Program

Historically, democracy (political equality) and capitalism (economic efficiency and inequality) have formed the foundation of modern society — in tension, yet mutually complementary.

But AI is now redefining that very equilibrium. The automation of knowledge work, the concentration of data and computing power, and the spread of algorithmic decision-making are raising new questions about the assumptions underlying market competition and democratic consensus-building.

At the same time, AI carries risks such as accelerating division and concentrating power — but it also holds the potential to advance public administration, expand civic participation, and create new forms of the public sphere.

This session examines, from multiple angles, how AI is transforming the structures of capitalism (markets, labor, and the distribution of wealth) and shaking the foundations of democracy (information, consensus-building, and legitimacy).

Furthermore — both democracy and capitalism rest on the culture that underpins them: what people regard as good, how they trust one another, and what they consider valuable in life. However sophisticated our institutions, if that culture erodes, both will collapse. Populism and extreme polarization are a degradation of democratic culture; treating money as the sole measure of a life is a degradation of capitalist culture. And AI reaches deep into this cultural layer too — fueling competition for attention, stoking division, and potentially compressing a diversity of values into a single metric. The question extends beyond institutional design: in the age of AI, how do we cultivate the culture that sustains democracy and a "good capitalism"?

Who will lead rule-making in the age of AI — the state, Big Tech, or civil society? And what equilibrium can "political equality" and "economic efficiency" find under this new variable called AI? We will discuss these questions.

Delivering citizens' voices to politics and government — PoliPoli and Liquitous share this goal, yet each has pursued it through a different mechanism: one through publishing and championing policy proposals, the other through dialogue and consensus-building. Drawing on hands-on work with roughly 60–70 municipalities, the two will share what that experience has revealed about citizen participation — what works, where the walls are, and the unexpected uses and insights along the way.

By bringing these different approaches together, the session invites participants to consider what it really means for a voice to "reach" someone, and what the citizen-participation infrastructure of the future will require.

Ask ChatGPT or Claude about policy or elections, and you'll get an answer that looks balanced at first glance. But is that seemingly neutral response truly neutral?

LLM bias arises in three layers: (1) the biases present in the web data used for training; (2) the values of human evaluators introduced through Reinforcement Learning from Human Feedback (RLHF); and (3) response controls imposed by corporate policy. These layers intertwine, embedding particular political and cultural positions as "common sense."

Even more serious is the problem of LLM poisoning. Research underway at Code for Japan is beginning to show that a model's political responses can be manipulated through deliberate contamination of its training data. Attack methods are diversifying — mass-generating fake news, repeatedly injecting particular narratives, and mixing malicious documents into RAG systems, among others. If citizen-participation platforms or government AI assistants are exposed to such attacks, there is a risk that the democratic decision-making process itself could be distorted.

Centered on the question "who defines the neutrality of an LLM, and who protects it," this session brings together research findings, implementation experience, and policy perspectives for discussion.

Discussions of broad listening and digital democracy tend to drift toward technical questions of "which tool, used how." But what practitioners repeatedly run into in the field is that the preconditions prior to any tool are not yet in place. This session digs into that precondition — what actually makes a "participatory society" possible in the first place — from the vantage points of local government and young people on the ground. Rather than a venue for handing down conclusions, we hope it will be a time for participants to leave with a "question" to bring back to their own work.

The discussion is organized around the three elements of the "architecture of participation" laid out in the Broad Listening book: (1) making good use of existing mechanisms; (2) attention to workflow as a process of handling and processing information; and (3) facilitation that leverages digital tools. Our position is that whether these three are in place on the ground — more than skill in the technique of "listening" itself — is what determines whether participation succeeds.

Moving between real changes seen among residents in local government settings (unexpected voices, and words that mean different things across generations), the experience of young people going out to listen to the raw voices of their peers, and the example of Barcelona’s participatory budgeting reaching those who "don’t normally have a voice," we will revisit, together with the audience, the questions "what must we put in place before we listen?" and "is this designed so that everyone can participate?"

This is a panel discussion where municipal officials, elected representatives, researchers, and civic-tech practitioners — each coming from a different position — can each leave with concrete leads to act on before introducing any technology.

Using AI to collect, summarize, and classify large volumes of opinions is already becoming technically feasible. Mechanisms for gathering tens of thousands of citizen proposals, and implementations where AI organizes the points of discussion at municipal citizen assemblies, are starting to take shape. Yet a deep gap remains between "gathering a large volume of voices" and "building consensus with legitimacy."

This session brings together researchers and practitioners at the frontier of AI-mediated communication and consensus-building technology, each presenting, at high resolution, what is "already possible" and what "remains impossible" in their respective fields. Rather than a lineup of demos, the goal is to share a cross-disciplinary research agenda.

The discussion starts from three current vantage points: (1) technology that extracts trustworthy opinions and important minority viewpoints from a large volume of input, aiming for aggregation unswayed by vocal or majority influence; (2) technology that uses LLMs to transform the format of information and dialogue, supporting understanding between people and between people and AI; and (3) the obstacles that have emerged from socially implementing citizen dialogue and consensus-building infrastructure in municipalities and political parties.

Building on this, the session takes up shared unresolved problems: the absence of metrics for good deliberation or good consensus, which stalls the feedback loop for technical improvement; who decides, and how, that a minority opinion is "important"; how far AI can go, beyond summarization and classification, in "translating" between people whose positions and values are fundamentally different; and, when AI mediates the formation of consensus, whose consensus is it — and where is the line between that and manipulation?

Attendees, technologists especially, will leave with a "map of unresolved problems in digital democracy" — where the next challenges lie, and in what order they might be tackled. Putting that into words and sharing it is the aim of this session.