Democracy's Compression Problem

This blog series, co-published with the Informational Democracy Substack, is written in partnership with the Max Planck Institute for Political and Social Science.

Beth Simone Noveck opens with a provocation: grounding, focused on why democratic institutions struggle to solve public problems in the age of AI.

Two more blogs are forthcoming:

“People talking without speaking / People hearing without listening.”

— Paul Simon, The Sound of Silence (1964)

Governments need to get much better at listening. 

Thirty public servants from across Australia’s federal government gathered recently at the Australian National University in Canberra for a workshop hosted by the APS Academy and InnovateUS to learn how to do just that. The workshop drew on Designing Democratic Engagement with AI, a course I have been developing with Danielle Allen and colleagues at Harvard.

We developed the course because listening is a skill. Public servants are routinely asked to consult communities, engage stakeholders, solicit submissions and conduct co-design. Yet most have never been systematically taught how to do it: whom to involve, what to ask them to do, how to make sense of what they contribute, or how to connect what they hear to an actual decision.

In a recent Australian Resilient Democracy Network paper with Nicholas Biddle and Alex Fischer, we argued that this gap matters because artificial intelligence is rapidly lowering the cost of public engagement and public listening. AI can translate, classify, cluster and summarize large volumes of public input. Tasks that once took teams of people days or weeks can increasingly be performed in hours. The constraint is shifting from whether governments can afford to listen to whether they know how to do so well.

But there is a complication to that argument that deserves much more attention.

Even if governments genuinely want to listen, what does listening mean when there is more information than any human being could possibly absorb?

There is no shortage of talking. Residents fill out surveys, lodge complaints, call contact centers, submit comments, attend town halls and participate in consultations. Experts write reports and evaluations. Frontline workers accumulate knowledge every day about where policies and services are succeeding or failing. Administrative systems continually generate information about how government is performing.

Today’s institutions are awash in signals.

Yet much of this becomes administrative exhaust. Governments collect input they cannot process. Residents contribute ideas that disappear into bureaucratic silence. Frontline knowledge remains trapped in organizational silos. As our ARDN paper put it, engagement that does not connect to action risks becoming performative listening, producing cynicism rather than trust.

Peter Lewis and colleagues describe the same institutional failure in broader terms. Much conventional public consultation remains linear: government frames the question, solicits input and issues a report.

The more ambitious democratic project is to replace this static model with feedback loops in which public knowledge can shape and reshape decisions over time via a multiplicity of channels.

The purpose of participation is not participation itself. The democratic value lies in bringing knowledge, experience and judgment into consequential decisions. A person with a disability does not experience humane government because there was a co-design workshop somewhere upstream. They experience it when they do not have to prove for the seventh time that they have a disability. A parent does not experience responsive government because an agency has a participation strategy. They experience it when the institution knows what it already knows and can help solve the problem in front of them.

The relevant democratic question is therefore not simply: Were people heard?

It is: What became possible because they were heard?

That is the difference between voice and agency.

From scarcity to abundance 

For much of democratic history, the central informational problem was scarcity. Institutions needed mechanisms for gathering information from relatively small numbers of people and transmitting it through recognizable channels.

The Internet overturned that world.

It became vastly easier to speak, publish, comment, organize and contribute. But the capacity to produce information expanded much faster than the institutional capacity to make sense of it.

A government consultation can attract tens of thousands or millions of responses. A minister can receive information from hundreds of programs, agencies and jurisdictions. Contact centres hear recurring problems through millions of interactions. Complaints, applications, correspondence, inspections, service records and public submissions continually produce information about how policies are functioning.

The number of sensors has multiplied. The feedback loops have not.

Artificial intelligence could change this equation.

Hamburg’s digital participation system, DIPAS, illustrates the shift. Reviewing feedback from a single public engagement once occupied five employees for more than a week. The city subsequently integrated AI into its open-source participation platform so that thousands of comments can be clustered by theme, distinguished as problems or proposed solutions, linked geographically and summarized for planners.

The UK government’s Humphrey system similarly aims to reduce the considerable time and cost involved in reviewing consultation responses. AI can help institutions translate across languages, identify recurring patterns, connect information across organizational silos and make institutional memory searchable.

The ability to synthesize knowledge changes the economics of listening.

When we use AI to listen and learn, it becomes possible to derive knowledge from thirty thousand online comments, millions of service interactions or hours of in-person dialogue without requiring a person to read every sentence individually. That is an enormous collective intelligence opportunity for institutions to get smarter faster and, at the same time, an enormous democratic opportunity.

