Redistributed

AI governance is currently shaped by three actors: companies, governments, and technical experts. Missing from this list: the public.

This is not an oversight. It reflects a structural assumption that runs through most technology policy: that complex systems require expert governance, and democratic participation is a luxury that slows things down.

The assumption is wrong. Participatory models for technology governance exist. They have been tested. They work. They just are not being scaled — because scaling them would redistribute power from those who currently hold it.

Who Governs AI Now

Consider the actors who shape how AI systems are built, deployed, and regulated.

Companies make the foundational decisions: what to train, on what data, with what safety measures, for what purposes. OpenAI's decision to release ChatGPT publicly, Google's choice to rush Bard to market, Meta's decision to release Llama as open-weight — these were not democratic choices. They were corporate decisions, made by small leadership teams, that affected billions of people.

Governments set regulatory frameworks — the EU AI Act, the White House Executive Order, China's algorithm regulations. But government processes are slow, and the officials involved are often dependent on industry for technical information. Governments regulate AI; they do not govern it in any meaningful sense of directing its development.

Technical experts — AI researchers, safety teams, ethics boards — provide guidance that shapes both corporate and government decisions. But expertise is concentrated in a handful of institutions, mostly in wealthy countries, mostly funded by or employed by the companies being advised.

What all three have in common: they are not accountable to the people most affected by AI systems. A welfare recipient whose benefits are denied by an algorithm has no seat at the table where that algorithm was designed. A worker displaced by automation was not consulted about the productivity gains that made her job redundant. A community subject to predictive policing did not vote on whether to deploy the system.

As Langdon Winner established decades ago, artifacts have politics — technologies embed choices about who benefits and who bears costs (Winner, 1980). The question is not whether AI will be governed politically. It is whether that governance will be democratic.

The Democracy Gap

The gap between who decides and who is affected is not unique to AI. But AI widens it.

Technical complexity serves as a barrier to participation. When policymakers do not understand transformer architectures, they defer to those who do. When the public cannot evaluate claims about model capabilities, they accept expert assertions. Complexity becomes an argument for exclusion — even though the most important questions about AI (who benefits? who bears risk? what is it for?) are not technical questions at all.

Ursula Franklin diagnosed this dynamic in her Massey Lectures: prescriptive technologies separate planning from execution, concentrating decision-making authority in the hands of those who design the system rather than those who use it or are affected by it (Franklin, 1989). AI is the prescriptive technology par excellence — designed by a small technical elite, deployed upon everyone else.

Speed is another exclusionary factor. AI capabilities advance faster than democratic processes can respond. By the time a citizens' assembly deliberates, the technology has moved on. This is treated as an argument against participation. But speed is itself a choice — companies choose to race rather than wait, to deploy first and address consequences later. A different choice is possible.

Corporate concentration means that the relevant decisions are made in a handful of boardrooms. OpenAI, Anthropic, Google DeepMind, Meta — you can count the frontier AI labs on one hand. This concentration is not a natural feature of the technology; it reflects compute economics, talent markets, and deliberate strategies of consolidation. But its effect is to reduce the number of decision points, making democratic input harder to route.

Global deployment creates jurisdictional complexity. A model developed in San Francisco, trained on servers in Nevada, deployed through APIs hosted in Ireland, accessed by users in Jakarta — where does democratic governance happen? National democratic institutions struggle to assert authority over systems that operate transnationally.

The result is what political theorists call a "democracy deficit" — consequential decisions made outside democratic accountability. Virginia Eubanks documented how automated systems create a "digital poorhouse" that affects the most vulnerable without their input or consent (Eubanks, 2018). The deficit is not accidental. It is a feature of how AI governance has been constructed.

What Democratic Governance Would Require

Democratic governance of AI is not impossible. But it requires taking seriously what democracy actually means.

Affected communities must have standing. The people subject to AI systems — not just as consumers but as workers, welfare recipients, criminal defendants, job applicants — must have formal roles in governance. This is not consultation, where their views are solicited and then ignored. It is participation, where their input has binding force.

Elinor Ostrom's work on commons governance provides a framework. Her research showed that communities can successfully manage shared resources when they have clear rules, monitoring mechanisms, conflict resolution processes, and recognition by external authorities (Ostrom, 1990). AI governance could be structured the same way — if we chose to structure it that way.

Expertise must be democratized. Technical knowledge cannot remain the exclusive property of industry. This means public investment in AI research outside corporate labs. It means educational programs that build AI literacy. It means creating institutions where non-industry experts can develop and share evaluative capacity.

The alternative — where regulators depend on industry for information about the systems they regulate — is not neutral. It is a structural advantage for industry, allowing it to define the terms of the debate.

