Skip to content
Citizen Infrastructure

Public AI

Public AI is the idea that artificial intelligence can and should be a form of public infrastructure — closer to libraries, parks, electricity, water, or the internet than to a private product. It’s less an organisation than a movement: a growing group of researchers, national labs, and nonprofits who think a powerful general-purpose technology shouldn’t be controlled entirely by a few profit-maximising companies.

Not corporate, not just open-source

Section titled “Not corporate, not just open-source”

The security technologist Bruce Schneier draws the sharpest line. If we want to use AI for democracy, he argues, we need trustworthy AI, and the market won’t supply it on its own. What’s needed isn’t a corporate model the public is free to use, nor a corporate model licensed by government, nor even an open-source model: it’s a public model built by the public, for the public — with political accountability, not just market accountability, and free for anyone to build on. The point of government, on this view, is to create the social trust a society runs on; making AI trustworthy is part of that job.

This connects public AI to the wider argument that knowledge is a commons: today’s large models are trained on humanity’s shared knowledge, so the value they create arguably belongs to the whole of society rather than to whoever enclosed it.

What it looks like in practice

Section titled “What it looks like in practice”
  • The Public AI Inference Utility — a nonprofit platform that hosts public AI models behind a free chat front-end and a developer API, backed by donated compute from partners worldwide. The aim is an option that is publicly accessible, publicly accountable, and permanent — “a highway that can’t disappear tomorrow.” It’s fiscally sponsored by Metagov and funded by Mozilla, the Future of Life Institute, and the Center for Cultural Innovation.
  • Apertus, built by the Swiss AI Initiative (EPFL, ETH Zurich, and the national supercomputing centre CSCS), is the flagship public model: fully open (weights, data, and training recipe), released under a permissive licence, at 8-billion and 70-billion-parameter sizes. It’s deliberately multilingual, with effort spent on low-resource languages like Swiss German and Romansh, and trained only on permissively-licensed data — narrowing what its builders call the “compliance gap,” the supposed quality cost of not training on copyrighted material.

The bet, and the objection

Section titled “The bet, and the objection”

As models get small enough to run on a laptop or phone, why build a shared utility at all? The honest answer is that most people won’t self-host, just as most people don’t run a web server from their phone even though they could. So a genuinely public, non-terrible option has to exist — and stay public — or the default falls back to corporate control. Public AI is meant as a counterbalance: infrastructure whose direction is set by democratic governance, not only by an expansionist market logic.

“Public” can also mean ownership rather than infrastructure: holding governing equity in the big private labs through democratic institutions, the route One Project’s Global AI Assembly proposal takes. The two senses converge — a public option you can use, and a public stake in the systems you can’t avoid.

For the case that this counterbalance is urgent — that the current design of AI actively weakens democratic institutions — see the critique. For the different things “democratising AI” can mean, see democratising AI.

Sources

Section titled “Sources”