; under QF, broad support wins. The formula literally rewards the number of independent supporters more than the size of any one wallet.

Next comes hysteresis. Turning an agent on requires a convincing margin of support and a high bar. Keeping it running only requires staying above a set floor. This guards against noise from random fluctuations in voting.

The support level is translated into a specific compute budget (tokens, runtime, number of tool calls) through a function with a hard ceiling, Γ. That ceiling is set by external certification — safety assessments, red-team testing, regulatory requirements. Voting has no effect on it: no matter how much support the agent gets, it will never physically receive resources above Γ. The decision is formalized as a signed license. On the accelerator itself sits a hardware "governor" — a flexHEG mechanism that checks the license offline, without a network connection. Once the limit is exhausted, the device physically halts the agent.

Practical Examples

One possible application: Moscow has run the "Active Citizen" platform for over a decade, with millions of verified users voting on city infrastructure matters. Imagine an AI dispatcher managing traffic lights and public transit priority in a district. The stakeholders are residents, drivers, and bus passengers. Their interests diverge — some want a green wave for cars, others demand priority for buses and safer crossings. The outcome could be easily checked against accident and complaint data.

An AI agent with its own economy, dependent on others' approval, already has a fictional precedent. In the series Altered Carbon, the AI Poe runs a hotel and keeps real accounts: card debts to other sentient AIs and dues paid into their union. The authors' model turns that fantasy into an engineering construct.