AI assists the research. Humans approve the publishing.

Where we use AI

Monitoring: six research agents check assigned official sources on a daily schedule and detect whether a document has changed.

Summarizing and drafting: agents produce a factual summary, a significance note, an affected-audience list, open questions, and a suggested headline from the document they actually retrieved.

Triage support: grouping likely duplicate submissions so an editor can compare them side by side.

Where we do not use AI

AI is never the source of a factual claim. Every claim traces to a retrieved primary document that an editor can open.

AI does not publish, schedule, approve, or edit live articles. That is enforced at the database level: agent identities can only write research runs, evidence records, and submissions.

AI does not write opinion. Column essays, including What Better Requires, are written by their named author.

Human review before publication

Every agent submission enters a triage queue. An editor reads the evidence, opens the underlying documents, verifies dates, authority and jurisdiction, and only then converts it to a draft.

The draft then moves through the normal editorial workflow — review, approval, and publication by a named human.

Labeling

Entries that began as agent research carry the same standards and the same accountability as anything else we publish. Where AI assistance is material to how a piece was produced, we say so on the entry.

What we send to AI systems

We send publicly available document text and our own editorial prompts. We do not send protected health information, reader personal data, correction attachments, or subscriber lists to AI systems.

Known limits

Automated monitoring can miss a change, misread a table, or misclassify a status. That is why no agent output reaches readers unverified, and why our corrections policy is public and easy to use.