Processes must match the pace of technology. Democratic participation takes time. But "time" is not fixed. Processes can be designed to be faster: standing panels rather than ad hoc consultations, continuous engagement rather than periodic review, delegated authority within democratic constraints.

The claim that democracy is too slow for AI is a self-fulfilling prophecy. We have not built fast democratic institutions because we have not tried. The will to do so is lacking precisely because the current arrangement suits those in power.

Scale must be addressed. AI affects billions of people across national boundaries. No single democratic process can encompass all affected communities. But this is an argument for multi-level governance, not for abandoning democracy.

Ostrom's concept of "polycentric governance" — multiple overlapping institutions at different scales, with coordination mechanisms between them — offers a model. Local, national, regional, and international institutions each with appropriate authority, connected through mutual recognition and shared principles.

Participatory Precedents

The claim that democratic governance of complex technology is impractical ignores the precedents that already exist.

Citizens' Assemblies

A citizens' assembly gathers a randomly selected, demographically representative group of people to deliberate on a specific issue. Participants receive information from experts, discuss among themselves, and produce recommendations.

Ireland's Constitutional Convention (2012–2014) used this model to address constitutional questions including same-sex marriage. Citizens received expert briefings, deliberated over multiple sessions, and produced recommendations that went to national referendum. The process worked: complex, contested issues were resolved through structured public deliberation.

France's Citizens' Convention on Climate (2019–2020) brought 150 randomly selected citizens together to develop climate policy proposals. Their recommendations were ambitious and specific. The government's failure to implement most of them is instructive — it shows that the obstacle is not public capacity to deliberate but political will to act on deliberation.

The UK's Climate Assembly (2020) similarly gathered citizens to deliberate on net-zero policy. Participants engaged with technical material, considered trade-offs, and produced nuanced recommendations.

These assemblies demonstrate that ordinary citizens can deliberate effectively on complex technical and policy questions — when given adequate information, time, and structure. The claim that AI is "too complex" for democratic input fails against this evidence.

Deliberative Polling

Developed by James Fishkin at Stanford, deliberative polling combines survey research with structured deliberation. A representative sample is polled, then brought together for informed discussion with experts and each other, then polled again.

The method consistently shows that opinions change when people deliberate — and change in the direction of more considered, less partisan, more nuanced positions. "Informed public opinion" is not an oxymoron; it is what emerges when the conditions for informed deliberation exist.

Deliberative polls have been conducted on issues from energy policy to European integration. They could be applied to AI governance — if there were institutional will to do so.

Municipal and Regional Governance

Some of the most promising democratic AI governance is happening at local and regional levels.

Amsterdam and Helsinki's Algorithm Registers require city governments to publish information about the algorithmic systems they use. Citizens can see what systems affect them and how. This is transparency as democratic infrastructure — a precondition for meaningful participation.

Barcelona's Decidim platform is an open-source participatory democracy tool used by the city government. Citizens propose, debate, and vote on city policies. The platform is publicly owned and developed, demonstrating that civic technology can be built as public infrastructure rather than purchased from vendors.

Kak and West's work on "AI localism" documents how cities and states are implementing their own AI policies, procurement standards, and oversight mechanisms (Kak & West, 2023). This creates space for democratic experimentation at scales where participation is more tractable — and builds institutional capacity that can inform larger-scale governance.

New York City's Local Law 144 requires bias audits of automated employment decision tools. It is imperfect — industry lobbying weakened the law, and enforcement is challenging — but it represents democratic governance asserting authority over AI deployment.

Experiments in Participatory AI

Some attempts have been made to incorporate public input into AI development itself.

Collective Constitutional AI

Anthropic's "Collective Constitutional AI" experiment (2023) used a deliberative polling methodology developed by the Collective Intelligence Project to gather public input on AI system values. This experiment, whatever its limitations, represents one of the few cases where an AI company subjected model development to genuine public deliberation — a precedent worth building on. A representative sample of Americans deliberated on what principles should guide AI assistant behavior, and the resulting "constitution" was used in model training.

The experiment was limited. Participants could influence how the model was trained to respond, not what the model was built for or whether it should exist. The scope of democratic input was constrained to downstream questions while upstream decisions remained corporate.

But the experiment demonstrated technical feasibility. Public input can be incorporated into model development. The question is whether companies will expand such processes — and whether such processes would have binding force rather than advisory status.

Participatory AI Design

Academic researchers have explored participatory approaches to AI system design. Suresh et al. and others have developed frameworks for involving affected communities in the design of AI systems that will affect them — not as test subjects but as co-designers (Suresh et al., 2022).

These approaches draw on participatory design traditions from Scandinavian workplace technology and community-based research. They treat affected communities as experts in their own circumstances — experts whose knowledge is essential to building systems that work for them.

