Editor’s note: Noema is transparent about AI use in its pieces. We publish original, human-generated ideas but allow authorized, disclosed use of AI in certain cases. Please see details at the end of this piece.

TOKYO — In the run-up to Japan’s lower house election earlier this year, a new party called Team Mirai — or “future” in Japanese — was doing something different: actively listening. Its AI-powered chatbot walked voters through policy proposals, answering their questions and probing their thinking about what they cared about. In the end, the party fielded more than 38,000 questions and collected over 6,000 suggestions from voters that helped surface one of the central dilemmas in Japanese politics.

Voters were worried about rising prices and wanted immediate relief. But a significant minority were skeptical of the solution other parties were offering — cutting the consumption tax — partly because of potential impacts on social programs funded by the tax. The concern is reasonable given Japan’s debt burden, now over 200% of GDP and the highest in the G7.

Team Mirai chose to represent voters skeptical of the tax cuts other parties were promising. The party won 11 seats on over three million votes (almost 7% of all votes cast) — more than double what it had aimed for in its first national election.

Using AI to listen more effectively helped Team Mirai find constituencies that conventional politics were not representing. But the chatbot didn’t listen to everything. Some conversational boundaries had already been set — whether intentionally or through the constraints of the systems the chatbot relied on.

The Listening Assumption

Democratic innovators, from political scientists to civic technologists, believe that democracy’s problem is, at least in part, a listening problem: if institutions could hear what citizens really think and care about, political disagreement could become more tractable and solutions clearer. The assumption is a hopeful one. But it is also, implicitly, an exercise of power, because someone must decide what gets heard: which voices are included, how opinions are interpreted, what counts as consensus and which problems are important.

After the turn of the century, a wave of innovation aimed at improving listening began. In the 2010s, a Citizens’ Assembly in Ireland helped break political inertia on abortion, producing recommendations that were approved by around 66% of voters. The Assembly consisted of 99 citizens, randomly selected to be broadly representative of Irish society. But that choice embeds assumptions about representation: who counts as a meaningful voice, and what diversity is.

In 2019, France tried democratic listening on a national scale. President Macron launched the Great National Debate after the Yellow Vest crisis, which began over a fuel tax increase and widened into anger over living costs and the distance between governing elites and citizens. The response, as political scientist Hélène Landemore documented, ran over two months: around 10,000 town-hall meetings, almost 19,000 local grievance books — reviving the cahiers de doléances of 1789, where citizens recorded their complaints and proposals on the eve of the Revolution — 21 randomly selected citizen assemblies and 1.9 million online contributions. But scale came with its own problems. While the smaller assemblies moved closer to deliberation, the online platform mostly gathered views. Hearing more people did not mean hearing them better.

Quadratic voting — advanced by Glen Weyl and Eric Posner — tries to capture what assemblies do well but at a lower cost. It gives voters a budget of credits to express how much they care about a particular issue, rather than letting them cast a single equally weighted vote. The cost of casting more votes on the same choice rises quadratically, so that preference intensity influences the outcome. But it can’t hear people who don’t know how much they care, who don’t have the time to figure that out or who express care differently.

Polis, the platform used in Taiwan’s vTaiwan process, largely removes the selection problem: It allows all participants to submit statements and vote on others’ contributions at scale. But the choice of what counts as meaningful hasn’t disappeared — it has moved to the algorithm. Polis maps clusters of opinion and highlights statements that gain support across groups, so less-supported or more-divisive views fade into the background. Persistent disagreement isn’t excluded, but it is — by design — not the center of attention.

Each of these innovations refines the information that’s heard. But none of them changes the process through which people reason together, challenge each other and sometimes change their minds.

“Using AI to listen more effectively helped Team Mirai find constituencies that conventional politics were not representing. But the chatbot didn’t listen to everything.”

But that is what AI now promises. In 2024, a team of researchers led by Michael Henry Tessler, a research scientist at Google DeepMind, built a system to explore it. The “Habermas Machine” — named after the philosopher of communicative rationality — was tested on more than 5,700 participants deliberating on divisive political issues, including Britain’s Brexit vote, immigration and climate change. The results were impressive: participants preferred AI-generated statements that synthesized group views over human-written ones; groups converged; divisions narrowed.

But those outcomes were also the measure of success. The designers built the system to move people toward agreement, treating convergence as evidence that deliberation had worked. They settled the question of whether democracy’s goal is consensus before anyone logged on.

