I built AI Policy Docs in spring 2024 to keep up with how the US government was regulating AI. Every federal agency publishes its proposed rules, notices and final rules in the Federal Register. That’s dozens of documents a day, and only a few of them are about AI. Finding those, and spotting the ones still open for public comment, meant a lot of reading.
AI Policy Docs did the reading. It was active at aipolicydocs.org through 2024.
How it worked
- Every day, a script pulled each new document from the Federal Register. Over the site’s life it read more than 36,000 of them.
- Claude read each one and decided whether it touched AI. That meant AI itself, but also chips, compute, AI safety and the CHIPS Act. It wrote a few bullet points on what the document said and tagged it with a topic, such as Geopolitics & Defense, Healthcare, or Policy & Standards.
- The AI list grew to 2,310 documents. You could filter it by type, search it, show only documents open for public comment, or sort by how many people had viewed each one on the Federal Register.
- Clicking a row opened the document: its summary, the official abstract, the comment deadline and a link to the original.
The search assistant
You could also ask a question in plain English, like “what rules cover AI in healthcare?”. The assistant found the closest documents and wrote an answer that pointed to them.
What I planned next
The tabs for US Congress, the UK, the EU and China were there from the start, waiting to be filled. I also wanted to read the public comments on each rule, to see which organisations were weighing in and what they asked for. That part never got built here. I came at it from the other side at a hackathon in June 2024, showing how easily LLMs can flood public comment systems.
It was built with React and Ant Design, a Node server, a Postgres database, Algolia for search and Claude 3 Haiku for the summaries, hosted on Railway. The code is on GitHub.