The challenge is scaling participatory design beyond individual projects. A research prototype is not governance. Moving from "we involved the community in this project" to "communities are systematically involved in AI governance" requires institutional infrastructure that does not yet exist.

Community-Based AI Governance

Some communities have developed their own frameworks for AI governance outside official channels.

Indigenous data sovereignty movements assert tribal authority over data about indigenous peoples and communities. The CARE Principles for Indigenous Data Governance provide a framework that emphasizes Collective benefit, Authority to control, Responsibility, and Ethics — centering community self-determination rather than external governance.

These movements demonstrate that democratic AI governance need not flow from state institutions. Communities can assert authority over technologies that affect them, develop their own principles, and demand that external actors respect those principles.

Challenges

Democratic AI governance faces real obstacles. Acknowledging them honestly is necessary to address them.

The Expertise Problem

AI systems are technically complex. Citizens cannot be expected to understand transformer architectures or RLHF training procedures. Does this mean democratic governance is impossible?

The argument proves too much. Citizens do not understand the technical details of pharmaceutical development, nuclear reactor design, or financial derivatives — yet we govern these through democratic institutions. The solution is not to exclude the public but to create institutions that translate between technical and democratic domains.

Expert bodies advise democratic institutions. Technical decisions are delegated within frameworks set by democratic processes. Citizens do not need to understand every detail; they need to understand the stakes, the trade-offs, and the values at issue.

The harder challenge is information asymmetry. AI companies control access to their systems. Even expert regulators struggle to evaluate proprietary models. Democratic governance requires transparency that companies resist.

This is not an argument against democracy. It is an argument for mandated transparency, public research capacity, and regulatory authority to compel disclosure. The information problem is structural, and structural solutions exist.

The Scale Problem

AI affects billions of people across national borders. No democratic process can include everyone affected.

But scale is a challenge for all global governance, not specifically for democratic governance. The question is whether international institutions can be made more democratic — incorporating public input across borders — not whether democracy should be abandoned at scale.

Precedents exist: international environmental governance has developed mechanisms for NGO participation and public comment. The challenge is extending these mechanisms and making them more binding.

Multi-level governance — local, national, regional, global — can address scale by allocating different questions to different levels. Not every decision needs global democratic input; many questions can be resolved at smaller scales, with coordination mechanisms between them.

The "Who Is the Public?" Problem

"The public" is not a unified actor. Different communities have different interests. Workers displaced by automation have different stakes than consumers who benefit from cheaper goods. Users in wealthy countries have different concerns than those in the Global South.

This is not a bug; it is democracy. Democratic processes exist precisely to negotiate among competing interests. The alternative — letting corporations or experts decide unilaterally — does not eliminate competing interests; it simply suppresses them.

The design challenge is ensuring that marginalized voices are not drowned out by powerful ones. This requires intentional structures: reserved seats for affected communities, resources for participation, protection against retaliation.

Ruha Benjamin's work on "race after technology" shows how supposedly neutral technologies encode and amplify existing inequalities (Benjamin, 2019). Democratic AI governance must be designed with these dynamics in mind — not naive about power, but structured to counteract it.

The Speed Problem

Technology moves fast. Democratic deliberation moves slowly.

But "slowly" compared to what? Corporate decisions are also slow — product development takes years. The difference is that corporate timelines are treated as fixed constraints while democratic timelines are treated as infinitely compressible.

Democratic processes can be designed for speed: standing panels with delegated authority, continuous engagement rather than periodic review, fast-track procedures for urgent decisions. The question is whether speed is genuinely necessary or whether it serves as an excuse for excluding publics.

Moreover, "fast" deployment often means "reckless" deployment — releasing systems without adequate testing, addressing harms after they occur rather than before. Slowing down for democratic input might prevent harms that fast deployment causes.

Institutional Designs

If democratic AI governance is possible, what institutions would it require?

Public AI Councils

Representative bodies, composed of randomly selected citizens and affected community representatives, with authority over AI policy in their jurisdictions. Not advisory boards whose recommendations can be ignored — bodies with actual power.

Functions could include: setting priorities for public AI research, reviewing AI deployments in public services, establishing standards for private AI systems, hearing complaints from affected communities.

The model is closer to elected school boards or zoning commissions than to ethics advisory committees. The authority must be real or the participation is performative.

Participatory Standards Bodies

Technical standard-setting currently happens in industry-dominated venues. Democratic governance would require opening these processes.

This means funding civil society participation in standards bodies. It means requiring that standards affecting the public undergo public comment periods with responses to comments. It means creating alternative standards processes that are not controlled by industry.

The Internet Engineering Task Force's open model — rough consensus and running code, with participation open to anyone — provides one precedent. It is imperfect but demonstrates that technical standards can be developed more openly than current AI governance allows.

