People Don’t Know What They Want

We Built the Tech Behind the AI-native political party. Here's What I learned.

In the summer of 2024, I helped collect thirty thousand opinions on urban policy in Tokyo. We built a platform, partnered with local government, ran the campaign. The opinions came in. Thousands of them. They looked like this:

“Lower taxes.”

“More support for single parents.”

“School lunches should be better.”

Each one a direction of feeling. Not one a position that anyone could negotiate with. We had built an extremely efficient system for gathering sentiments that were too vague to act on.

I remember sitting in a conference room afterward, scrolling through the results, and thinking: if this were a hundred thousand responses, would it matter more? If it were a million?

No. It wouldn’t.

---

Two years earlier, I had arrived at this work with a clear thesis: the bottleneck of democracy is scale. Too few voices are heard. Build systems that hear more of them, and democracy gets better. I believed this with the confidence of a twenty-year-old studying computer science and political science at Columbia who had just discovered deliberative theory and could see, with great clarity, the engineering problem underneath it.

I spent the next two years building tools in that lineage. Through a loose coalition called Digital Democracy 2030, and later through the founding effort behind Team Mirai—a new Japanese political party that would go on to win eleven seats on 3.8 million votes in the February 2026 election—I worked on systems for large-scale opinion collection, AI-assisted manifesto editing, and visualizing the shape of public sentiment across thousands of participants.

The tools worked. The most ambitious was an AI interview system we built for Team Mirai’s manifesto. The manifesto was version-controlled on GitHub, and any citizen could open a chat, pick a policy area, and talk through the party’s position one-on-one with an AI. The AI would explain the current policy, ask what the person thought was missing or wrong, and help them formulate a concrete revision. Each proposal became a pull request, publicly tracked.

In seventeen days during the 2025 House of Councillors campaign, 9,688 proposals came in. 346 were adopted. Some were substantial: public funding for HPV vaccination for men, subsidized rubella antibody testing for partners of people hoping to become pregnant, updated measures for the support cliff that children with disabilities hit at eighteen. Among independent voters (Japan’s largest voting bloc), the party became the second most popular after the ruling LDP. We proved that you could lower the cost of voicing an opinion at a scale I had not thought possible.

We also proved, at least to me, that voicing an opinion is often the wrong target.

The thirty thousand responses in Tokyo. Ten thousand proposals for a manifesto. A Senate election survey with similar numbers. At every scale, the same pattern: the inputs were too blunt to use. Not because people were uninformed or apathetic, but because translating a felt dissatisfaction into a position precise enough to compare against alternatives, expose to tradeoffs, and revise in contact with other people is genuinely hard cognitive work. Almost no democratic process makes room for it. Almost every democratic process assumes it has already happened.

Here is what I now think the real bottleneck is: people often do not yet know what they think—at least not in a form that anyone, including themselves, can negotiate with.

I don’t mean people are stupid. I mean something more ordinary and more difficult. You can care deeply about education without knowing whether you would trade cost for quality. You can feel strongly that something is unfair without being able to say which part, exactly, bothers you most, or what you would accept in exchange for fixing it. That gap—between having a feeling and having a position—is where most collective decision-making quietly fails.

The clue

The clue came from an embarrassingly small conflict inside my own company.

I run a small civic tech outfit called Plural Reality. Last year, my cofounders and I spent the better part of a week arguing about the admin dashboard for our product. The argument kept shapeshifting. One conversation it was about information density. The next it was about user onboarding. Then it was about design philosophy.

Real tension. The kind where you start avoiding the Slack channel. Where you draft messages, delete them, draft again. Everyone slightly convinced the others were missing something obvious, but unable to say what the obvious thing was.

Then I noticed something. We weren’t actually arguing about the dashboard. We were arguing about what the word good meant.

When one cofounder said “this is the right design,” he meant: this interface respects the conceptual model of the system. When another said “this is bad,” he meant: this interface makes a first-time user feel lost within three seconds. The phrase “too much information” was doing especially treacherous work. For one person, it meant every element on this screen is necessary; nothing is extraneous. For another, it meant the screen feels overwhelming. These are not different answers to the same question. They are answers to different questions wearing the same words.

