Our editorial standards

Every story on AI Broadsheet comes from a named newsroom, lab, government or community source, through its public feed. We show the publisher's headline (or a labelled AI desk headline), a short excerpt and, where we may, the publisher's own photo with credit, and we link to the full article.

Videos play in the publisher's own YouTube player. Podcast episodes play from the show's own audio file. Both are credited to the channel or show.

Our software sorts stories into desks using published keyword rules, and groups headlines that describe the same event so you can see every outlet's coverage side by side.

The lead story is the event the most newsrooms are reporting right now. “Developing” means three or more outlets reported it in the last six hours. “Breaking” means three or more outlets reported it within the last two hours. These labels are counted, never chosen.

For the stories leading the site, Claude, an AI model made by Anthropic, writes a new headline and a brief of at most two sentences, in English and French. It works only from the headlines and excerpts the publishers themselves published for that story.

It may not add any name, number, date, place, quote, motive or consequence that isn't in that text. Before anything is shown, our software checks every number and every name in the new copy against the publishers' text. If one is missing, the copy is thrown away and the publisher's headline stays.

Every AI desk headline is labelled, the story page shows the publisher's original headline and excerpt, and the link to the original article is always there. If you find an error, use “Report an error”: we correct it and log it.

A Dispatch is an article our desk writes, with AI, when two or more outlets are reporting the same event. The AI model (Claude, made by Anthropic, or Google's Gemini when Claude is unavailable) works only from the headlines and excerpts those outlets published. It adds no background, no reactions and nothing from memory.

Before a dispatch is published, our software checks every number and every name in every part of it, in English and in French, against the outlets' text. One miss and the draft is thrown away. A point is shown as “reported by several outlets” only when two or more independent publishers state it as fact; a company's own announcement, or a single outlet's report, is shown as a claim in its name. Nothing is called independently verified unless we tested it ourselves. What the reporting leaves open is listed as unknown, never guessed.

Every dispatch is labelled “Written by the AI Broadsheet desk with AI”, names the outlets it comes from, and links each point and each paragraph to the report it came from. “Listen” uses a synthetic ElevenLabs stock voice, or your browser's own voice. If you find an error, use “Report an error”: we correct it and log it.

Translation: headlines and summaries without an AI desk version are translated for readers of the other language, labelled “Translated with Claude”, with the original shown.

Grouping: Claude checks whether headlines grouped together really report the same event. Funding pages: it flags official pages that may have changed; an editor verifies and updates every listing. Newsletter: it drafts The Morning Broadsheet; an editor reviews every draft before it is sent.

AI output can contain mistakes. Our checks catch many of them, not all: we write only about stories publishers have reported, we never make pictures that look like news photos, and when we get something wrong we correct it in public.

Sometimes an editor writes a sharper headline for a big story. We do that only after reading the original article, we keep it true to the reporting, and the story page always shows the publisher's original headline.

Opinion appears only on the Editor's desk, under the author's name, and is never mixed into the news feed.

Advertisers never choose, change or preview our coverage. Ads are labelled “Advertisement”; sponsored content is labelled “Sponsored” and is never placed in the news feed.

If something on our pages is wrong — a headline filed under the wrong desk, a broken credit, an error in our own writing — use the “Report an error” link on any story page. We fix it and log every correction to our own work, dated, on the corrections page. Errors in a publisher's article belong to that publisher; we'll point you to them.