But such curation also introduces a new kind of power.

The unavoidable compression problem

No minister can read 50,000 submissions. No civil servant can review every service interaction. No citizen can read everything written about a policy question. Information abundance requires selection and synthesis. And increasingly, the synthesis becomes the thing we encounter. Pew Research found that when Google users were shown an AI-generated summary, only 8 percent clicked through to a traditional search result.

The question is therefore not whether democratic institutions will compress information.

The question is how, by whom, according to what criteria, and with what consequences.

Every act of compression is also an act of curation. Which two statements are similar enough to become one category? Which difference is significant enough to preserve? Is an unusual account a weak signal that deserves attention or an outlier that can safely disappear? Is an angry contribution noise, evidence of systemic failure, or a perspective expressed in a register the system does not understand? Which disagreements remain visible after summarization? Which minority experiences get swallowed by the dominant cluster?

These are not merely technical choices. They are also choices about representation.

Between voice and decision now sits an increasingly powerful informational intermediary: a system that may select, classify, group, rank, summarize and represent what a public knows before any policymaker encounters the underlying material.

That system helps determine not merely what government hears, but what government comes to know.

Compression is therefore an exercise of epistemic and representational power.

Political scientists have long worried about the power of agenda setting: the capacity to determine which issues become available for political decision in the first place. Robert Dahl treated control over the agenda as a fundamental democratic question. AI introduces an analogous problem one step earlier in the informational chain. Before an issue can reach the agenda, systems increasingly shape what is legible as an issue at all.

What survives the synthesis can become institutionally real. What disappears from it may never become available for judgment.

Synthesis is not settlement 

A recent Noema essay by HennyGe Wichers, “Democracy Needs Friction to Function,” warns against assuming that better listening should culminate in agreement. Drawing on Isaiah Berlin and traditions of agonistic and deliberative democratic theory, she argues that some political disagreements reflect genuine conflicts among values rather than informational deficits waiting to be resolved.

This matters enormously for AI-assisted listening.

Systems such as Pol.is can identify clusters of opinion and surface propositions that bridge otherwise divided groups. Ovadya, Lewis and other colleagues emphasize the promise of precisely these kinds of digital feedback loops, which can help institutions identify common ground and connect public input to government decision-making.

That is a genuine service to institutions drowning in disagreement.

The democratic danger arises when an information system treats disagreement itself as an error to be optimized away. 

A system that produces a clean statement of “what the public thinks” may be useful. It may also conceal the fact that there is no such singular thing.

Optimizing for consensus assumes that the principal purpose of engagement is to aggregate preferences and find sufficient agreement to move forward. But democratic institutions listen for more than preferences. They listen to acquire knowledge: about how policies work in practice, where systems are failing, what consequences officials have overlooked, and what people with lived or professional expertise know that institutions do not.

Reducing citizens to holders of opinions therefore impoverishes the informational value of participation. A resident may know why a service repeatedly fails in her neighborhood. A disability advocate may understand an implementation problem invisible in administrative data. A frontline worker may recognize a pattern before it appears in any official metric. These contributions are not merely positions to be reconciled with competing positions. They are forms of knowledge that institutions need to act intelligently.

Some differences therefore matter because they contain different knowledge. Others matter because they reflect genuinely competing interests, experiences or values. There may be competing publics, incompatible experiences, unequal distributions of costs and benefits, or moral disagreements that should remain visible because they are themselves the substance of the political question.

Plurality has both a political and an epistemic value that we do not want to elide with compression. Sometimes we preserve disagreement because democracy must accommodate conflicting values. Sometimes we preserve difference because what looks like disagreement is actually distributed knowledge that should not be averaged away.

In information science, compression inevitably loses detail. In democratic institutions, the harder question is what kind of information may safely be lost. A repeated preference, a minority experience, an anomalous observation and an irreducible value conflict are not interchangeable forms of “signal.” Democratic compression requires judgments about which differences can be collapsed and which must remain visible.

From mediation to epistemic agency 

Audrey Tang offers another useful way to see what is changing.

For much of the history of media and democracy, the central concern was mediation: who owns the newspaper, who controls the broadcast network, who moderates the platform, who carries the message.

AI systems do something more.

They search, rank, classify, summarize, advise and increasingly act. In Tang’s formulation, the issue is shifting from information mediation toward epistemic agency: who chooses the question, determines when it has been answered and decides what follows from the answer.