We were having three separate monologues in overlapping semantic space and mistaking the overlap for communication.

So we did something slightly absurd: we used our own tool on ourselves.

The system had been built for deliberative settings—citizens’ assemblies, stakeholder consultations. It asks each participant a sequence of questions individually before group discussion begins: simple yes-or-no prompts first, then follow-ups generated based on each person’s own prior answers, gradually making their position more explicit.

Each of us spent about ten minutes answering questions on our phones. Alone, before talking to each other.

The result was startling. We all agreed that the interface should be developed iteratively. We all agreed that first impressions matter. We even agreed, somewhat to my surprise, that interface quality should ultimately be judged by user experience rather than architectural elegance.

The real disagreement turned out to be one clean axis: when clarity and efficiency conflict, which one wins?

One person would sacrifice efficiency for immediate legibility: the user should feel oriented from the first second, even if it means a slower workflow later. Another would tolerate some initial confusion if it produced a faster experience on the third use.

That was it. That was the fork. Once we could see it, the conversation unlocked in twenty minutes. Everything else—the circular Slack threads, the mounting frustration, the week of lost momentum—had been shared ground that we were pointlessly fighting inside.

Three people who had known each other for years and worked together daily could not name their own disagreement until each had spent ten minutes answering structured questions about their own preferences.

The mechanism

That experience reframed the problem for me. The hardest part of many disagreements isn’t the debate. It’s the work that should happen before debate: each participant figuring out what their own position actually is.

Deliberative theorists have long understood that good process changes minds. James Fishkin’s deliberative polling, for instance, shows that informed discussion and exposure to opposing views can shift people’s positions substantially. But even that framework assumes participants arrive with positions to shift. What I kept encountering was something more basic: people who had not yet formed the position that deliberation is supposed to refine.

Great facilitators know this intuitively—they pull people aside, ask probing questions, surface assumptions. Good therapists do it. Executive coaches do it. The Socratic method is, at bottom, this: asking a person the next question they would not have asked themselves.

You said you value simplicity, but also want all the information visible. How do you reconcile that?

You say you support universal provision, but you also care about quality. What would you accept as a tradeoff?

You say this matters to you a lot. Which part, exactly?

None of that is magic. It is structured self-reflection with a guide.

The problem is that you can’t give everyone a personal Socratic interlocutor. It doesn’t scale. You can’t station a skilled facilitator next to each of twenty residents in a town hall, or each of three hundred employees in a corporate strategy session, or each of ten thousand citizens contributing to a manifesto.

But you can build a system that does a reasonable version of it—asking questions, adapting based on answers, surfacing tensions in a person’s own stated views—and deploy it on a phone in ten minutes before a meeting starts.

That is what we built. Not an opinion aggregator. A pre-deliberation tool. Something that helps each participant cross the threshold from feeling to position before the group conversation begins.

Twenty strangers in a shrinking town

If this only worked for cofounder arguments, it would be a nice niche product. What convinced me it mattered more was watching it work in a citizens’ assembly.

Ota City is a shrinking manufacturing town about ninety minutes north of Tokyo. Like many Japanese municipalities, its population is shrinking. The city government, working with a think tank called Kouso Nippon that has facilitated randomly selected citizens’ assemblies in more than eighty municipalities, convened twenty residents to discuss childcare support.

This was not a case where the obvious answer was “spend more.” Ota already had some of the most generous childcare policies in Japan: free school lunches, free pediatric care, free diapers, non-repayable scholarships. The question was whether those resources were allocated well—which meant residents had to think about tradeoffs, not just preferences.

The meeting was held in a room on the upper floor of city hall. Three rectangular tables arranged side by side, five chairs each. The facilitator, a man named Maeda—a municipal official from Hokkaido by day, but one of the most skilled deliberation facilitators I’ve encountered—stood at the front with a projector screen behind him.

The room was mostly silent.

