The Information Layer

Before AI can act on our behalf, it needs to understand the political world. But today's AI systems carry embedded ideological biases, struggle to reason about forward-looking political questions, and operate within an information ecosystem — prediction markets, news media, social platforms — that has its own deep structural problems.

We built AI agents that trade political prediction markets, measured the ideological leanings of every major AI model, and tested whether collective deliberation among models can sharpen their political judgment. What we found: AI political intelligence is real but brittle, systematically biased, and far from the “truth machine” its boosters promise.

The Representation Layer

AI can now vote on your behalf, advise you on ballot propositions, and execute research programs at superhuman speed. JPMorgan has replaced its human proxy advisors with an AI system to vote $7 trillion in client assets. DAOs are building AI delegates. One in five Americans already asks ChatGPT for political advice.

We built a personal AI delegate that learned our political philosophy and voted the way we would have voted on a hard case. Then we broke it in ten minutes. We embedded invisible text into a proposal, and the delegate flipped its recommendation. The promise of AI representation is real. So is its fragility.

The Governance Layer

Even if individual AI agents can perceive the world and act faithfully, the hardest problem is collective: can a society of AI agents — acting on behalf of diverse humans with conflicting interests — actually govern?

We created an AI legislature where agents with competing goals had to negotiate and allocate scarce resources. They produced a 10,000-word constitution and almost no actual policy. They reinvented gridlock, process creep, and procedural complexity. Democracy is hard. AI doesn't make it easy. But it does let us run a thousand experiments to find what works — and stress-test constitutions before anyone has to live under them.