Using Artificial Intelligence to Build State Capacity

Start by getting rid of pointless reports. Build on that success.

It’s an unfortunate fact of modern American life that government officials are routinely required to do dumb jobs that nobody wants them to do.

The trouble is that it’s much harder than it sounds to figure out what those dumb jobs are. Legal requirements accrete over time, and because there’s little political glory to be had in getting rid of them, they just sit on the books for decades, sapping public time and resources.

So it’s very cool see that Stanford Law’s Dan Ho has joined forces with David Chiu, the San Francisco city attorney, to take a stab at this “policy sludge”:

A team from Stanford’s Regulation, Evaluation and Governance Lab trained an AI program to chew over a city’s legal code in search of every instance in which a city department is mandated to produce a report. Then Chiu’s team went to those departments to see which reports could be tweaked for efficiency, combined with similar requirements, or slashed altogether.

“This tool saved us countless hours of work,” Chiu said. “Because of the length of our code … it’s likely a project we would never have undertaken.”

Now Chiu is sponsoring city legislation that would change more than a third of the nearly 500 reporting requirements that can be altered by a city ordinance. Chiu wants to do away with 140 of them entirely.

“This isn’t just a San Francisco problem,” Chiu said, referencing a report that described the millions of pages produced by Congress every year as a black hole. “We need to be delivering results and services, not just churning out more reports,” he said. “Particularly in this era of budget scarcity we need to get up staff time to focus on the truly pressing issues of the day.”

This is exciting work on its own terms. But it’s also proof of concept—an early example of how AI might aid in building state capacity. As Ho says, “[t]he partnership demonstrates a repeatable playbook for how to leverage rapid advances in AI to identify obsolete requirements, to cut down on procedural bloat, and modernize code and regulation.”

At the core of the mission is something that Ho and his team call STARA—the Statutory Research Assistant. STARA is an LLM that they’ve trained to do statutory and regulatory surveys. They’re looking to build it out over time to do take on more complex tasks. You can read more about STARA here.

Think about permitting reform, for example. There’s wide agreement that we’ve made it much too hard to get permission to build housing, trains, highways, energy infrastructure, and more. But eliminating one permitting chokepoint—by, say, adopting a law exempting infill projects from the California Environmental Quality Act, as the state legislature is currently considering—will inevitably expose other chokepoints. Permitting reform is not something you just do in a day. It’s a process of repeatedly hacking away at the legal kudzu vines that strangle state capacity.

Just figuring out what those kudzu vines are is a challenge. You need to review thousands of federal, state, and municipal laws; pull through reams of regulations, building codes, union rules, and more; and understand how dozens of different agencies, commissions, and offices apply their processes on the ground. It’s a wildly complex task, and one for which AI may be especially well-suited.

More generally, Ho’s collaboration with San Francisco offers a refreshing counterpoint to the anxiety that official use of AI tends to provoke. Yes, incorporating AI into governance can present risks. Yes, the Department of Government Efficiency has acted recklessly in deploying AI tools to dismantle agencies and harass civil servants.

But AI also holds great promise, as Ho and I argued last year. “Humans are expensive, fallible and slow. Sometimes, AI can help do the job faster, more accurately, and at less taxpayer expense.” In a devilishly complex world, reformers can use all the help they can get to improve the legibility of our legal system. STARA is a good start.