This distinction helps explain why debates about AI and democracy cannot stop at misinformation, bias or content moderation.

The deeper issue is institutional cognition.

• What does a government notice?

• What does it treat as a pattern?

• What does it regard as exceptional?

• When does a question become settled?

• What information reaches the person with authority to act?

• And who can reopen the question when the representation turns out to be wrong?

Tang’s broader argument about the “maximisation operating system” is useful here too. Systems optimized for attention, engagement or a single measurable objective can quietly substitute a metric for judgment. Democratic institutions require something different: the ability to keep questions open, preserve disagreement and make the objective itself contestable.

What democratic compression should require 

If compression is unavoidable, the democratic task is not to reject it. It is to govern it. That requires a richer standard than accuracy alone. A summary can be factually accurate and still democratically inadequate.

At least five qualities matter.

Fidelity. A synthesis should accurately represent the underlying contributions rather than imposing themes that are artifacts of the model, prompt or classification scheme.

Plurality. It should preserve politically meaningful disagreement, minority experience, weak signals but also diverse know how and lived experience. A democratic summary should not equate “most common” with “most important.” It should surface all unique ideas worth hearing.

Traceability. Compression should not sever the relationship between the representation and the material it represents. Decision-makers should be able to move from a theme back to the underlying comments, cases or evidence from which it was derived.

Contestability. People affected by a synthesis should be able to question it. What was included? What was excluded? Why were these contributions grouped together? What disappeared? Where appropriate, participants should be able to challenge an account of what they said.

Consequence. Listening must connect to a receiver with the authority to act. An exquisitely representative summary that disappears into an institutional void is still consultation theatre.

These are the properties of an AI listening system. That system includes the model, but also the people who frame the question, choose the participants, procure the technology, set the categories, interpret the results and decide what happens next.

This is why the idea that governments can simply purchase a “summarization solution” is inadequate. Listening infrastructure requires product management, evaluation, data stewardship, institutional ownership and professional judgment.

Lewis’ work on civic infrastructure makes a related point: not every decision requires the same form or depth of participation. Some questions call for broad consultation, others for representative deliberation, collaboration, or forms of shared decision-making. What matters is that institutions are explicit about the influence participants can actually exercise rather than promising empowerment while offering consultation.

The same principle should apply to AI-assisted synthesis and compression

We should know what the system is being asked to do.

•  Map disagreement?

• Identify recurring service failures?

• Elicit root causes of problems?

• Surface unusual cases?

• Find solutions to problems?

Each is a different task with different democratic consequences.

There is no neutral compression because there is no neutral purpose for compressing and no one way to listen and learn.

Public infrastructure for listening 

This is why the tools through which governments listen should increasingly be treated as civic infrastructure.

That does not mean governments must build their own foundation models. They may use commercial models underneath.

But applications that determine how public input is classified, clustered, weighted and summarized should be inspectable and governable. Wherever possible, governments should retain the ability to compare systems, test their outputs, modify their behavior and understand how different technical choices change what becomes visible.

South Australia is already experimenting with different approaches to AI-assisted transcription, intentionally running multiple pilots with different products, including comparing commercial tools with locally developed alternatives. The significance is not that home-grown technology is inherently superior. It is that public institutions need the capability to compare representations rather than accepting whatever a general-purpose system happens to produce.

Governments also need multiple sources of intelligence.

No consultation can tell an institution everything it needs to know. Administrative data reveals one kind of reality. Public submissions reveal another. Frontline workers see things that senior officials cannot. Representative deliberations answer different questions from open participation. Experts and people with lived experience contribute different forms of knowledge.

AI makes this plurality more manageable. But the goal should not be to funnel every form of knowledge into one authoritative summary or one consultation process.

It should be to make patterns, differences, contradictions and uncertainties more legible to people who must exercise judgment.

Critical listening as a public profession

This brings us back to the public servants in the room in Canberra.

If listening is a professional skill, then in the age of AI it must include the ability to interrogate what machines produce: what went in, whose knowledge is missing, what disagreement has been smoothed away, and what decision will change because we listened.

There is no neutral way to listen. The task is not simply to hear more people, but to exercise judgment over how what we hear becomes what institutions know.

As information becomes abundant and synthesis increasingly automated, the democratic question is who controls that compression, according to what rules, and with what possibility of challenge.

The question is not simply whether government can hear us, but whether it can listen without compressing away the pluralism it is meant to represent.