If you have ever been dropped into a room of strangers and asked to express a view on public policy, you know the feeling. You may have a strong emotional orientation—something feels unfair, something seems off—but converting that into a coherent position, in public, on demand, surrounded by people you’ve never met, is not natural for most people. The usual result is either confident people dominating the conversation or polite silence hardening into disengagement.

Before the meeting, each participant had answered a sequence of AI-generated questions on their phone. About fifteen prompts—simple yes-or-no questions interspersed with open-ended follow-ups, dynamically generated based on each person’s earlier responses.

How do you feel about free school lunches?

You said you support them, but also mentioned concerns about quality. If you had to choose: free lunches at current quality, or a small co-pay if it improved the food?

It sounds like your concern is less about cost than about whether free provision might have affected quality. Is that right?

The sequence does something specific. A person who walks in thinking “school lunches should be better” walks out of the pre-survey knowing something more precise: I suspect quality has declined under the current system, though I’m not certain whether that’s fact or impression. If quality did decline, I’d consider partial cost-sharing as a solution. The system didn’t supply that view. It helped the participant arrive at it.

Generated Report - “What people agree/disagree about the topic”

When Maeda projected the aggregated pre-meeting results, something shifted. People could see, before anyone had spoken, where the group already agreed. The opening silence loosened. Not because of an icebreaker, but because visible common ground made it safer to speak.

Then the conversation reached school lunches. The pre-survey had already surfaced a split: some participants believed quality had declined; others didn’t.

A woman in her fifties—a parent whose children had gone through the public school system—said, firmly: “The quality has clearly gotten worse.”

Maeda asked one question: “Is that a fact, or an impression?”

The room paused. The woman paused. Then she said, slowly: “I think it’s an impression. I’m not sure I have evidence.”

That single exchange changed the meeting. The conversation stopped being a clash of slogans and became a shared inquiry: what do we actually know? What are we assuming? What tradeoffs follow from each possibility? By the end of the session, the group had produced three concrete, tradeoff-aware proposals—not a wish list, not a pile of complaints.

The pattern was the same as in my company, just in a different register. Self-clarification first. Then mapping where the real disagreements were. Then focused conversation on the questions that actually mattered.

The dangerous question

There is a serious problem hidden inside everything I’ve described.

If AI helps people clarify their views, where does helping end and shaping begin?

We ran into this directly. In an earlier version of our system, the AI was eager to help. After a short exchange, it would say, in effect: “So what you mean is this, right? Here, I drafted a proposal for you.” Users loved it. It felt fluent and empowering.

It also felt wrong.

The same capability that lowers the cost of self-reflection can quietly substitute the system’s framing for the participant’s own. Small choices in question order, follow-up phrasing, or summary language can nudge people toward particular formulations without ever looking like overt persuasion. Which follow-up question appears first? Which tension gets highlighted? Which of five concerns gets elevated into “the main point”? These are editorial choices, and the AI makes thousands of them per session, invisibly.

I do not think this problem is solved. I am skeptical of anyone who says it is.

Our current approach has two parts. First, the system prompts and question-generation rules are designed to be inspectable—not “trust us, we’re neutral,” but “here are the actual rules; audit them.” Second, the AI can ask, probe, mirror, and pressure-test, but it does not draft your final position for you. The last step—the actual articulation—has to come from the participant, even when that introduces friction.

That friction is not a flaw. It is part of what keeps the output yours.

But I want to be honest about the limits. Transparency addresses intentional bias; it doesn’t solve emergent framing effects. Question order matters. Summary language matters. Any system that mediates through language will shape thought to some degree, because language is not a neutral vessel.

The best name I have for the failure mode is this: ventriloquism with good UX. The system produces coherence on your behalf and hands it back to you as if it were self-expression. That is not empowerment. It is something much more dangerous, dressed in empowerment’s clothes.

I don’t yet know the right boundary. It is something I am actively researching at Sony Computer Science Laboratories—when a system structures your words, drafts your thoughts, or speaks partially on your behalf, at what point does the output stop being yours? I suspect this is one of the most important unsolved questions in this space—for deliberation tools, for AI assistants more broadly, for any system that structures human communication at scale.

