Redistributed

AI is built on public investment. DARPA funded the research that became deep learning. NSF grants supported the scientists. Public universities trained the engineers. The internet itself — the source of training data — was a government project.

The returns flow to private shareholders.

Mariana Mazzucato calls this the "entrepreneurial state" problem: the public takes the risks, the private sector takes the rewards. Her argument applies directly to AI (Mazzucato, 2013).

What would public AI look like?

The Public Investment Behind AI

The research foundation:

DARPA funded neural network research from the 1960s onward, including backpropagation
NSF supported academic ML research for decades when industry wasn't interested
NIH funded biomedical AI applications
DOE national labs developed high-performance computing

Geoffrey Hinton, Yann LeCun, and Yoshua Bengio — the "godfathers of AI" who won the 2018 Turing Award — built their careers on public funding. The breakthroughs that enabled GPT happened in universities, not corporate labs.

The infrastructure:

The internet was a DARPA project (ARPANET)
GPS was a military project, now freely available
Public universities trained most AI researchers
Government data (weather, census, maps) remains foundational

The training data:

Common Crawl, Wikipedia, academic papers, government documents — the raw material for AI was largely produced with public support or by volunteer labor, not corporate investment.

The calculation:

Mazzucato argues that when public investment seeds private profits, returns should flow back to the public — through taxation, equity stakes, or public ownership of resulting infrastructure (Mazzucato, 2021).

AI is a test case: will the technology built on public investment serve public purposes?

What Public AI Could Look Like

Option 1: Public Compute Infrastructure

The problem: Training frontier models requires massive compute — billions of dollars in GPU clusters. Only a handful of companies can afford it.

The public alternative: Government-funded compute infrastructure, available to researchers, startups, and public institutions.

Precedents:

CERN — International particle physics research, publicly funded, results open
National labs (Argonne, Oak Ridge) — High-performance computing for research
Jean Zay — France's national AI supercomputer, available to French researchers

The US version: The National AI Research Resource (NAIRR), proposed by Congress, would provide compute access to academic researchers. Pilot programs are underway, but funding remains limited.

What it would enable: Academic researchers, nonprofits, and public agencies could train models without dependence on corporate APIs. Research directions not profitable for industry could be explored.

Option 2: Public Foundation Models

The problem: Foundation models from OpenAI, Google, and Anthropic are proprietary. Even "open" models like Llama come with restrictions.

The public alternative: Publicly funded, truly open foundation models — trained transparently, governed publicly, available without restrictions.

Precedents:

BLOOM — BigScience's multilingual model, trained collaboratively with public compute
OpenLLaMA — Community effort to replicate Llama with open training
StarCoder — Code model with transparent data and training

Challenges:

Compute costs remain high (though falling)
Talent competition with well-funded labs
Governance complexity (who decides what to train?)

What it would enable: A baseline of AI capability outside corporate control. Research on safety, interpretability, and alignment without commercial constraints.

Option 3: Public AI Services

The problem: Useful AI applications require not just models but deployment infrastructure, user interfaces, and ongoing maintenance.

The public alternative: Government-run AI services for public purposes — healthcare, education, civic participation.

Examples emerging:

NHS AI Lab (UK): The UK's National Health Service has an AI lab developing applications for healthcare: diagnostic imaging, administrative efficiency, patient triage. The principle: AI developed for public health should be governed by public health institutions.

Barcelona's Decidim: A participatory democracy platform, publicly developed and open source. Not AI-native, but demonstrates the model: civic technology built by and for citizens, not sold by vendors.

Estonia's digital infrastructure: Estonia built public digital services (ID, voting, health records) as public infrastructure. The approach could extend to AI: government as provider, not just purchaser.

US examples:

AI.gov — Coordination of federal AI initiatives
Census Bureau — AI for demographic analysis
NOAA — AI for weather prediction
Limited compared to private sector, but foundations exist

Option 4: Public Alternatives to Corporate Platforms

The problem: AI is increasingly embedded in platforms (search, social media, productivity tools) controlled by a few companies.

The public alternative: Public or cooperative alternatives to dominant platforms, with AI capabilities.