What Democracy Is For

The listening assumption has a deeper problem than any of these innovations can fix. The Russian-British philosopher Isaiah Berlin argued that the problem is not that we don’t listen carefully enough, but that sometimes we don’t agree because some human values are irreconcilable. Liberty conflicts with equality. Justice conflicts with mercy. Security conflicts with freedom. These conflicts are irreducible, permanent features of the moral landscape. As Berlin put it, “the possibility of conflict — and of tragedy — can never wholly be eliminated from human life, either personal or social.”

This is an uncomfortable claim because it forecloses the idea that politics can get things right in the end. It means there is no final harmony. Politics does the work of “promoting and preserving an uneasy equilibrium” between conflicting values that is continuously threatened and in need of repair, Berlin wrote. If values genuinely conflict, then democracy’s job can only be to manage those conflicts openly and legitimately.

Seen this way, democracy’s merit is that it institutionalizes losing. It allows disagreement over values to continue without requiring a final resolution. Opposition rights, minority protections, judicial review and bicameralism are features by which a community ensures that a plurality of views retain political standing. Today’s decisions can be challenged and revised tomorrow.

The Categorical Problem

Now consider what an AI system must do. It must optimize for something. That something might be consensus, as in the Habermas Machine, or preference satisfaction, as in most recommendation systems (say, Netflix or Amazon). Whatever the objective, the system moves toward it. And in doing so, it privileges certain values over others by design, whether the designers intended to target that value or not. Ultimately, to optimize is to choose.

Unlike human deliberative institutions, the losing side of an algorithm has no standing. The value choices are made in the code, by people who are not on the ballot, and they are not contestable through any democratic procedure.

The fact that an algorithm’s code can be inspected or even modified in a public repository is not the same as its value choices being contestable by those subject to them. The code still determines what enters the political process before democratic contestation can begin.

The problem with the Habermas Machine is that the decision to have it optimize for agreement is itself a political act that settles in advance what democracy exists to hold open. This is the algorithmic form of what political philosopher Michael Sandel calls the “tolerance of avoidance” — rather than engaging with moral disagreement, the system moves past it. On those terms, the machine does not fail democracy by producing bad outputs; it substitutes for democracy by resolving, before anyone even deliberates, whether agreement is the goal at all.

This is not an alignment problem that can be solved by better objective functions or more careful training. It is categorical. A system that optimizes for consensus will systematically underweight dissensus, smoothing away the edges where the work of politics actually happens. And in doing so, it embeds choices about whose views count as agreement that are, in fact, political.

The Habermas Machine’s designers were aware of this tension. Their paper cites Chantal Mouffe and Ian Shapiro — two of the most prominent theorists of agonistic democracy, who argue that the drive toward consensus is precisely what must be resisted — in a brief acknowledgment that states: “Even after deliberation, a plurality of views may persist.”

But acknowledging this is not the same as designing for that possibility. The Habermas Machine is designed to require a single, collectively endorsed statement; non-agreement is deemed a failure of the process. Proceeding anyway is a political choice that the system then enacts invisibly at scale.

“Unlike human deliberative institutions, the losing side of an algorithm has no standing. The value choices are made in the code.”

In a 2025 analysis, political science researcher Nicolás Palomo Hernández argued that the Habermas Machine “overemphasises the desirability of agreement in deliberative processes” and includes “a very narrow conception of deliberative democracy embedded in its design.” Habermas himself treated non-agreement as a possible, even legitimate, outcome of deliberation. But the machine named for him treats agreement as necessary. Habermas viewed political decisions as open to contestation; the machine sees them as the output of a completed process. Habermas placed deliberation in the exchange between participants; the machine reduces them to supplying and evaluating inputs within a process whose terms are set in advance. The system is named after a misreading of the very philosopher it invokes.

Henrik Skaug Sætra, a philosopher in the ethics of technology, warned at the end of 2024 that the Habermas Machine can become “not a deliberation tool, but a consensus machine.” His concern is not only that the system steers groups toward agreement, but that it presents the agreement as unusually neutral. If participants come to see AI-generated statements as “inherently more neutral or fair,” they may be “less likely to critically engage” with their substance and treat the system itself as a solution to political disagreement rather than a tool for working through it. The appearance of neutrality can begin to substitute for the political work of contestation.

The late Robert Dahl, a leading theorist of modern democracy, argued that people must be able to decide what issues are placed on the agenda. But in the Habermas Machine, the researchers generated the questions using an LLM and then vetted the results to minimize “the risk of provoking offensive commentary.” Participants did not control the deliberative agenda. The question of which conflicts are worth having was, again, resolved in advance.