What I believe now

Two years ago, I would have said something like: AI can fix democracy.

I don’t believe that anymore, at least not in that form.

What I believe now is narrower and, I think, stronger. The highest-leverage intervention in many collective decisions is helping people understand their own positions before group discussion begins. That intervention is upstream of debate quality. Upstream of facilitation technique. Upstream of information access. In many cases, upstream of institutional design itself. Skip it, and the downstream failures will present themselves as polarization, stubbornness, or noise—when the more basic problem is that the participants were never ready to negotiate in the first place.

I also no longer think the right image for “scaling democracy” is one conversation that gets bigger—more voices, more data, more inputs flowing into a single process. I think the more promising model is making the sequence I described—self-clarification, issue mapping, focused disagreement, decision—available in many places at once. Municipal assemblies, corporate leadership teams, product reviews, community conflicts. Not one master facilitator running one miraculous process, but tools that let many ordinary facilitators run processes that are substantially better than today’s default.

But I want to say something beyond the practical point.

The thing that struck me most over the past two years was not any single result. It was the sameness. Three cofounders fighting about a dashboard. Twenty strangers in a shrinking city discussing childcare. Thirty thousand respondents giving directions of feeling about urban policy. The scale varied by four orders of magnitude. The failure was identical every time: people who had not yet formed positions precise enough to negotiate with, thrown into processes that assumed they had.

Once you see that pattern, you start seeing it everywhere—because it is everywhere. Wherever multiple agents with different interests share finite resources, there is a negotiation to be had. A family dividing responsibilities. A team allocating headcount. A city deciding between a park and housing. A coalition of nations dividing emission cuts. The situations differ enormously; the underlying structure does not. Someone wants something. Someone else wants something incompatible. The resources do not stretch to cover both. A decision must be made, somehow, that enough parties can live with.

This is the problem of politics in its most general form. Not partisan politics—the deeper thing. The fact that coexistence requires ongoing negotiation among beings who do not want the same things.

It is tempting to believe that this problem will eventually dissolve. That with enough information, enough intelligence, enough optimization, the right answer will become obvious and agreement will follow. I think this is a fundamental misunderstanding. Better information can resolve factual disagreements—whether school lunch quality has declined, whether a policy has the effect its proponents claim. But it cannot resolve the value conflicts underneath: how much efficiency are you willing to trade for equality? How much freedom for safety? How much present comfort for future resilience? These are not problems with solutions. They are tensions that must be negotiated, again and again, by the people who live inside them. No amount of artificial intelligence changes this. If anything, as AI concentrates the power to make decisions at unprecedented speed and scale, the question of whose values guide those decisions becomes more urgent, not less.

That is why I think this work—building tools for the moment before negotiation, the moment where a person crosses from feeling to position—is not a niche within civic tech. It is infrastructure for what may be the last problem standing. Logistics can be optimized. Information can be indexed. Prediction can be automated. But the act of a group of people with conflicting interests arriving at a decision they can all live with—that cannot be optimized away. It can be supported, structured, made more intelligent. But the outcome has to be owned by the people inside it, or it is not agreement at all.

The twenty residents in that room in Ota City went from silence to concrete, tradeoff-aware proposals in a single session. Not because the AI was brilliant. Because it helped each of them cross a threshold that most collective processes ignore entirely: the threshold between having a feeling and having a position.

That threshold may be smaller than democracy. It may also be upstream of most of it.

---

Shutaro Aoyama studies Computer Science and Political Science at Columbia University. He is the founder of Plural Reality, which builds AI-mediated deliberation tools, and a researcher at Sony Computer Science Laboratories, where he works on human-AI interaction and the agency problem in AI-mediated communication. He previously helped build the deliberation technology behind Team Mirai, a political party that won eleven seats in Japan’s 2026 general election. He can be reached at shutaro.aoyama@gmail.com or @blu3mo on 𝕏.