Tasks where AI tools are already reliable enough
- Transcription of interviews and press conferences, with a listen-through of the quotes you will actually use.
- Summarising long documents to decide what deserves a full read. The summary is a map, not the territory.
- First drafts of routine items built from source material you already trust: a results table, a council agenda, a weather warning.
- Translation for understanding, not for publication, unless a fluent editor reviews the output.
Tasks where they fail in predictable ways
- Facts without sources. A model asked for a number will produce a plausible number. Ask instead for the number with the document it came from, then open the document.
- Quotes. Never publish a quotation a model produced. Quotes come from recordings and documents.
- Names, dates and places in less covered regions and languages. Error rates rise where training data is thin.
- Recency. Models do not know what happened this morning unless the tool feeds them today's sources.
A five-minute check that catches most problems
- Every number and name in the draft is traceable to a source you can open.
- Every quotation exists in a recording or document.
- The draft does not assert intent or guilt about a named person.
- Images have rights information.
- Someone with their name on the story has read it start to finish.
How this works in Doxx
Doxx is not a general writing assistant. It prepares drafts around a standing brief and delivers them with the sources & checks used to build them, so the five-minute check has something to check against. Editors rewrite, add reporting, cut, or reject. The methodology page describes what stays attached to each delivered item.
See what a draft with its sources attached looks like on a topic you cover.