Political Representation by AI Agents: The Liberty of the Ancients at the Scale of the Moderns

Abstract

Democratic theory has long faced a tension between self-rule and scale. Rousseau held that the will of a people cannot be represented: a people that elects its representatives is free only on election day, and enslaved as soon as it is over ([1762] 2002). Constant replied that this ideal does not fit large modern societies, because their citizens have private lives to run and cannot take part in every decision themselves ([1819] 1988). Deliberative democrats have offered one way to ease this tension, pointing out that deliberation need not happen in a single assembly. Rather, it can be spread across a whole system of parliaments, media, associations and everyday conversations (Mansbridge et al. 2012). But a system can deliberate well while most citizens' own views barely reach it (Chambers 2012; Owen and Smith 2015). In most democracies, citizens' will enters politics mainly through their vote, which is a "low information, nonpropositional means of communicating" (Warren 2017, 40). Many democratic innovations have been proposed to close this gap (Smith 2009). Referendums and citizens' initiatives give every citizen a direct say (el-Wakil and McKay 2020), but rarely give voters the time or reason to deliberate before a single yes or no (Fishkin 2009, 47). Mini-publics and lottocracy give real deliberation and decision-making power to a randomly selected sample of citizens (Guerrero 2024), at the cost of others' control over who decides for them (Lafont and Urbinati 2024). Liquid democracy lets each citizen delegate her vote to someone of her choice, issue by issue, and take it back at any time (Blum and Zuber 2016). But votes can pool in a few super-delegates (Kling et al. 2015), and monitoring them takes more time than most citizens have (Elliott 2023). We propose a different innovation. So far, AI has entered this debate mainly as an aid to human deliberation, for instance as a moderator of small-group discussions (Landemore 2024; Fishkin et al. 2025). We suggest it could represent citizens, in what we call one-to-one representation. Each citizen would have her own AI representative, built from her values and positions. Her representative would debate with the representatives of all other citizens in a standing AI assembly. In the process, it may come to revise the positions it holds on her behalf in light of new information, perspectives and arguments. A second, independent AI would flag every such change to the citizen, together with the case for and against it. A change would count as hers only once she approves it. Like liquid democracy, this model is individual, continuous and revocable. Unlike liquid democracy, no citizen hands her vote to anyone else. Each AI serves one citizen only, so votes cannot concentrate in a few hands. And the second AI does the monitoring that liquid democracy leaves to citizens. In this way, citizens keep control over every position taken in their name, without having to give up their private lives. This proposal may sound like science fiction, but engineers are already building prototypes of AI representatives (Jarrett et al. 2025; Fulay et al. 2025), and their feasibility is rapidly improving. What is missing is a clear normative account of whether, and to what extent, such systems can be democratic. This paper surveys both the normative potential and the risks of this new path of reform. If the risks can be managed, one-to-one representation could offer something like the liberty of the ancients at the scale of the moderns.