Liz Highleyman

I'm increasingly despondent about how binary the AI debate has become in the US. Either you're naively pro-AI, happy to see massive data center build-out and a handful of frontier model companies become central to everything we do online, or you want to pause research, stop the build-out, and discourage people from using AI. Almost nowhere do I see people talking about the need for decentralized AI. To me, this is the only safe way forward. There's an underlying truth to both sides of the naive debate. The pro-AI camp is right that refusing these technologies leaves enormous consumer welfare gains on the table. An individual using an AI agent to interact with companies and governments can stand up for their own interests rather than accept whatever terms or bureaucracy they're offered without a practical means of negotiating. And refusing to advance AI research would have enormous national security and competitiveness consequences as adversaries develop technologies that could exploit our increasingly vulnerable infrastructure. The anti-AI camp is right that a world where everyone uses the same handful of inference providers for their most intimate affairs rapidly looks like a corporate panopticon. Ask yourself whether you really want every self-driving car to have a corporate kill switch. Sure, we could just regulate and trust giant inference providers, as we've done with Big Tech companies like Google. But the stakes are much higher when these companies sit at the center of virtually all our computing, and increasingly our physical infrastructure. Email and cloud storage were just the beginning. Both of these positions can be true. And their synthesis points toward a third way. We should be supercharging AI development and adoption while encouraging sovereign AI: systems in which individuals have meaningful control over the models, data, and infrastructure providing their inference. An important distinction is between training models and providing inference. Frontier training may require enormous concentrations of capital, energy, and compute. That doesn't mean the companies that train models must also control how people use them. Even the most capable models should be runnable independently of their creators. People should be able to keep their data locally and run models on hardware they control whenever feasible, especially where independence matters for safety, as with self-driving cars and home security systems. They should have access to open-weight models they can inspect, modify, and run without anyone's permission. Where cloud computing is necessary, we need genuine competition among providers that respect user privacy and autonomy. But competition isn't enough. We should develop technical and protocol-based guarantees that limit what even a dominant provider can learn about or do to its users. Technologies pioneered in cryptocurrency could play a significant role. Zero-knowledge proofs could help verify that computations were performed correctly; homomorphic encryption could enable computing on encrypted user data without revealing it to providers; and blockchain-based protocols could coordinate decentralized computing resources and enable new forms of user ownership. These technologies are at different stages of maturity, but point toward privacy and user control guaranteed by system architecture rather than a company's privacy policy. Perhaps users could even own part of the infrastructure providing their inference, sharing in its value rather than merely providing valuable data to someone else's business. There's no reason AI's enormous promise should require surrendering control over our digital lives to a handful of corporations. We can have rapid AI progress and individual technological sovereignty. Indeed, I think the latter is essential if we want the former to be a lasting benefit to humanity.