We measure how companies write about artificial intelligence in their regulatory filings: whether AI appears as a technology they use, a product they sell, a market they depend on, or a risk they carry, and how specifically they say so.

The data are audited, legally mandated corporate reports. They are published on a fixed schedule in a regulated context, which makes them a reliable baseline for tracking corporate AI exposure at scale, unlike surveys or voluntary disclosures.

The first phase covers US filings (Form 10-K and related SEC forms). Other jurisdictions follow once the US pipeline is validated. The target scale is several hundred thousand reports.

1. Filings are collected and parsed into text blocks with their report section.
2. Candidate passages are found with a core AI vocabulary plus industry-specific glossaries (one per ISIC Rev. 5 division). Passages that name AI go to classification; passages that only touch AI-linked topics are kept as hints for a reader.
3. Each passage is labelled against a published codebook: frame, role of AI, risk type, adoption stage, kind of AI, disclosure traits, named entities, verbatim evidence.
4. Several independent coders, human and model, label the same passages; labels are reconciled by majority vote and agreement is reported.
5. Results are aggregated per report, year and industry.

The taxonomy is versioned; every change is recorded with what it replaced and why.

Early. The pipeline runs end to end on single filings; the codebook is at version 0.4; human annotation and the first large-scale run are being prepared. Research questions are being defined with the policy team and are not final.

• classification-pipeline: the pipeline, glossaries, codebook, research plan and change log.

Issues and pull requests in the repositories above. Report data cannot be redistributed; the repositories contain code, glossaries and derived labels only.