Method, caveats, and every source

What this is

A citizen’s field guide, built from the public record

This site was put together in September 2026 by a private citizen who wanted the companies, the models, the money, the risks and the politics of artificial intelligence in one place, in plain language, with every number traceable to a document. It follows the same approach as Power & the Sound, an earlier field guide on a regional utility.

How it was made. Research, data compilation, charts, maps, the message-drafting tool and the code were produced with the assistance of Claude, Anthropic’s AI model, working from the documents listed below. Anthropic is one of the companies profiled here; the site treats it exactly as it treats every other company, with the same sourcing rules, and readers should weigh that relationship for themselves. The AI did not browse freely and summarize from memory: for each section it ran targeted searches and retrieved specific pages, and every fact on the site is tied to a retrieved page or a page that appeared in search results, labeled accordingly.

Method

• Primary sources first. Company announcements, SEC and court filings, official pricing and documentation pages, peer-reviewed papers, government reports (IEA, Lawrence Berkeley National Laboratory, Census Bureau, Federal Reserve), and the official pages of Congress and state governments. Where those were not retrievable, major outlets (Reuters, AP, NPR, CNBC, Bloomberg, Fortune, TechCrunch, The Verge) and, as a last resort, Wikipedia or trade compilations, which are always labeled as such.
• Three confidence labels. Company profiles, the homepage headline figures and the verification list use verified (a primary source or a fetched major-outlet page), reported (a major outlet reported it but the page could not be retrieved), and unverified (only secondary compilations carry it). Forecasts are labeled projected and named by scenario where the source gives one. Nothing unverified appears as a headline figure.
• Lab claims are labeled as claims. A benchmark score published by a company about its own model is presented as that company’s claim, with the reporter named. Independent evaluations are named as independent. Results from controlled experiments, such as Anthropic’s 2025 blackmail simulation, are described as simulations, not as deployed behavior.
• Balance by evidence, not by symmetry. Where credible sources disagree (whether AI is a bubble, whether it is displacing jobs, whether data centers are raising bills), both are presented with their sources. Where the weight of evidence is lopsided, the site says so.
• Dates are as of Sept 13, 2026. This field moves weekly: models are replaced, valuations change, bills advance. Always follow the source link before relying on a number, and check the change log for corrections.

What it is not

It is not a news organization: nobody here interviewed anyone, filed a records request or verified a claim by phone. It is not neutral about the value of evidence: sourced claims are preferred to unsourced ones regardless of who makes them. It does not endorse any of the positions on the take-a-position page; those are offered as starting points across the spectrum, and each names the real bills and actions a constituent can ask about. It does not collect any data about visitors: the message drafter runs entirely in your browser and sends nothing anywhere.

Worth double-checking

These items rest on secondary sources, single reports, aggregator figures, or facts that were unresolved at publication. They are the least certain claims on the site. Each is phrased so a reader can check it. The policy page carries its own short version of this list.

Corrections

Found an error, a newer number or a better source? Note the page, the figure and the source, and email David Geller <[email protected]>. Corrections are logged, with dates, on the change log.