Research

Original research into AI products and the companies behind them: what they cost, what they publish, and what a customer or anyone else can actually find out from the public record. Every paper here ships with the dataset behind it.

Wave 3 · Preprint · Dataset frozen 14 October 2026

EU AI Act Transparency Audit: what 26 AI model providers publish under the Act's transparency duties

A documents-only audit of what 26 providers of general-purpose AI models publish under the EU AI Act's transparency duties, read from their own public documents and checked against the date each duty falls due. Of the 21 providers whose training-content summary deadline had passed for the model studied, 11 published a summary that presents itself as such and opens without an account, on the final, exploratory coding (7 of 21 on the coding fixed before collection). In the places we looked, we found no published Commission list of models with systemic risk.

26 providers1 model per provider~74 min readOpen dataset on Zenodo

Read the full paper

Wave 1 · Preprint · Dataset frozen 17 August 2026

I spent 4.1 billion tokens to investigate 76 AI platforms. They publish what they're selling — not what you're buying.

A documents-only audit of 76 consumer AI products across 15 categories, coded against an instrument published before collection began. 67 of 72 publish a headline price. 7 of 64 document whether you are charged when a generation fails. The open dataset ships with the audit trail, including 90 dated deviations.

76 products15 categories~48 min readOpen dataset on GitHub

Read the full paper

What we hold ourselves to

Findings come from data we collect and verify ourselves, not from figures reported elsewhere.
Every figure states the date it was collected and the size of the sample behind it.
The method and its limits are published alongside the findings, including what the data cannot show.
Every paper ships with the underlying dataset, so a reader can redo the arithmetic.

Every paper here ships its dataset publicly, on GitHub or Zenodo: the records, the instruments, the tools, and every dated correction, so a reader can disagree with us on the evidence rather than on our word.

The standards behind our tool investigations are a separate document: how we investigate. Corrections to research follow the same corrections policy as everything else on this site.