← AI investigation
Methodology

How the investigation is built, and how to read it.

A short description of where the numbers come from, how the claims are tracked, and what we are intentionally not doing.

Source hierarchy

For every load-bearing claim, we prefer primary sources in this order:

  1. Government / standards-body publications (NIST, IEA, EU AI Act, BIS, BLS, EPA, ICO)
  2. Court filings and government inspector reports (Raine v. OpenAI, Robodebt RC, OHCHR)
  3. Tax filings (Form 990s) and EU transparency disclosures
  4. Earnings calls, 10-K filings, official ESG reports
  5. Peer-reviewed academic work (FAccT, NeurIPS, NBER, Nature Computational Science)
  6. Investigative journalism with named sources (Bloomberg, +972, Karen Hao's reporting, MIT Tech Review, FT)
  7. Self-published company documents and lab safety frameworks (treated as primary about the company, not as third-party verification)

Where a claim depends on a single source we say so. Where it depends on a contested estimate (training-run carbon, OpenAI valuation, expected AI productivity) we say that too and link the dispute.

The receipt cards

Every numbered claim on the four act pages is wrapped in a receipt card with a small dot in one of three colors:

  • live — re-verified within the last 30 days. The number is the latest official disclosure.
  • current — verified within the last 90 days. Probably still right but worth a re-check before citing.
  • stale — older than 90 days. Re-check before citing; flag in maintenance pass.

Maintenance cadence

The investigation is treated as a living document, not a paper. Per topic the cadence is roughly:

  • Quarterly — hyperscaler capex (after each earnings cycle), NVIDIA revenue, Anthropic / OpenAI valuation, MoneyFlow data, MIT NANDA-style adoption surveys.
  • Annual — IEA Energy and AI report, Stanford HAI AI Index, Google AI energy disclosure, Pew + Reuters/Ipsos polling, UK AISI Frontier Trends, Open Philanthropy & SFF grants.
  • As-they-happen — court filings (Raine, Character.AI, NYT v. OpenAI), AISI red-team disclosures, EU AI Act enforcement guidance, executive orders, sovereign-AI deals.

Every act page has a footer naming the next earliest expected staleness — this is the single most useful pre-flight check before citing any number.

Editorial stance

The investigation has a thesis (Act IV — discourse displacement) but it is built on documentary primary sources, not on opinion. Two specific commitments:

  • Disagreement is shown. Where credible sources contest a number — Strubell vs Patterson on training carbon, Acemoglu vs Goldman bull case on productivity, the “ChatGPT bottle of water” framing — we cite both sides and explain the methodological gap.
  • Funding is named. When an org is referenced — METR, Apollo, Redwood, Anthropic, AI Now, DAIR — its primary funder is stated. This is not an attack; it is information you need in order to read the org's output.

What is intentionally not in scope

  • Predictions. The capability climb chart in Act III shows extrapolation as extrapolation. We do not bet on AGI timelines.
  • Endorsements. No vendor recommendations. Where models or hardware are named (Llama on M-series, Whisper on iPhone NPU), it is to make a measurement intelligible, not to recommend a stack.
  • Speculative harms without documentation. Act IV's displaced-harms atlas only lists cases with named victims and primary-source coverage.
  • Live-fetched data. All numbers are checkpoint-cited with a date. The page does not call APIs at runtime — that's an editorial choice, not a technical one.

Repository / source notes

The full research notes — every claim with every URL — live in this repository at research-notes/ and research-notes/deep-dives/. Each act page's data files live next to the page itself in _components/. Adding a new pin to the atlas, a new node to a money flow, or a new finding to the bias showcase is a one-file change.

Errors, missing sources, broken receipts: the GitHub link in the top-right is the fastest way to flag them.