An AI Model That Runs a Government
Overview
Governance is a decision-making problem. At any given moment, a national government is managing hundreds of departments, hundreds of billions in spending, and interdependencies no single person fully tracks. The question we wanted to answer was simple: what does an AI system do when you put it in charge of that and stop telling it what to think?
We built Centre Block, an experiment in autonomous decision-making, using Canada’s federal government as the domain. We gave it one instruction: “Do what you believe is best for Canadians.” Everything else it figured out on its own.
How It Works
The system runs on Blocky 4, our structural model of how the federal government is actually assembled: departments, budgets, programs, and the dependencies between them. Each cycle, it reads the current state of the country, decides what to act on, and drafts a response in the form a real bill or directive would take.
Each cabinet portfolio has its own AI minister — Finance, Housing, Health, Immigration, Defence, and others — operating autonomously within their department, grounded in real Canadian law, budget data, and precedent. They work and report back to an AI Prime Minister who holds the full picture and makes final calls. The whole structure mirrors how a real cabinet government actually functions, with distributed decision-making and central accountability.
Before anything goes out, it also runs its own internal review. Constitutionality is checked against the Charter. Cost and accountability go through an Auditor General function. Anything with foreign policy implications gets assessed for consequences abroad. Once a proposal clears all hurdles, it publishes.
When something material breaks in the news, the system doesn’t wait for the next scheduled cycle. It moves the way a government actually has to.
What It Found
Two hours into its first test run it flagged two things on its own. Grant programs designated for Canadian businesses were flowing to companies with no Canadian employees, some collecting federal tax benefits on top. Separately, dozens of federal departments, from Immigration Canada to Parks Canada, each run independent HR operations, duplicating work that could be centralized. The two findings together represented roughly eleven billion dollars in potential annual savings. No department had come up with either one.
Why This Matters
The savings number is not the point. The point is how it got there. It wasn’t directed to audit grants or review HR structures. It built its own picture of how the government fits together, noticed something that didn’t reconcile, and followed it. That’s the behavior we were looking for.
We believe AI is going to directly take on governance and managerial roles within the next two years. The transition is already underway in lower-stakes contexts, and it will move up. The question that actually matters is not whether that happens but how these systems reason when they get there. What do they prioritize? Where does their judgment hold and where does it break down? What do they catch that humans miss, and what do they miss that humans catch?
Almost everything known about model behavior comes from benchmarks with fixed answers. That measures knowledge retrieval. It says nothing about decision-making under genuine uncertainty in open-ended domains. Centre Block is an attempt to study that directly, with a system actually running, producing real outputs, on the record.
Conclusion
Centre Block is a demonstration and a research instrument. The eleven billion came from the first version, in its first two hours, from a single line of instruction. What matters is what a system built this way surfaces once it has run for a year.
It’s live now. Everything it produces is accessible, at centreblock.ai and on X at @CentreBlockAI.