Possibilities:

Public search engine (not dependent on Google/Bing)
Public social media infrastructure (interoperable, non-extractive)
Public productivity tools (document editing, communication)

Challenges: Network effects favor incumbents. Building competitive alternatives requires sustained investment. But the vision is clear: platforms as public utilities, not private monopolies.

Governance Questions

Public AI raises governance challenges that private AI can ignore:

Who decides what to build? Democratic input takes time. Technical decisions require expertise. Balancing participation with competence is hard.

How to handle dual-use? AI capabilities can serve good and bad purposes. Public institutions face political pressure that private companies can deflect.

How to ensure quality? Without profit motive, how do you ensure continuous improvement? Civil service constraints can limit agility.

How to maintain independence? Political interference in public AI could be worse than corporate control. Insulation matters.

Possible models:

Independent agencies (like central banks) — expertise-led, politically insulated
Public benefit corporations — mission-driven, with public accountability
International bodies (like CERN) — pooled resources, distributed governance
Cooperative structures — user-owned, democratically governed

Elinor Ostrom's work on commons governance is relevant here: successful commons management requires clear rules, monitoring, conflict resolution, and recognition by external authorities (Ostrom, 1990). Public AI needs institutions that embody these principles.

The Politics

Public AI faces political obstacles:

Industry opposition: Companies benefiting from the status quo lobby against public alternatives. "Government can't innovate" is a useful myth for incumbents — one refuted by the history of the internet, GPS, and the AI research pipeline itself.

Austerity constraints: Public investment requires public spending. In an era of budget pressure, new programs face scrutiny.

Ideological resistance: In the US especially, "public" is politically contested. Public AI would face attacks that corporate AI doesn't.

Coordination problems: AI doesn't respect borders. National public AI initiatives may fragment rather than coordinate.

But also opportunities:

Backlash against Big Tech creates openings for alternatives
National security concerns motivate public investment
Academic community supports public research infrastructure
International competition (China's state-directed AI) prompts responses

Kak and West (2023) document growing "AI localism" — cities and states implementing their own AI policies, procurement standards, and oversight mechanisms. This creates space for public alternatives to emerge bottom-up, not just top-down.

What Redistribution Looks Like

The status quo: Public investment → private capture → public pays (as users, workers, data sources).

With public AI: Public investment → public capability → public benefit.

Concretely:

Researchers get compute access without corporate dependency
Public agencies can deploy AI without vendor lock-in
Citizens have alternatives to extractive platforms
Workers face AI systems governed by public interest, not profit
Taxpayers see returns on their investment

This isn't about eliminating private AI. It's about ensuring alternatives exist — that the AI ecosystem isn't entirely captured by a few companies, that public purposes have public tools.

What You Can Support

Policy:

National AI Research Resource funding
Public compute infrastructure investment
Open government data and models
Procurement standards favoring open and accountable AI

Institutions:

Academic AI research (universities, national labs)
Public interest AI organizations (AI Now, DAIR Institute)
Open source AI projects (Hugging Face, EleutherAI)

Local:

Municipal AI policies and oversight
Public agency AI capacity building
Community input on AI deployment

The question isn't whether AI will be powerful. It's whether that power will be publicly accountable.

Sources

Green, Ben. The Smart Enough City. MIT Press, 2019. Publisher
Kak, Amba and Sarah Myers West. "AI Localism: Local Governments Are Taking the Lead in Algorithmic Accountability." AI Now Institute, 2023. AI Now
Mazzucato, Mariana. The Entrepreneurial State: Debunking Public vs. Private Sector Myths. Anthem Press, 2013. Publisher
Mazzucato, Mariana. Mission Economy: A Moonshot Guide to Changing Capitalism. Penguin, 2021. Publisher
Ostrom, Elinor. Governing the Commons. Cambridge University Press, 1990. Publisher
Whittaker, Meredith et al. "AI Now Report." AI Now Institute, 2018-2023. AI Now
National AI Research Resource Task Force. "Strengthening and Democratizing the U.S. Artificial Intelligence Innovation Ecosystem." 2023. NAIRR

General

Public AI