Where AI is used
A language model does a lot of the reading and writing here, and none of the deciding that matters. This page says which is which, because a site that marks other people's claims ought to be plain about its own.
What a model does
- Screens incoming reports. A cheap model decides whether a report is worth a full explanation, so routine fixtures and promotional items do not consume the desk.
- Writes the explanation. The headline, the summary, why it matters and what happens next are written by a model from the reports, in original wording.
- Extracts the claims. It separates the specific factual assertions out of the reporting.
- Reads the other outlets. For each claim it reports, per outlet, whether that outlet supports it, contradicts it, or does not address it.
- Answers your questions about a story, using only that story's own material — and says so when the answer is not there.
- Translates, when you choose another reading language. That is our English explainer translated, not the outlet's own words.
What a model is never allowed to do
- Set a verification status. Status is computed from who reported what. A confident-sounding explanation cannot talk a claim up.
- Clear a story for publication. Moderation can only ever add caution. It can hold a story the rules would have released; it can never release one the rules held.
- Mark a claim false. That is a person's decision, and it is recorded with their name against it.
- Invent a source. An outlet a model names that is not actually in the story is discarded before it can count as corroboration.
- Answer beyond the sources. If a story's own material does not cover your question, the answer is that it does not cover it.
When the machinery fails
Every one of these steps can fail, and each failure leaves the story in the safer state rather than the more convenient one. If corroboration cannot run, claims stay unverified. If moderation cannot run, the story waits for a person. If translation fails, you get the English. If the day's model budget is exhausted, the desk stops writing rather than writing badly.
Who is accountable
A model is a tool this desk uses, not an author it hides behind. Everything published here is the desk's responsibility. Where we get it wrong, it goes on the corrections page with what changed and when.