Should AI Be a Public Good? Arguments For and Against

1/ GPS started as a military project. The internet was publicly funded. The Human Genome Project was a public investment. Should frontier AI be a public good too? The arguments are more nuanced than you'd think. 🧵

2/ WHAT IS A PUBLIC GOOD? In economics: non-rival (my use doesn't reduce yours) and non-excludable (hard to prevent access). AI models are partially rival (compute is finite) and easily excludable (API keys). AI isn't a pure public good. But neither was clean water.

3/ THE CASE FOR: Public goods argument. AI is becoming essential infrastructure. When something is essential, markets alone produce underinvestment (not enough people get access) and misallocation (wrong problems get solved). Markets optimize for profit, not need.

4/ HISTORICAL PRECEDENT: Public investment created the internet, GPS, weather satellites, and most foundational research behind AI itself. Government-funded research at Stanford, MIT, and DARPA built the field. Private companies captured the returns.

5/ THE EQUITY CASE: Private AI optimizes for paying customers. Medical AI targets wealthy-market diseases. Language models work best in English. Public AI could prioritize neglected diseases, smallholder farming, low-resource languages.

6/ Elinor Ostrom proved commons can be governed without privatization or state control. Community-managed fisheries, forests, and irrigation systems thrive worldwide. Could AI be governed as a commons — with shared rules and distributed stewardship?

7/ THE CASE AGAINST: Innovation argument. Private competition drives rapid improvement. The AI race between OpenAI, Google, and Anthropic produced GPT-4, Gemini, and Claude in three years. Would a public project have moved as fast? Probably not.

8/ GOVERNANCE RISK: Who controls a public AI? Governments get things wrong too. A state-controlled AI could entrench power as easily as a corporate one. China's AI governance shows how "public" can mean "state-directed" rather than "citizen-owned."

9/ COST REALITY: Training frontier models costs $500M-$1B. Operating them costs billions more. This isn't like funding a library. It's like funding a space program — perpetually. Taxpayers would need to accept staggering costs with uncertain returns.

10/ THE MIDDLE PATH: Not purely public or private. Options include: public compute infrastructure (EU's plans for sovereign AI), mandatory licensing (like pharmaceutical compulsory licensing), regulated utilities (like electricity), or data trusts (community-governed data).

11/ FRANCE'S MODEL: Mistral received significant government support and strategic backing. The EU is funding public compute through EuroHPC. These aren't full public goods, but they establish that AI sovereignty requires public investment.

12/ The binary "public vs. private" obscures the real question: What governance structures ensure AI serves broad public interest, not just shareholders? That might look like public investment, private innovation, and democratic oversight — together. /end

---

LinkedIn version:

GPS was military. The internet was publicly funded. The Human Genome Project was a government investment. Should frontier AI be a public good? The arguments are more nuanced than partisans on either side suggest.

The case for: AI is becoming essential infrastructure, and markets alone produce underinvestment (too few people get access) and misallocation (wrong problems get solved). Private AI optimizes for paying customers — medical AI targets wealthy-market diseases, language models work best in English, agricultural AI serves large farms. Public investment could prioritize neglected needs. And let's not forget: government-funded research at Stanford, MIT, and DARPA built the field that private companies now monetize.

The case against: Private competition drives remarkable speed. The AI race produced GPT-4, Gemini, and Claude in three years. A public project likely wouldn't match that pace. Governance is a real concern — state-controlled AI could entrench power as easily as corporate AI. And the costs are staggering: $500M-$1B per frontier model, billions more to operate.

Elinor Ostrom showed that commons can be governed without privatization or state control. The middle path is promising: public compute infrastructure (as the EU is building), mandatory licensing frameworks, regulated utility models, and community-governed data trusts.

The binary of "public vs. private" obscures the real question: What governance structures ensure AI serves broad public interest, not just shareholders? That likely means public investment, private innovation, and democratic oversight working together — not choosing one over the others.