AI in the Newsroom: What It Can and Cannot Do

Quick answer — Newsroom AI divides into settled and unsettled. Translation, transcription and archive search are largely settled and in production. Generation, summarisation and anything with a byline attached are not.

The settled half

Translating a filed story through document translation. Transcribing an interview. Searching twenty years of archive. Generating subtitles for a video package.

These are mechanical, verifiable, and the failure modes are visible. They are in production at serious mastheads and the argument about them is over.

The unsettled half

Drafting copy. Summarising a story for a newsletter. Writing a headline. Anything that produces text a reader will attribute to a journalist.

The objection is not quality. It is that a masthead's value is the promise that a person checked, and automation at the point of authorship weakens exactly what the reader is paying for.

Translation sits on the safe side, with conditions

A translated story is still the reporter's work. The claim, the sourcing and the attribution are unchanged, which is why translation raises fewer editorial objections than generation does.

The conditions are real though. Names and titles have to match local convention, quotes have to survive intact, and a mistranslated attribution is indistinguishable from a fabricated one to the person quoted. quality control with review on named individuals is the control that matters.

Speed has to be tiered

A results table can publish unattended. A court report cannot. The tier belongs to the desk rather than the language, and it has to be enforced by the workflow rather than by a policy document nobody reads at eleven at night.

Where to start

Transcription and archive search, which nobody objects to, then translation on one low-risk desk.

News media localization covers the audience case.

FAQ

Which newsroom AI uses are settled? Translating filed copy, transcribing interviews, searching the archive and generating subtitles. All mechanical, verifiable, with visible failure modes, and in production at serious mastheads.

Why is generation different from translation? Because a masthead's value is the promise that a person checked. Translation leaves the claim, sourcing and attribution unchanged; generation touches the point of authorship the reader is paying for.

What is the specific risk in translated reporting? Attribution. A mistranslated quote is indistinguishable from a fabricated one to the person quoted, which makes review on named individuals the control that matters most.

How should speed be managed in a newsroom? By tiering per desk rather than per language, and enforcing it in the workflow. A results table can publish unattended; a court report cannot, and a policy document will not stop that at eleven at night.