Democratising AI

“Democratising AI” sounds unifying, but it hides at least four different goals — and companies tend to advertise the easy ones while skipping the one that actually redistributes power. A widely-cited paper by Elizabeth Seger, Aviv Ovadya, Ben Garfinkel, Divya Siddarth, and Allan Dafoe (Centre for the Governance of AI, 2023) pulls them apart:

• Democratising use — making it easier for more people to use the technology. Free tiers of ChatGPT or Claude.
• Democratising development — letting more people contribute to building AI. Public APIs, and open-weight models like Llama or Mistral.
• Democratising profits — distributing the value that accrues to whoever builds and controls advanced AI. Mostly unrealised: profit caps, “windfall clause” ideas, UBI funded by AI labs, data dividends.
• Democratising governance — distributing influence over if, how, and by whom AI is built, used, and shared. Collective constitutional AI, and the alignment assemblies run by the Collective Intelligence Project.

Why the distinction matters

Use and development are the obvious, competitive things for a company to do — and they’re real. But governance is the one that navigates the trade-offs and risks, and it’s the hardest. As OpenAI’s Teddy Lee (Collective Alignment) puts it: if you don’t get governance right, none of the rest is sticky, because anything can change later. So when someone says they’re democratising AI, the useful question is which of the four — and which they’re quietly leaving out.

The sharper provocation

Amba Kak (AI Now Institute) pushes further: “democratic AI” can itself become a fuzzy consensus that makes AI feel inevitable, smuggling in the assumption that AI belongs at the centre of every public problem. Her flip: don’t start from AI. Start from the society we want — fair wages, good jobs, education and healthcare, a sustainable planet — then ask where, and what kind of, AI plays a part. The goal is to move from a “seat at the AI table” to a table about the public interest, where it’s the AI companies that have to demonstrate why they get a seat.

Democratising development and profits, concretely

Karya is one of the few examples that touches the harder two meanings at once. A nonprofit data cooperative working with rural communities in India, it pays workers over 20 times the Indian minimum wage to build speech, text, and image datasets — and gives workers de-facto ownership of the data they create, so they earn royalties every time it’s resold. It has reached 35,000+ workers across tens of millions of paid tasks. Its founder Manu Chopra frames the wages most data workers earn as “a failure of the market” that nonprofits exist to correct.

A concrete agenda: rules, incentives, wealth, ownership

Where the four meanings classify claims, Justin Rosenstein (co-founder of One Project) turns them into a programme. His starting point is structural: a handful of companies are building civilisation-shaping systems “accountable to no one but their investors,” and no individual CEO can slow down without losing to a rival — so the fix has to be collective too. He sets out four democratic claims on AI, each meant to reinforce the others: the public should set AI’s rules (through citizens’ assemblies, not popularity-contest votes); redirect its incentives through democratic procurement and conditional licensing; share the wealth via a “public AI dividend” allocated by participatory budgeting; and hold governing ownership stakes in large AI companies. The throughline is that rules alone get captured — as radio’s “public trustee” obligations were gutted over decades — so the public needs equity and standing authority, not just conditions someone else enforces.

To carry it, One Project proposes GAIA, a Global AI Assembly, modelled on the International Atomic Energy Agency: a participatory core of citizens’ assemblies that sets goals and red lines, paired with an expert technical body accountable to it, nested by subsidiarity from global standards down to local data-centre decisions. It wouldn’t need every state — only “enough coordination among the democratic states that control the computer supply chain.” The appetite looks real and cross-partisan: the essay cites polling that 66% of Americans support citizen panels helping set AI rules.

The cooperative answer: a “solidarity stack”

A different route to the harder two meanings — development and profits — starts from labour rather than law. Trebor Scholz and Mark Esposito describe today’s AI as a vertically integrated “extraction stack” (hardware, cloud, models, labour, applications) controlled by a few firms, and its hidden workforce — content moderators in Nairobi, Manila and Hyderabad paid little to absorb a daily stream of traumatic material — as the people it “isolates and breaks.” Their alternative is a “solidarity stack”: reclaiming each layer through cooperative ownership instead of renting space on the extractive one. They point to pieces already working — MIDATA, a Swiss health-data cooperative governed by the patients whose records it holds; the Gamayyar African Tech Workers Cooperative in Kenya, started by a former Meta moderator; rural electric cooperatives as the historical proof that communities can own infrastructure — and put the count at 1.2 million workers across 53 countries. Like the social-technology reframe, they reject “the notion of artificial intelligence, which implies a magical, autonomous force,” renaming it collective intelligence to keep the human labour in view. It’s the commons logic applied to the AI stack, and it meets public AI on shared infrastructure like Apertus.

For the deeper critique of whether any of this is enough, see does AI weaken democratic institutions?; for the infrastructure angle, public AI; for the market-inspired mechanisms behind profit-sharing, plural mechanisms.

Sources

• Elizabeth Seger, Aviv Ovadya, Ben Garfinkel, Divya Siddarth & Allan Dafoe, “Democratising AI: Multiple Meanings, Goals, and Methods,” 2023: arxiv.org/abs/2303.12642 · governance.ai
• Collective Intelligence Project “Democratic AI” panel (Divya Siddarth, Saffron Huang, Manu Chopra, Teddy Lee, Amba Kak, Yoshua Bengio), 2024: youtube.com/watch?v=OQxMGMB5kL8
• “The Indian Startup Making AI Fairer — While Helping the Poor” — TIME on Karya, 2023: time.com
• Justin Rosenstein (One Project), “How to Make AI Serve the Public,” 2026 — the four democratic claims and the Global AI Assembly (GAIA): oneproject.org
• R. Trebor Scholz & Mark Esposito, “Building a Solidarity Ecosystem for AI” — Stanford Social Innovation Review, 2026: ssir.org