The United States needs adaptable and responsive government institutions to effectively tackle the complex challenges facing the country both domestically and internationally. One major obstacle to efficient and agile policy implementation is the accumulation of “policy cruft”– outdated, redundant, overly-complicated, or conflicting elements within our laws and regulations. Over time, this policy buildup has grown beyond the scope of manual review. In response, potential solutions from emerging technology tools are cropping up, chief among them being Stanford’s Regulation, Evaluation, and Governance Lab (RegLab).
Each year, governing bodies at the federal, state, and local levels approve new laws and regulations to fulfill their many responsibilities in service to the public. However, without consistent efforts to revise or eliminate outdated language, policies accumulate over decades, leading to vast and cumbersome bodies of law. At the federal level, the U.S. Code (all the general and permanent federal laws passed by Congress) spans nearly 53,000 pages. This cascades down into tens of thousands of additional pages as the White House and federal agencies issue policy guidance for implementation, not to mention the over 200,000 pages of regulations in the Federal Code of Regulations. The same thing happens at the state and local levels. For example, San Francisco’s municipal code alone comprises almost 19,000 pages.
“Policy cruft” makes it more difficult for lawmakers to create streamlined policies, for civil servants to faithfully implement programs, and for the public to understand and navigate government services. The advancement of generative AI and large language models (LLMs) offers a powerful potential tool to cut through this inefficiency, streamlining policy implementation and making government institutions more responsive and capable.
Recent work by Stanford’s RegLab demonstrates the significant role LLMs can play in public sector reform. RegLab partners with federal, state, and local government agencies to leverage AI and data science to modernize government and enable it to more effectively and fairly serve the public. For example, they have developed an AI tool to identify, redact, and map racially restrictive covenants in 5.2 million deed records for Santa Clara County. Other projects include using AI to assist staff at Colorado’s Department of Labor and Employment with unemployment insurance fact-finding, and creating an LLM-based search tool to quickly analyze the U.S. and municipal codes.
The potential applications for LLMs for public sector reform are both exciting and diverse. Imagine using these tools to identify duplicate or conflicting statutory requirements, detect tensions between proposed legislation and existing code, or even reimagine how citizens engage in the notice-and-comment rulemaking process. However, as we explore these possibilities, it is crucial to prioritize prototyping and rigorous evaluation, especially regarding legal reasoning (given the well-documented risks of LLMs). An LLM’s simplification process could overlook or oversimplify decades of nuanced legal, regulatory, and policy understanding.
That said, in certain cases, this simplification can be beneficial. For instance, the infamous (and ill-fated) Enterprise Service Bus – long considered a necessary contracting requirement – proved a misguided solution. By carefully harnessing the strengths of LLMs while accounting for their complexities and limitations, we can strike a balance, ensuring that we use these technologies responsibly.
Cutting through policy cruft supports better policy implementation, but we are still in the early stages of discovering the most valuable use cases. Further exploration is necessary, which is why Niskanen and RegLab are working with innovative government partners and delivering real value through pilot projects. As we absorb hard-earned lessons over time, the next step will be to share case studies, playbooks, and resources to scale impact across the country. Keep an eye on this space for bureaucratic breakthroughs.