Community Technology Councils

At the local level, community bodies could review AI deployments affecting their areas. Before a police department deploys predictive policing, before a school district adopts AI tutoring systems, before a welfare agency implements automated eligibility determination — community review with authority to reject or modify.

This is the logic of environmental impact review applied to AI. We do not let developers build wherever they want without community input. Why should AI deployment be different?

International Coordination Mechanisms

Global AI governance requires international institutions. Current proposals (UN AI advisory bodies, OECD guidelines) are expert-driven and advisory. Democratic alternatives would incorporate public input across borders.

Citizens' assemblies could be convened internationally — randomly selected participants from multiple countries deliberating together. The Global Assembly, convened during COP26, provides a model. Such assemblies could inform international AI governance rather than leaving it to diplomats and industry representatives.

The Redistribution Stakes

Governance is not separate from redistribution. It is how redistribution happens.

Who decides how AI systems are built determines who benefits from them. If corporate executives decide, value flows to shareholders. If technical experts decide, value flows to those who can deploy expertise. If affected communities decide, value might flow more broadly.

The current governance structure — corporate decision-making, expert advice, weak regulation — has produced a particular distribution of AI's benefits: concentrated among technology shareholders, high-skilled workers, and wealthy consumers; costs dispersed among displaced workers, surveilled communities, and exploited data sources.

A different governance structure could produce a different distribution. But that structure will not emerge automatically. It requires political struggle — against the interests that benefit from the current arrangement.

The companies building AI systems have not invited democratic oversight. Governments have not prioritized building democratic AI institutions. The technical community has largely accepted its privileged advisory role. Change requires pressure from outside — from communities demanding a seat at the table, from movements insisting that those affected by technology should govern it.

What Would Have to Change

Democratic AI governance is not a technical problem awaiting a clever solution. It is a political problem requiring political action.

Resources must be reallocated. Building democratic AI institutions costs money. That money could come from taxing AI companies, redirecting public research funding, or international coordination. The will to allocate resources signals the will to make governance democratic.

Authority must be transferred. Corporate decision-making authority must be constrained; public institutions must gain jurisdiction they currently lack. This requires legislation, regulation, and potentially constitutional change.

Culture must shift. The assumption that technology governance belongs to experts must give way to the recognition that affected communities have standing. This is not anti-expertise; it is anti-monopoly on legitimacy.

International cooperation must deepen. AI governance cannot be solved nationally. But international institutions will not become democratic unless nations commit to making them so.

None of this is easy. All of it is possible. The precedents exist. The knowledge exists. What is lacking is political will — and political will can be built.

What You Can Do

Locally:

Advocate for municipal AI oversight bodies
Demand community review before public agencies deploy AI systems
Support local participatory budgeting and extend it to technology decisions

Nationally:

Support legislation requiring public participation in AI policy
Advocate for funding civil society engagement in standards processes
Push for transparency requirements that enable democratic oversight

In your community:

Build organizations that can participate in governance processes
Develop AI literacy programs that prepare people for participation
Document how AI systems affect your community and make that knowledge public

The question is not whether AI will be powerful. It is whether that power will be democratically accountable. The answer depends on whether publics organize to demand it.

Sources

Benjamin, Ruha. Race After Technology: Abolitionist Tools for the New Jim Code. Polity, 2019. Publisher
Eubanks, Virginia. Automating Inequality: How High-Tech Tools Profile, Police, and Punish the Poor. St. Martin's Press, 2018. Publisher
Fishkin, James. Democracy When the People Are Thinking: Revitalizing Our Politics Through Public Deliberation. Oxford University Press, 2018. Publisher
Franklin, Ursula. The Real World of Technology. House of Anansi Press, 1989. Publisher
Kak, Amba and Sarah Myers West. "AI Localism: Local Governments Are Taking the Lead in Algorithmic Accountability." AI Now Institute, 2023. AI Now
Ostrom, Elinor. Governing the Commons: The Evolution of Institutions for Collective Action. Cambridge University Press, 1990. Publisher
Suresh, Harini et al. "Towards Intersectional Feminist and Participatory ML: A Case Study in Supporting Wage Theft Advocacy." ACM Conference on Fairness, Accountability, and Transparency, 2022. ACM
Winner, Langdon. "Do Artifacts Have Politics?" Daedalus 109, no. 1 (1980): 121-136. JSTOR
Anthropic. "Collective Constitutional AI: Aligning a Language Model with Public Input." 2023. Anthropic
Collective Intelligence Project. "Alignment Assemblies." 2023. CIP
Decidim. "Free Open-Source Participatory Democracy." Decidim
City of Amsterdam. "Algorithm Register." Amsterdam
CARE Principles for Indigenous Data Governance. "CARE Principles." GIDA

General

Democratic AI Governance