AI is already routine in most newsrooms, even among journalists who distrust it. In the Reuters Institute's survey of 1,004 UK journalists (fieldwork August–November 2024), 56% used AI for professional tasks at least weekly and 27% used it daily, while only 16% had never used it for a journalistic task. The most common uses are unglamorous language work: 49% use AI for transcription or captioning monthly or more, 33% for translation, and 30% for grammar and copy-editing. Fewer lean on it for the reporting itself — 22% for story research, 16% for headlines, and just 10% for generating first drafts.
The same journalists are wary of the tool they keep reaching for. In that survey, 62% called AI a "large" or "very large" threat to journalism, and 60% were extremely concerned about its effect on public trust, 57% about accuracy, and 54% about originality. That tension is the right posture. The wins are concrete — a 90-minute interview transcribed in minutes, a dense public report turned into a list of angles to chase, a technical story rewritten for a general audience — but every one of them is raw material a reporter still has to verify.
The guardrails are not optional. The Associated Press tells staff to treat any generative-AI output as "unvetted source material" and does not use it to create publishable content. Reuters will not publish AI-generated stories or images, requires a human to greenlight anything AI touched, and forbids putting unpublished stories into open tools like ChatGPT. Those rules exist because the failures are real: Ars Technica fired a reporter in 2026 over AI-fabricated quotes, and The New York Times ran a correction after a reporter rendered an AI-generated summary as a direct quotation.
56% of UK journalists use AI for professional tasks at least weekly and 27% daily; only 16% have never used it for a journalistic task (Reuters Institute survey of 1,004 journalists, 2024)Source ↗
The most common uses are transcription/captioning (49% monthly or more), translation (33%) and grammar/copy-editing (30%); only 10% use AI to generate first draftsSource ↗
62% of journalists call AI a 'large' or 'very large' threat to journalism, and 60% are extremely concerned about its impact on public trustSource ↗
The AP treats any generative-AI output as 'unvetted source material' and does not use it to create publishable content; Reuters will not publish AI-generated stories or images without human greenlightingSource ↗
Once a story is filed, you still owe the desk a headline, a standfirst, and platform-specific social posts — often against a clock. Headlines are already one of the more common AI uses (16% of journalists), and it works precisely because the reporting is done: the AI only rearranges words you have already reported and verified, so there is nothing new to fact-check.
Prompt
You are a subeditor helping a reporter write headlines and social copy for a story that is already finished, edited, and fact-checked.
The full article text (the ONLY source of facts you may use): {{article_text}}
Publication and audience: {{outlet_and_audience}}
Produce:
1. Six headline options — a mix of straight/informative and more engaging, all under {{headline_char_limit}} characters, counting characters for each.
2. One standfirst/subhead of 20–30 words.
3. Three social posts: one for X, one for LinkedIn, one for Instagram, each in the outlet's voice.
Rules:
- Use only facts, names, numbers, and quotes that appear in the article text. Do not add, sharpen, or extrapolate any claim — if a headline would need a fact that is not in the text, do not write that headline.
- No overstatement: the headline must be supportable by the body of the story, not a stronger version of it.
- Preserve the exact spelling of every name and place.
- Flag any option that leans on the single strongest claim, so I can check it is not overselling.
Fill in your details and the prompt updates live — then copy.
Transcription is the single most common AI task in journalism — 49% of journalists use it monthly or more — because turning a 90-minute recording into searchable text by hand eats an afternoon. The catch is that AI transcription mishears names and figures, so the transcript is a finding aid, not a citable record: every quote you publish still gets checked against the audio.
Prompt
You are helping a reporter mine a cleaned interview transcript. Work only from the transcript I paste — do not add context or interpretation from outside it.
Interview transcript with timestamps: {{transcript}}
What the story is about: {{story_focus}}
Produce:
1. A 150-word summary of what the source said, in neutral language.
2. The eight most quotable passages, each copied word for word with its timestamp, so I can locate and verify it against the audio.
