Seven million pages of local law became searchable. The hard part is knowing which rule still applies

A resident trying to work out whether a fence is legal may find that the answer lives in a poorly scanned PDF from a council meeting years ago. A journalist comparing noise rules across neighbouring towns faces the same obstacle multiplied hundreds of times. Researchers at the University of California, Berkeley, have built an AI-assisted public archive meant to bring that scattered layer of government into the search age. It is as much a project about access to law as about language models.[1]

Announced on 2 October, the project aggregates millions of local ordinances from across the United States. Its research team gathered nearly 10,000 large documents — approximately seven million pages — from government and other hosting sites. The United States has more than 3,000 counties and roughly 6,000 additional local jurisdictions making rules. State and federal legal databases have long existed, but comprehensively searchable county and municipal law has been far harder to obtain.[1]

The first problem was not answering questions

Before any chatbot could summarise a rule, the researchers had to locate the original texts, download them responsibly and extract their contents. Local ordinances appear in inconsistent forms: a municipal code website, a scanned meeting binder, a multipage PDF with skewed lettering or a version with confusing amendments. These are document-engineering problems. Even the strongest reasoning model cannot reliably compare what it cannot read.[1]

The team used LightOnOCR, a vision-language optical character recognition system, to reconstruct text and structural information from difficult source documents. Subsequent AI and data-processing tools helped standardise and classify the material. That transformation allows a researcher to search across jurisdictions that never agreed on the same digital format. It is not evidence that every law has been interpreted correctly, and it does not automatically resolve whether an older ordinance has been repealed.[1]

Berkeley provides an open-access entry point through the Responsibility database and demonstrates how others can build narrower search tools around particular questions. A housing researcher could compare how cities regulate accessory dwelling units. A public-health investigator could examine local rules about tobacco retailing. A small publisher might track rules affecting street markets in a group of towns. The value is the ability to discover patterns that were previously obscured by the difficulty of gathering the texts.[1][2]

More searchable does not always mean more authoritative

Legal information has an unusual dependence on dates and jurisdiction. The same address can sit within city, county and state rules that interact or conflict. A council may amend one paragraph without reproducing the entire code, and a court may change how a provision is interpreted. An AI system that assembles an apparently neat answer from old documents can mislead a reader more efficiently than an inaccessible PDF ever did.[1]

A trustworthy product built on this database therefore needs to do more than generate fluent prose. It should return the relevant jurisdiction, ordinance number, enactment or amendment dates and a link to the primary text. It should distinguish an ordinance from a proposed measure, and signal when the source material cannot establish the rule currently in force. Readers making consequential decisions still need the appropriate official confirmation or professional advice.[1]

There is a further distinction between discovering a rule and measuring its effect. A comparison might reveal that one city adopted a particular zoning restriction and another did not. That does not prove the restriction caused differences in rents, housing supply or health. Such questions require careful causal research, including controls for local economics, demography and enforcement. The new database supplies a much better starting point for those studies, not the final statistical answer.[1]

The project's origins are revealing. Researcher Diag Davenport wanted to study bias in the laws themselves, but first encountered the absence of a usable inventory. A longstanding political-science problem became solvable partly because document extraction and computing grew capable enough to process millions of pages. In this case, AI is not replacing legal scholarship; it is lowering the cost of seeing the evidence at all.[1]

An opportunity for local accountability

For a local newsroom, the archive could make cross-border reporting less prohibitively expensive. A reporter investigating rules that disproportionately affect street vendors might identify comparable ordinances across dozens of jurisdictions before requesting enforcement data or interviewing affected residents. The most valuable outcome would be better questions for officials, not an automated article claiming that every local code has already been understood.[1]

The project is a reminder that public access is not secured simply by publishing a document somewhere online. Information can be technically public and practically inaccessible at the same time. AI did not write these laws, and it cannot determine whether they are fair. It helped build the instrument through which citizens can finally examine far more of them together.[1]

Report an error or suggest a correction