Methodology

Entries move through a four-stage lifecycle: intake, review, publication, and ongoing verification. Intake has two channels, and both feed the same human review.

Every change to every entry is recorded in the public edit history visible on each company profile page. This append-only log is the primary mechanism against astroturfing.

Every claim must be backed by at least one cited source, and every source is classified into a three-tier taxonomy. The tier determines what a source can do on its own.

Three sourcing rules apply. We distinguish which are enforced by code and which by editorial review, because the difference matters to anyone auditing us:

We reject claims based solely on speculation, rumors, or sources that cannot be independently verified. When evidence is disputed, the disputed status is shown on the entry. Where sources do not support a specific value, fields are honestly marked unknown rather than filled with plausible guesses.

We do not tell you how to feel about AI. We document what companies do and let you draw your own conclusions.

We do not accept payments from companies to influence their entries. Companies cannot submit corrections directly; they must go through the same community process as everyone else. All edits are logged and public.

Labor impact classifications describe real-world effects, not moral judgments. A practice classified as automating a prior human task is factual, not an accusation, and it carries the strictest evidence standard on the platform.

The record is three tables: companies, practices, and evidence. A company has many practices, a practice has many cited sources, and every claim is traceable to at least one source row. Three more tables hold the public records about a company, the fields of each record with the quoted words behind them, and the none-found lines that state where a search for records found none. The full codebook defines every field and value; this is the summary.

Every structured field takes one of a fixed set of values, and each set includes an explicit sentinel (unknown, unclear, other) for when the sources do not say. The values and their labels:

Two fields are easy to misread. The completeness score is computed from which fields the cited evidence supports; it is set automatically and is not an editorial judgment. A company's “last verified” date is the most recent time a moderator published or edited a practice or a cited source under that company; it does not mean every claim was re-checked on that date.

Read the full codebook, or fetch the same definitions as a machine-readable schema at /datapackage.json.

The record is built from public documents, reviewed by people, and it has the limits that implies. Anyone relying on it should know these before drawing conclusions.

One citation form is used on every surface: an APA-shaped reference with Trace Foundation as the author, the date the data was current, and a stable link that survives a company rename. Live pages add the date you accessed them; a dataset release carries a DOI instead.

Trace Foundation. ({year}). {Company name}: documented AI practices (data as of {YYYY-MM-DD}) [Data set record]. AI Trace. https://www.aitrace.org/r/company/{id}. Accessed {Month D, YYYY}.

Trace Foundation. ({year}). {Company name}: {practice title} (data as of {YYYY-MM-DD}) [Data set record]. AI Trace. https://www.aitrace.org/r/practice/{id}. Accessed {Month D, YYYY}.

Trace Foundation. ({year}). {Briefing title} (data as of {YYYY-MM-DD}) [Briefing]. Generated from AI Trace, https://aitrace.org/. Accessed {Month D, YYYY}.

Trace Foundation. ({year}). AI Trace export {export id} (data as of {YYYY-MM-DD}) [Data set extract]. AI Trace. https://aitrace.org/. Accessed {Month D, YYYY}.

Trace Foundation. ({year}). AI Trace dataset (Version {YYYY.MM}) [Data set]. Zenodo. https://doi.org/{doi}

Braced items are filled in from the record. The stable link uses the record's id and redirects to the current page, so a citation keeps working after a rename; the data-as-of date is the record's most recent content change. A BibTeX entry with the same fields is offered beside every citation.

License

The factual data (company names, AI practice descriptions, evidence citations, and the fields exported with them) is licensed under the Creative Commons Attribution-ShareAlike 4.0 International license (CC BY-SA 4.0). You may share and adapt the data for any purpose, including commercial and machine-learning purposes, on the license's terms: give appropriate credit to AI Trace, link to the license, note any changes you made, and distribute adapted material under the same license.

Trace Foundation will not assert the ShareAlike condition against machine-learning model weights or model outputs trained on the data. Attribution still applies where the data itself is redistributed. The license applies to the factual data only; the site software, design, and branding remain the property of Trace Foundation, Inc.

If you believe an entry is inaccurate, you can file a dispute directly from the practice detail page. Disputes are reviewed by moderators and resolved transparently.

Dispute types are: wrong or fabricated information, source inaccuracy, miscategorization, outdated information, and biased framing. Every reviewed dispute's outcome, whether accepted or rejected, is published on the entry's page with a resolution summary written by the reviewer, and an accepted dispute's changes additionally appear in the public edit history.

The dispute submission itself is never republished. What appears publicly is our own text: the dispute category, the dates, the verdict, and the reviewer's summary. Submissions that are not genuine disputes, such as spam or abuse, are dismissed without a public record; the transparency promise applies to disputes reviewed on their merits.

We do not silently remove contested information. If an entry is disputed, it carries a “Disputed” badge until the dispute is resolved.

Corrections follow the same transparency rule. When an error is verified, the entry is corrected and the change is recorded in the public edit history. An entry that turns out to be materially wrong is marked Retracted rather than deleted: the record of what we got wrong stays public, because a transparency platform does not get to hide its own mistakes.

The most valuable thing you can do is submit evidence. If you notice a company using AI, report it via the Report a Sighting form. It takes under 90 seconds. No account required.

You can also help by sharing news articles about corporate AI practices, linking to company profiles in relevant discussions, and filing disputes when you spot inaccuracies.

AI Trace tracks corporate AI use. We hold ourselves to the same standard. Our own AI practices are documented on our company profile at /company/trace-foundation.

Every entry here starts with automated discovery and ends with human judgment. No human team can continuously track the public record of hundreds of companies, so our automated research pipeline, built on commercial AI models, does the finding and the first draft: it watches company newsrooms and regulatory filings for new developments, searches public sources, and drafts candidate entries with citations. Human reviewers decide what is accurate enough to publish, and every published entry records who reviewed it and when, field by field, on the entry itself. This disclosure does not depend on watermarking or AI-detection technology; we describe how our system works whether or not tools exist that could detect it.

Nothing AI-assisted is published without human moderator approval. The specific commercial models in use are documented on our company profile and updated as they change, so this page describes the architecture and its guarantees rather than any one vendor.