Skip to main content

1 of 17 desks have not been read in over 48 hours. Their pages say when each was last read. Anything those desks show as upcoming or underway may have changed — check the source before relying on it.

Articles edited by a language model

Three different mechanisms have edited articles on this site, and this page lists every article any of them touched. Between Aug. 9 and Aug. 18, 2026 a language model was handed 37 articles and published the prose it returned; that mechanism was withdrawn, and no person reviewed either version of those 37. Since Sept. 1, 2026 a second one runs on request: the model is shown an article's own numbered sentences and replies with the numbers to delete, so it cannot write, rephrase, reorder or add text — every word that remains was already in the article, sentences carrying a caveat, an attribution or a source link cannot be cut at all, and a person asks for each edit and approves the result before it publishes. Since Oct. 30, 2026 a third runs on the roads desk only: the model may rewrite an article from the same record, so a rewritten article can contain wording the template did not write; the template's version publishes first, a person approves each rewrite before it replaces that text, and the article is stamped ai-drafted. Every article not listed below is exactly what its template wrote from the record. At the time the first 37 were published this page said a machine check proved the model added no fact and removed none; the check was real but weaker than that sentence — it compared tokens, not claims, and it could not see a figure recombined from digits already present, a number reattached to a different thing, or a name removed. For 10 of the 37 the template's own wording survives in the article's version history; for the rest the edit was applied before the article was first published, so the template's original wording was never written to any version. Each article says which mechanism edited it, and which of those two cases it is.

This page exists so that claim can be checked rather than taken on trust. Every article below was edited, either shortened or, on the roads desk, rewritten by a language model; every article not below was published exactly as its template wrote it.

38 of 6,099 published articles were edited — 0.6%.

Two kinds of edit are listed. In a shortening the model never adds a word. It is shown the article’s own numbered sentences and replies only with the numbers to delete; the article is reassembled from its remaining text, unchanged. Sentences carrying a caveat, an attribution or a source link are locked and cannot be cut at all. A program then re-checks the result against the article as the template wrote it, and refuses the edit if a source link, a caveat, or a figure, identifier or date that distinguishes this record from its siblings has gone — it does not check every name, and a figure this desk repeats on every article is not treated as distinguishing. An edit that fails the check is not published and the article stays exactly as it was. A person asks for each edit and approves the result before it publishes. A roads rewrite is different: on the roads desk only, the model may write new prose from the same record, so a rewritten article can hold wording the template did not write. The template’s version publishes first, a person approves the rewrite before it replaces that text, and the article is stamped ai-drafted with editedBy: language-model. The 37 articles from the withdrawn trial of Aug. 9–18, 2026 were made a different way: a language model wrote the prose, and no person read either version. How this is verified is described on why you can verify this.