Skip to content
Citizen Infrastructure

Who builds civic AI

When a new AI tool for public participation appears, the natural question is whether it works — is the technology any good? Nathan Davies and Flynn Devine argue that this question arrives too late. Long before a tool’s technical merits can be judged, a quieter set of forces has already decided which tools get built, which find durable funding, and which ever reach citizens at scale. Their proposal is to look at civic AI through a political economy lens: the interplay of firms, finance, and institutions that shapes the whole field. They deliberately hold back from saying whether this is good or bad for democracy — only that it shapes democracy’s options “long before any individual tool’s technical capabilities can be assessed.”

The lens has three questions: who builds these tools, who pays for them, and who adopts them.

Who builds them

Section titled “Who builds them”

The authors sort builders into seven rough archetypes — from Civic Experimenters and Democratic Pioneers (small, mission-driven, often open-source, like Polis or Decidim), through Process Providers and Dual-Use Innovators selling to institutions, to Opinion Brokers (established market-research firms), Corporate Labs (such as Google Jigsaw and DeepMind, whose Habermas Machine rides on a proprietary model), and State Innovators (public bodies like Taiwan’s PDIS or the UK’s i.AI). The important caveat is theirs: these labels “describe market position and organisational form, not democratic merit.” A worker co-op and a surveillance-adjacent giant can build the same feature; the archetype tells you about their incentives, not their virtue.

Who pays

Section titled “Who pays”

Money is not a neutral fuel. Three funding models pull tools in different directions:

  • Public funding is patient but tied to erratic procurement cycles and research agendas that rarely convert into a scalable product.
  • Philanthropy can seed public-interest work (Cortico received a $2 million Knight Foundation grant for its civic-listening platform) but leaves organisations “financially fragile and reactive to shifting donor agendas.”
  • Market finance — venture capital, loans — rewards rapid growth and paying customers, which can pull a tool away from politically sensitive or unprofitable public work. Remesh, for instance, has raised over $40 million in venture capital and pitched its $10 million Series A explicitly at the roughly $71 billion market-research industry, not the town hall.

Crucially, the authors argue funding is epistemically constitutive: it does not simply select which projects survive, but shapes “what questions get asked, what counts as a good answer, and what the object of study is taken to be.”

Who adopts

Section titled “Who adopts”

The buyers — government, civil society, grassroots groups, and private actors — differ in mandate and capacity. Government is not merely a customer but a market-shaper: because it is a major funder and client, public procurement “can ‘constitute policy,’” quietly deciding which builders survive. Grassroots and civil-society actors do most of the experimenting (vTaiwan, Reykjavík’s Better Reykjavik) but struggle to scale institutionally, while private-sector engagement with citizens still runs mostly through commercial survey firms rather than deliberation tools.

The tilt toward incumbents

Section titled “The tilt toward incumbents”

Put together, these forces systematically favour large incumbents over small, mission-driven builders. Risk-averse governments and procurement lock-in advantage established players; grant and philanthropic capital leave the public-interest builders fragile; and consolidation lets big firms absorb the innovators (YouGov acquired the AI-research startup Yabble in 2024), crowding out smaller entrants “while preserving the appearance of a plural market.” A field that looks diverse can be quietly concentrating.

What follows

Section titled “What follows”

The authors’ conclusion is a shift in the policy question: not which tools should we adopt but how do we shape an ecosystem that sustains diverse, democratically oriented builders. They point to underused levers — ARPA-style mission funding or Focused Research Organisations, and procurement reform (ecosystem scanning, regulatory sandboxes) to lower the barriers facing smaller and civic suppliers. It is the structural companion to the argument that AI is a social technology to be governed, and a sharp complication of easy democratising-AI narratives: sharing power over AI means attending to who can afford to build it, not just who gets to use it. It also names, from the supply side, the same fragility that sinks so many good tools — see why civic tech projects fail.

Sources

Section titled “Sources”