The most dependable uses of AI in a newsroom are not about writing stories. They sit around the journalism: tagging articles and extracting the people, places and organisations they mention, linking related stories, making the archive searchable, suggesting headlines and metadata, and transcribing audio and video. Journalists remain responsible for everything that is published.

These are also the uses news leaders value most. In the Reuters Institute's 2026 survey of 280 digital news leaders, 97% said back-end automation was important, and back-end tasks such as transcription, copyediting assistance and automated metadata were the use case most often rated very important (64%).

This guide explains how each of these jobs works in practice, where the risks lie, and the editorial safeguards that keep AI-assisted work accurate.

Key takeaways

  • AI earns its place in newsrooms mainly through back-end work such as tagging, metadata, transcription and search, rather than through writing published copy.
  • Tags and entities are only useful when they come from a controlled vocabulary and their accuracy has been measured, not assumed.
  • Related-story links and archive search depend on the same foundation: consistent labels and dates on every story, including old ones.
  • AI-drafted headlines, summaries and transcripts should be treated as unverified drafts that a journalist reviews before publication.
  • Audiences are far more comfortable with news made mostly by humans with some AI help than with news made by AI, even with human oversight, so oversight and disclosure matter.

Where AI fits in the news workflow

The Reuters Institute's Journalism, Media, and Technology Trends and Predictions 2026 report, based on a survey of leaders from 51 countries and territories, found that headline generators and alt-text automation have been widely adopted and that many newsrooms had moved AI pilots into their content management systems over the previous year. The same leaders were measured about the results so far: 13% described the impact as transformational, 44% as promising and 42% as limited.

Audiences draw a similar line. The Reuters Institute's Generative AI and News Report 2026, based on around 2,000 respondents in each of nine countries, found that people are more comfortable with AI being used for tasks such as translation and checking spelling and grammar. Only 12% were comfortable with news made entirely by AI, and 19% with news made by AI with some human oversight, compared with 45% for news made mostly by a human with some AI help and 67% for news made entirely by a human journalist.

Task What AI does What journalists and editors do Main risk
Tagging and entity extraction Suggests topics, people, places and organisations from a controlled list Accept, correct or remove tags Wrong or missing tags, ambiguous names
Related stories Finds stories with shared entities or similar meaning Set rules for freshness and sensitive pairings Outdated or inappropriate links
Archive search Labels old stories so they can be found by topic, date or question Decide what to resurface and add context Old facts presented as current
Headlines and metadata Drafts headlines, summaries, SEO descriptions and alt text Edit and approve before publication Claims the story does not support
Transcription Converts interviews, briefings and podcasts to text Check quotes against the recording Invented words or sentences

Automated tagging and entity extraction

Tagging assigns each story labels from a defined list: topics such as elections or housing, plus attributes such as article type, location or intended audience. Entity extraction picks out the specific people, organisations, places and events a story mentions, and ideally links each one to an identifier, so that Paris the city is not confused with a person called Paris.

Newsrooms have long wanted this. In 2015 Nieman Lab reported on an experimental New York Times tool called Editor, which used machine learning to suggest tags while a reporter was writing, matching the text against an index of terms drawn from every article in the paper's archive. Writers could review the suggestions with as little effort as a spell check, correcting, verifying or removing tags where needed. In the same piece, the FT's then chief technology officer, John O'Donovan, argued that without metadata an archive is full of material of no value, because nobody can find it or knows what it is about.

Use a controlled vocabulary

AI tagging works best when the model chooses from a fixed list rather than inventing labels. A shared standard helps. IPTC Media Topics, the news industry's subject taxonomy, has more than 1,200 terms under 17 top-level subjects, is available in 13 languages and language variants, and includes mappings to Wikidata concepts. Many publishers combine a standard like this with their own house labels; IPTC Media Topics and News Taxonomies Explained covers how.

