I built an interactive map of global AI policy
An experiment to figure out which AI laws actually matter before they matter, with help from two AI agents and a healthy distrust of their output.
AI policy is happening everywhere right now. Over a thousand policy initiatives across more than 70 countries, according to the OECD. Forty-something countries with active AI-specific legislation, according to Stanford’s 2026 AI Index. The EU is rolling out enforcement guidelines for the AI Act. Colorado’s law takes effect this summer. France’s CNIL and Italy’s Garante are setting precedents through enforcement that will get cited for years. Texas, New York, Utah, and Tennessee all passed substantive AI laws over the past year. South Korea, Japan, and Singapore each took different approaches to the same questions.
I built a tool to map all of it. It is live now at pulse.humanactually.com.
It is called AI Policy Pulse. It updates itself weekly. It tracks enacted laws, proposed bills, regulatory guidance, enforcement actions, and the kind of frameworks practitioners actually have to track (NIST RMF, EU AI Act guidance, CNIL advisories, that kind of thing). The part I am most interested in talking about is how I got the system to do this without hallucinating. More on that below. First, the basics.
What it is
A visual map of where AI policy activity is concentrating, with commentary on what is driving the pressure. There are three views: a pressure map showing where regulatory activity is intensifying right now, a heat calendar showing how that activity has built up over the last twelve months, and a pulse wave showing what changed in the last thirty days. I built all three because they answer different questions, and I am genuinely curious which one people will reach for most often.
This is V1. Coverage is real, but not exhaustive. I started with the jurisdictions where AI policy is happening most actively and where I can verify developments from authoritative sources. More jurisdictions will get added as I prove the pipeline can handle them without compromising quality.
The verification problem
Here is the part I spent the most time on, and the part I am most interested in talking about.
The whole thing collapses if even a small percentage of entries are wrong in a way someone can screenshot. AI summarization is useful for reading hundreds of pages of regulatory text every week. AI summarization also hallucinates. If I publish summaries of regulations that do not quite exist, or that exist with different terms than what I summarized, the credibility is gone the first time someone clicks through to the source.
So I had to figure out how to use AI for the heavy lifting while catching its mistakes before anything goes public.
The approach I landed on uses two independent AI passes over each source. The first agent reads the source and pulls out structured facts. The second agent reads the same source separately, without seeing the first agent’s output, and pulls out the same facts. I compare what each one extracted, field by field. Jurisdiction. Status. Effective date. What the rule actually requires. If they agree, the entry is eligible to publish. If they disagree, it queues for editorial review and I look at it before anything goes live.
A single LLM pass can be confidently wrong. Two independent passes that agree on extractable structural facts are much harder to fool. Where they disagree, that disagreement itself is the signal that something needs human attention.
Every entry has a confidence score reflecting how strongly the agents agreed. The score is a deterministic calculation based on field-by-field agreement. The whole methodology is on the methodology page and I will keep it updated as the system evolves.
The Pressure Index
Every jurisdiction on the map has a Pressure Index score reflecting how much regulatory pressure is currently active there. The score weights three things: how much new activity has emerged recently, how many of that jurisdiction’s entries are flagged as highest urgency, and how many describe active enforcement rather than proposed rules.
The weights are judgment, not science. I picked them based on what I think people need to see first. The trend arrow next to each score shows whether pressure is rising, stable, or falling compared to the previous month.
What it does not claim to be
It is not exhaustive. AI policy is being made in lots of places I cannot reliably track yet. The methodology page is honest about coverage gaps.
It updates weekly, not in real time. For breaking news, follow regulators directly.
It will sometimes be wrong. Verification reduces error rates, it does not zero them out. Every entry links to its primary source, which is the actual trust mechanism. If something looks wrong, click through and check. When you find errors, tell me and I will fix them.
Where I am uncertain
This is part of an ongoing thread of experiments at Human Actually where I have been exploring what trustworthy AI built pipelines actually look like in practice. Earlier pieces in that thread are smaller and more interactive — the Trust Dossier, the Deepfake Family Drill, and a trust-yourself-versus-the-AI quiz. This one is bigger in scope, but the same spirit: a hands-on experiment in whether AI tools can be built to be genuinely useful while staying honest about their limits.
I am calling this V1 because I am genuinely curious how it will land. I have hypotheses about who will care most. I have hypotheses about which view will get used most. I do not actually know.
If you use it and find a missing jurisdiction, an error, or a feature that would make it more useful, I want to hear about it. That feedback is what makes the next version better than this one.