A few months ago I published a position paper arguing that deliberative democracy’s real problem is not scale but consequence. Citizens’ assemblies can produce thoughtful recommendations, but rarely political cost for ignoring them. The bottleneck is not the absence of public agreement. Policies like universal background checks, medicaid expansion, campaign finance reform enjoy broad popular support. The bottleneck is that representative institutions respond to wealthy and organized minorities, rather than disorganized majorities. Responsiveness does not follow from agreement.
I proposed what I called “the algorithmic hand,” a coordination infrastructure that converts deliberative consensus into organized pressure. Unlike the invisible hand of markets or the visible hand of the state, the algorithmic hand coordinates voluntary collective action at scale. Think AI systems that identify legislative windows when pressure is most effective, contribution ledgers that make civic effort visible and enable conditional cooperation, and coalition architectures that detect latent alignments among groups who rarely collaborate. The goal is strategic leverage that preserves the educative function of organizing, not frictionless coordination. Democracy’s calculator dependency, where citizens get efficient outputs while losing the capacity to think strategically, is a real risk. If citizens become click-to-deploy amplifiers, they may win reach while losing strategy.
Today I came across a preprint by Jonas R. Kunst and colleagues that confronts me with the dark mirror of exactly this logic.
They describe “cyborg propaganda”: partisan coordination apps where AI generates personalized messages for thousands of verified human users, who post them as their own thoughts. The identity is authentic. The voice is synthetic. The coordination is invisible. Platforms already sell coordination infrastructure openly, mostly for commercial ends. Tools that let organizations distribute ready-to-post content to large networks of affiliates and advocates are now standard in marketing and in political campaigning. Combine that distribution layer with engagement analytics and AI text generation, and you get a closed-loop system: message out, reaction measured, next message optimized.
What makes Kunst and his coauthors’ framework sharp is a two-by-two matrix that clarifies where cyborg propaganda sits. Traditional botnets use synthetic identities with automated articulation. Troll farms use fake personas with human-crafted messages. Grassroots action involves real people sharing self-authored views. Cyborg propaganda keeps verified human identity, but industrializes the voice. It hybridizes the authenticity of grassroots with the scale of the botnet.
The paper then poses something I should have spent more time on: is this manipulation or empowerment? The same architecture can be read as coercion or as a collective prosthetic. In one reading, users become “cognitive proxies,” human relays whose function is to bypass platform filters designed to catch automated content. Participation shifts from authorship to ratification. In another, for citizens drowned out by paid advertising and algorithmic asymmetry, coordinated sharing could function like a union for attention. AI-generated prose could serve as an accessibility bridge, allowing people who hold policy views but lack rhetorical tools to compete with professional lobbyists. In authoritarian contexts, the architecture could resolve the classic first-mover problem of dissent: synchronized release of thousands of messages overwhelms the state’s capacity for targeted enforcement.
The convergence of both papers is uncomfortable. The same infrastructure that could help democratic publics convert consensus into leverage can, with a different governance structure, manufacture the appearance of consensus where none exists. My paper’s “verifiable contribution ledgers” and Kunst et al.’s “network harvesting” (where users are incentivized to map the political leanings of friends and neighbors) are separated by governance design, not by technology. The civil rights movement’s WATS lines, which I discuss in my paper, make the point historically. The same telephone infrastructure that enabled activists across the American South to coordinate responses to violence was also wiretapped by the FBI. The technology was identical. Who governed it determined the outcome.
Kunst and colleagues identify a regulatory problem that compounds this. Current legal frameworks rely on the distinction between human and automated speech. Cyborg propaganda sits in the loophole: you can ban botnets, but restricting verified citizens who click “post” on AI-drafted content raises immediate free speech objections. Their proposal to redefine coordination hubs as undisclosed political action committees shifts the regulatory target from individual speech to the industrial manufacturing of consensus. It targets the right layer, and it has teeth.
Both papers converge on a conclusion that neither of us finds comfortable. AI can already coordinate collective action. It does so today, mostly for commercial and partisan ends. Whether democratic movements can build, govern, and claim these systems before those threatened by democratic coordination deploy them first, and more effectively, remains an open wager. As I put it in the paper: the information highway is littered with good democratic intentions. Refusing to test the coordination hypothesis is just another way of accepting unresponsiveness as a permanent feature of democracy.