Earlier, I discussed my friend in UK attending the Oxford Civic AI Conference, where one of the key themes was “civic infrastructure.” This included attempts to develop agent-based deliberation systems, in which AI agents represent citizens in public discussion and decision-making. Habermolt, as a kind of “Habermas machine,” was one of the notable experiments highlighted at the conference.
These seemingly novel—even revolutionary—models not only enhance AI accountability to humans and responsiveness to community needs, but may also fundamentally challenge long-standing assumptions about democratic governance. The real question, however, is this: beyond serving as intellectually satisfying academic explorations, can these ideas be widely implemented in the real world? Can they stand firmly in front of vested interests? Can they take root and ultimately form part of a new political imagination for the future of human society?
Polycentric AI Governance
I therefore asked ChatGPT to further examine, based on my research in my new book Demotopia: An Exploration of Democracy and Artificial Intelligence (2026), the feasibility of a bottom-up, incrementally evolving model of Polycentric AI Governance. Within minutes, it produced the lengthy article appended below, which readers are welcome to explore in full.
Here, however, I would first like to use my own language—yes, human language—to present a more intuitive interpretation of the article’s key ideas, so that readers may gain an initial impression. I also passed the article to NotebookLM and asked it to generate a concise infographic for comparison.
Ober’s Demopolis
The article begins by drawing on the ideas of two major scholars. The first is Josiah Ober of Stanford University, whose work on Demopolis takes Athenian democracy as an ideal type and reinterprets democracy as a mechanism for the aggregation and accumulation of knowledge—quite distinct from modern liberal democratic imaginaries, and strikingly aligned with multi-agent AI systems.
Romer’s chartered cities
The second is Nobel laureate Paul Romer and his concept of chartered cities. Drawing inspiration from China’s special economic zones such as Shenzhen and Zhuhai, Romer proposes creating sites of institutional innovation without requiring comprehensive political reform, offering replicable models for different regions and scalable experimentation. Notably, Romer has long envisioned chartered cities as a kind of software—capable of being copied and rapidly replicated.
With the rocket-speed advancement of generative AI, the ideas of Demopolis and chartered cities have now become practically feasible. Local communities, civil society organizations, and even loosely connected citizen networks no longer require massive capital investment or centralized technical authority. With low-barrier AI tools, they can quickly build their own platforms for deliberation and decision-making. Institutional innovations, once only exist in limited experimental settings, can now be simultaneously developed within existing urban neighborhoods, parks, schools, and even online communities.
These AI-assisted micro-institutional units retain local knowledge, values and beliefs while also enabling cross-regional exchange and learning through data and models, gradually accumulating momentum for institutional evolution. Rather than waiting for top-down, system-wide reform, different localities can experiment at their own pace. In this way, AI-assisted democracy can develop through multiple points of origin in the real world, eventually into a polycentric, incrementally evolving governance network, connecting through an interoperable ecosystem of innovation.
Swarm polity
Within such a “swarm polity,” authority is no longer monopolized by any fixed or rigid entity. Instead, it becomes situational, distributed across multiple agents, and continuously generated through interaction. Different AI agents embody knowledge and concerns from specific domains, exerting varying degrees of influence depending on the issue at hand. Decision-making thus emerges and adjusts dynamically through interaction, rather than being predetermined by fixed structures. In reality, Web3’s DAOs are already seen as early-stage prototypes of such a system.
Correspondingly, the authority and legitimacy of future democracy will rely less on formal delegation or electoral procedures, and more on procedural justice grounded in openness, transparency, and broad participation, as well as on citizens’ ability to provide meaningful and effective feedback to the system. Citizens are no longer merely “represented”; they remain continuously engaged in processes of monitoring, coordination, and adjustment, co-constituting with AI a governance ecology that is constantly evolving and improving itself.
As I write this, I am struck by the recent electoral success of Japan’s “Team Mirai,” which surged from a single seat to eleven seats in last month’s parliamentary election, becoming the country’s sixth-largest political party. Polycentric AI governance is no longer a distant fantasy. More on this next time.