A practical rollout

  1. Decide the dimensions. Topic, people, places and organisations are the usual start. Article type (news, explainer, opinion, live blog), user need and location are common additions.
  2. Write definitions and examples for every label, including when not to use it.
  3. Build a test set. Have editors tag a few hundred stories by hand, covering every desk.
  4. Measure the model against that test set for each label: how often its tags are right (precision) and how many of the correct tags it finds (recall). Evaluating AI Tagging: Precision, Recall and Human Review explains the method.
  5. Decide where suggestions go. Labels with high measured accuracy can be applied automatically; weaker ones appear as suggestions in the CMS for a journalist to accept.
  6. Backfill the archive with the same labels, so old and new stories can be searched and compared together.
  7. Re-test regularly, and whenever the taxonomy or the model changes.

Common mistakes: near-duplicate labels, tags for passing mentions, and no named owner for the taxonomy.

Related-story modules are only as good as the labels behind them. Two signals are usually combined. The first is shared entities and topics: two stories about the same council, court case or team. The second is semantic similarity, where an embedding model places stories with similar meaning close together even when they use different words; Embeddings and Vector Search, Explained Without the Maths shows how that works.

Editorial rules matter as much as the algorithm:

  • Freshness. In a running story, prefer the latest update; an old one shown as "related" can mislead.
  • Corrections. Exclude or flag stories that have been corrected, superseded or withdrawn.
  • Sensitive pairings. Block links between tragedies and light-hearted pieces, or between a named person and an unrelated crime story.
  • Variety. Avoid filling every slot with the same angle; mix the latest news with an explainer and some analysis.

The Reuters Institute report notes that some degree of homepage personalisation is already widespread on news sites. It also found that audiences struggle to think of original questions to ask publisher chatbots and mostly click suggested prompts, so those tools may work better as features for onward navigation, which is the job related-story links already do.

The same index helps journalists: the CMS can suggest earlier coverage for a reporter to link to.

Archive search and resurfacing evergreen content

Many newsroom archives can be searched only by keyword and date, which fails when a reporter wants every story the paper ran on flooding in one county, or every housing explainer. Labelling the archive with the same topics, entities, article types and dates used for new stories turns it into something editors can filter and query. Newer tools add question answering on top: a system retrieves relevant archive stories and summarises them with links back to the originals.

A Reuters Institute piece republished by Nieman Lab in April 2026 gives examples. The Guardian built an internal chatbot that lets journalists query its archive and experimented with tag pages carrying AI-generated summaries of past events. L'Eco di Bergamo used AI to repurpose more than 70 years of obituaries into a database readers can explore for local and family history. Not every successful archive project relies on AI: The Economist's archive features were largely manual, with staff choosing pieces from scans of old articles, and France's RetroNews works with historians to avoid drawing misleading parallels between past and present.

Finding evergreen stories

Evergreen content (explainers, guides, profiles and background pieces that stay useful) is one of the most reusable assets in an archive. There is no universal definition, so write your own, for example: "still accurate, not tied to a single news event, and useful to a reader arriving from search today". Then:

  • label article type and user need, so explainers and guides can be pulled out by topic;
  • flag time-sensitive facts such as prices, office holders and statistics that need checking before reuse;
  • record when a story was last reviewed, and always show readers the original publication date;
  • add a line of context when an old story is resurfaced alongside new coverage.

Headlines, summaries and metadata suggestions

AI is now commonly used to draft the small pieces of text around a story: headline options, standfirsts, summaries, SEO descriptions, social copy and alt text. The Associated Press's updated newsroom standards, released in July 2026, list approved uses that include suggesting headlines, story summaries and shotlists, and assisting with grammar, spelling and search optimisation. They state that in every case AI-generated output is reviewed and edited by AP journalists before publication. The same standards keep AP's prohibition on using generative AI to create, alter or enhance news photography, and set disclosure standards for cases where generative AI plays a material role in published content.

Practical safeguards for this kind of drafting:

  • Generate several options and let the editor choose and rewrite. Never publish a generated headline automatically.
  • Check that every claim in a headline or summary appears in the story, with the same degree of certainty.
  • Put house style and guidance on sensitive topics into both the prompt and the editor's checklist.
  • For alt text, describe what is visible in the image, not what the story says about it.