Similar Failures

The Habermas Machine is not an isolated case. Digital technology like social media and, more recently, generative AI have been changing how disagreement appears in public life for years now, eroding the vibrancy of democracy.

Platforms have, at times, algorithmically amplified outrage, accelerated polarization, elevated misleading content and degraded the shared epistemic commons that democratic deliberation depends on. Social media didn’t just make politics angrier; it severed political conflict from accountability. Outrage generates engagement — clicks, shares, time on platform — but rarely representation or institutional response. For platforms, conflict produces revenue before it produces consequences. Conflict may go viral, yet generally nowhere democratic.

Large language models can homogenize public discourse with their neutral-sounding prose and scalable synthetic content. More troubling still is their fluent reconciliation of divergent views: these systems can smooth disagreement into synthesis before it takes a political form.

LLMs can also influence political choices. In a recent working paper on AI voting advice for the same Japanese election, political economy researchers Sho Miyazaki and Andrew Hall found that five major AI chatbots consistently recommended the Japanese Communist Party to simulated voters with left-leaning policy preferences, even though other, more moderate, parties held similar positions on the issues tested. This was because the AI models could freely access the Japanese Communist Party’s website and newspaper, which they sometimes misclassified as independent journalism, while most major Japanese news outlets have blocked AI crawlers.

Similar problems can appear even in purpose-built tools used with the best intentions.

After Team Mirai released parts of the chatbot’s source code on GitHub, the public repository showed that the system treated certain terms as “NG words” — Japanese shorthand for “no good,” or blocked terms — limiting the responses it would generate on some issues. Team Mirai developed the chatbot application, but it relied on Claude, a proprietary model whose behavior and guardrails were set by Anthropic. Whether the party knew about the specific word restrictions is unclear. But once democratic tools are built across layers of application code, platform rules and foundation models, the boundaries of political conversation may be set in more than one place, and the question of which conflicts they will engage with — and which ones they won’t — may not have a simple answer.

Where Politics Begins

What made Team Mirai’s technology democratic was that it did not try to produce agreement. The chatbot explained the party’s proposals and gathered questions and suggestions from the people who engaged with it. But Team Mirai still had to decide what to do with that feedback — which concerns to represent and which trade-offs to defend publicly. Those choices remained political, and voters could respond to them at the ballot box.

“Large language models can homogenize public discourse with their neutral-sounding prose and scalable synthetic content … smooth[ing] disagreement into synthesis before it takes a political form.”

This is the distinction between tools that feed into democratic conflict and tools that stand in for it. The process matters. Citizens’ assemblies and vTaiwan structure disagreement and refine what gets heard, but leave the final judgment to political actors. Team Mirai’s tools — for the most part — did the same. The Habermas Machine does not.

Holding It Open

None of this is an argument against using AI in democratic contexts. It is an argument about what kinds of uses are compatible with democracy — and whether they leave something for politics to do.

AI can illuminate the structure of disagreement, mapping the lines of conflict and trade-offs for politicians, courts, the legislature and citizens, without choosing a side. Berlin wrote in his last essay that “the enemy of pluralism is monism — the ancient belief that there is a single harmony of truths into which everything, if it is genuine, in the end must fit.” And those who believe they know that harmony may feel entitled to impose it, which Berlin saw as one of the oldest justifications for despotism.

AI systems tend toward monism. They have objectives. They optimize. The question is whether we build the institutional and political will to keep them in their place — subordinate to the messy and inefficient, irreplaceable human work of deciding what we value and who gets to say so.

That also requires us to accept that some conflicts should stay unresolved. The design challenge today is to build institutions and technologies that can hold disagreement open long enough for something democratic to happen. The value choices embedded in technological systems are being made regardless — whether visibly or invisibly and perhaps with the best intentions — but often before anyone thinks to ask. The democratic task is to make those choices visible and contestable before they close what politics exists to keep open.

Editor’s Note: Noema is transparent about the use of AI in its pieces. We publish original human-generated ideas but allow authorized, disclosed use of AI in certain cases. For the initial submitted draft of this piece, the author used Claude Sonnet 4.6 as a preliminary research assistant and for note-taking. Specifically, it was used to help identify additional books and sources for the author to review, summarize selected materials and work through a large body of notes accumulated over a five-month period. The author wrote the essay from this original research.

The author wrote the essay from her own reading, notes and analysis. All facts, claims and sources were reviewed by the author first. The essay’s ideas and arguments are the author’s own. Errors are the author’s own. The essay has received numerous human reviews since its first submission. Noema verified the author’s identity and the piece’s conceptual originality using various scanners and review processes and conducted a detailed human fact-check. See our AI policy here.