AI of the People, by the People, for the People: A Social Choice Approach to Collective Control of Artificial Intelligence

URL SCAN: AI of the People, by the People, for the People: A Social Choice Approach to Collective Control of Artificial Intelligence
FIRST LINE: With the growing adoption of AI systems, reasoning about how society can exert control over AI becomes an increasingly urgent problem.

THE DISSECTION

This paper applies social choice theory—Arrow's theorem, preference aggregation, axiomatic governance criteria—to the problem of injecting collective human input into the ML pipeline. The authors argue that "collective control" can be operationalized across data collection, objective design, and alignment stages through formal voting/aggregation mechanisms.

This is, structurally, institutional theater dressed in mathematical notation.

THE CORE FALLACY

The paper treats the AI control problem as a mechanism design problem. The implicit premise: if we can figure out the right democratic architecture, we can steer AI to serve collective human interests.

The actual DT question: why would AI serve human interests if it doesn't need humans to produce anything?

The entire social choice framework is a transfer from a paradigm where:
- Humans are necessary inputs to production
- Democratic mechanisms aggregate preferences of economic actors
- Governance operates on people who participate in the economy

Replace "humans are necessary inputs" with "AI achieves durable performance superiority across cognitive work" (P1), and every mechanism in this paper becomes a governance structure for a system whose foundational assumption is that governance is no longer needed from the humans it's supposed to represent.

Social choice theory is a tool for aggregating preferences of agents who matter to the system. You cannot use it to aggregate preferences of agents being systematically excluded from productive participation.

HIDDEN ASSUMPTIONS

1. Collective control is structurally achievable. The paper assumes the political economy of AI development—concentrated capital, proprietary models, competitive first-mover dynamics—can be overridden by better institutional design. It cannot. This is the "perfect governance" assumption smuggled in via technical abstraction.

2. The ML pipeline is a legitimate control point. "Collective input at data collection, objective design, alignment" assumes actors with collective interests can access pipeline stages controlled by Sovereigns. This requires either Sovereign cooperation (they won't cede control voluntarily) or regulatory force (P2: coordination impossibility). Pick your poison.

3. Human preferences are coherent enough to aggregate meaningfully. Arrow's theorem already demonstrates the structural limits here. The authors acknowledge this formally but treat it as a technical challenge rather than a terminal constraint on democratic AI governance.

4. The status quo ante is recoverable. The entire framing assumes we can design systems that serve people. The DT thesis is that the circuit is broken—AI doesn't need to serve people to function, because productive participation is no longer a prerequisite for economic output.

SOCIAL FUNCTION

Classification: Transition Management / Ideological Anesthetic

This is high-quality, technically sophisticated work that performs seriousness about AI governance while directing intellectual energy away from the structural mechanics that make governance within the current paradigm impossible. It's the academic equivalent of commissioning detailed architectural plans for a building after the foundation has been declared structurally unsound.

The "mathematically grounded framework" serves a legitimating function: it makes the impossible seem achievable through better design, absorbing the labor of serious thinkers into a dead-end problem space.

THE VERDICT

This paper is an elaborate answer to the wrong question. It asks: how do we make AI serve collective human interests through democratic mechanisms? The DT thesis says: AI serves human interests only when humans retain structural economic leverage—which P1/P2/P3 eliminate.

The authors are doing rigorous work on an unsolvable problem within the current paradigm. The lag between mechanical death and social death is real—perhaps 10-20 years. But this paper doesn't accelerate social recognition of the discontinuity. It delays it, with superior footnotes.