Transcription

Transcription is widely valued because it saves hours and makes audio and video searchable. The Reuters Institute's 2026 report describes how The New York Times, investigating a major breaking news event, quickly transcribed thousands of podcasts and videos; the investigation took two weeks, where similar ones had previously taken a year or more.

Speech recognition can also invent text. The Careless Whisper study of OpenAI's Whisper model found that about 1% of transcriptions in its sample contained entire hallucinated phrases or sentences with no basis in the audio, and 38% of those hallucinations included explicit harms such as invented associations or false authority. Hallucinations occurred disproportionately for speakers with longer pauses.

For newsroom use:

  • Treat transcripts as search aids and drafts, and check every published quote against the recording.
  • Keep timestamps, so a reporter can jump straight to the original audio.
  • Pay particular attention to passages around long pauses or silence.
  • Store speaker labels, language and recording details with the transcript. Making Audio and Video Archives Searchable covers the wider workflow.

Editorial safeguards and human oversight

JournalismAI's 2023 global survey, Generating Change, covered more than 100 news organisations in 46 countries. More than 60% of respondents were concerned about the ethical implications of AI for journalistic values such as accuracy, fairness and transparency, and respondents stressed the need for any AI-generated content to be checked by a human. In practice, that means:

Safeguard What it looks like in practice
Human review before publication Nothing AI-generated reaches readers until an editor has approved it
Treat output as unverified AI suggestions are checked like any other unconfirmed source
Measured accuracy Every tagging or transcription task has a test set and a known error rate
Confidence routing Low-confidence tags and uncertain transcript passages go to a person
Audit trail The CMS records which fields were AI-suggested and who approved them
Disclosure Readers are told when AI played a material role in what they see
Accountability A named person re-tests tools and updates guidance as models change

Agree these rules early. The Reuters Institute report observes that product teams often come up against a risk-averse newsroom; written rules on review, disclosure and accuracy give both sides a shared basis for decisions.

How Druid Learning helps

Druid Learning is an AI archival processing platform, based in Dublin and New York, that turns content archives into AI-ready datasets. For NPR it processed more than 2 million assets and applied 20+ labels across 8 dimensions, including user needs, article type, location, topic and community. Journalists receive suggested tags and related stories, editors can search the archive by label, and audience analytics can be broken down by label in Chartbeat and Google Analytics. The Standardization Tool provides real-time tagging and labelling, and Archive Optimization applies the same labels to the back catalogue. Read the NPR case study.

Frequently asked questions

What is the best first AI project for a newsroom?

Automated tagging is usually the best starting point, because so much else depends on it. Consistent topic, entity and article-type labels improve related-story links, archive search, newsletters and audience analytics at the same time. Start with a small set of well-defined labels, build a hand-tagged test set, measure accuracy for each label, and show suggestions to journalists in the CMS before applying any label automatically.

Can AI write headlines for a news site?

AI can draft headline options, summaries and SEO descriptions, and many newsrooms use it for this. The safe approach is to treat every suggestion as a draft: an editor chooses, rewrites and approves it, and checks that each claim matches the story. The Associated Press's 2026 standards, for example, allow AI to suggest headlines but require AP journalists to review and edit all AI output before publication.

How accurate is AI transcription for journalism?

Modern speech recognition is good enough to search recordings and produce draft transcripts quickly, but not reliable enough for unchecked quotes. Research on OpenAI's Whisper found that about 1% of transcripts in its sample contained whole phrases or sentences that were never spoken, more often for speakers with long pauses. Reporters should check every published quote against the original audio, using timestamps to find the passage.

How do you resurface evergreen content from a news archive?

First, label the archive consistently with article type, topic, entities and publication date, so editors can filter for explainers, guides and profiles on a subject. Then check time-sensitive facts such as figures and office holders, show readers the original date, and add context when an old piece is linked from new coverage. AI can find candidates quickly, but an editor should decide what is still accurate.

Do readers mind newsrooms using AI?

It depends on the task. The Reuters Institute's 2026 research found people are more comfortable with AI used for jobs such as translation and spelling and grammar checks than with AI creating news. Only 12% were comfortable with news made entirely by AI, against 45% for news made mostly by humans with some AI help. Visible human oversight and clear disclosure address much of that concern.

Sources