How are newsrooms actually using AI?
Far more for production and research tasks than for writing articles — and the distinction matters, because public discussion focuses almost entirely on the least common use.
Where it is genuinely established:
Transcription of interviews, which was among the first uses and saves enormous amounts of time.
Translation, particularly for monitoring foreign-language sources.
Document analysis. Searching very large leaked or obtained document sets is an obvious fit, and has been part of major investigations — finding candidates for human examination, not drawing conclusions.
Structured-data stories. Automated generation from reliable structured sources — financial results, sports scores, weather, election counts — has been in production for years and is unremarkable.
Headline and summary variants, tested for performance.
Tagging, archiving and search, improving the value of back catalogues.
Accessibility, including captions, audio versions and summaries.
Content moderation of comment sections.
Where it is contested:
Drafting articles, where accuracy and attribution problems are serious and errors have been published, damaging trust disproportionately to the volume involved.
Generated imagery, largely banned in news contexts.
Synthetic presenters and voices, where disclosure obligations are unsettled.
What responsible newsroom policies typically require:
Human accountability for anything published, with a named person responsible.
Disclosure where AI materially contributed.
No AI-generated facts or quotes, ever.
Verification of anything AI-produced, treating it as an unreliable source rather than a tool that returns answers.
Restrictions on inputting confidential material, particularly source-identifying information — which is a serious and underappreciated risk, since pasting unpublished material into a third-party service may breach source protection.
The commercial conflict running alongside. Publishers are simultaneously licensing archives to AI companies, suing them over training use, and blocking crawlers — sometimes all three at once. The unresolved question is whether AI answers that summarise journalism without sending traffic destroy the business model that produces it.