3. Three follow-up questions the interview left unanswered.
Rules:
- Quote verbatim. Never smooth, paraphrase, or "clean up" a quote and present it inside quotation marks — reproduce the transcript's exact words.
- Do not attribute any statement the transcript does not clearly attribute to the source.
- Where the transcript marks a word as [inaudible] or the meaning is unclear, keep the marker and flag it — do not guess what was said.
- Add no facts, dates, or names that are not in the transcript.
Fill in your details and the prompt updates live — then copy.
A 200-page budget, a court filing, or a regulator's report lands in your inbox and the news could be on page 140. Journalists use AI for story research (22%) to get through dense public documents faster — reading for structure, contradictions, and numbers that do not add up. The output is a to-do list of leads to verify, never a set of facts to publish.
Prompt
You are a research assistant helping a reporter find leads in a long public document. Work strictly from the document text I paste — treat everything you produce as an unverified lead I will confirm myself.
Document (public record): {{document_text}}
What I am looking into: {{beat_or_angle}}
Produce:
1. A one-paragraph plain-language summary of what the document is and covers.
2. Up to eight potential story leads. For each: the claim or figure, the exact page or section where it appears, and why it might be newsworthy.
3. Any internal contradictions or figures that do not reconcile, quoting both places.
4. A list of what is notably missing or unexplained.
Rules:
- Cite a page or section reference for every claim. If you cannot point to where something is in the document, do not include it.
- Use only the document. Do not add background, context, or figures from your own knowledge — if outside context is needed, write [VERIFY: needs external reporting] instead.
- Do not infer intent or wrongdoing; describe only what the document states.
Fill in your details and the prompt updates live — then copy.
You have an interview in two hours with someone you have never covered, and prep is where reporters quietly reach for AI. It is genuinely useful for organizing what you already know into a briefing and a sharp question list — but dangerous the moment you let it supply biographical "facts," because models confidently invent details about real people, and publishing those is a defamation risk.
Prompt
You are helping a reporter prepare for an interview. Organize the material I give you into a prep brief — do not add biographical facts, career history, or quotes about this person from your own knowledge.
Who I am interviewing and their role: {{interviewee}}
Background I have gathered and verified (the only facts to rely on): {{verified_background}}
The story and what I need from this interview: {{interview_goal}}
Produce:
1. A half-page brief organizing my verified background into: who they are, why they matter to this story, and known points of tension or sensitivity.
2. Twelve questions ordered from rapport-building to the hardest accountability questions, each open-ended.
3. For each tough question, a likely deflection and a follow-up that closes the door on it.
Rules:
- Use only the background I provided. Do not state any fact about this person that is not in my notes — where a useful fact is missing, write [RESEARCH: confirm independently] instead of supplying one.
- Do not fabricate prior quotes, votes, dates, or affiliations.
- Keep questions neutral and non-leading.
Fill in your details and the prompt updates live — then copy.
A story reported for the business desk needs a plain-language version for a general newsletter, or a piece has to reach readers in another language. Translation (33%) and copy-editing (30%) are among the top AI uses in newsrooms. The rule that makes it safe: the AI may change the wording, never the facts — and any machine translation gets a human check and a disclosure, exactly as Reuters labels its own.
Prompt
You are an editor adapting an already-published, fact-checked story for a different audience. You may change wording, structure, and reading level. You may not change, add, or remove any fact.
Original story (the fixed source of every fact): {{original_story}}
New audience and format: {{target_audience}}
Reading level / length target: {{level_and_length}}
Produce the adapted version, then a short "changes log."
Rules:
- Preserve every name, number, date, and direct quote exactly as written. Do not round figures, update them, or rephrase anything inside quotation marks.
- Explain jargon in plain terms without introducing any new claim or example not in the original.
- If a sentence cannot be simplified without a fact I did not provide, leave it and note it in the changes log rather than inventing detail.
- If translating, mark the output as machine-translated and flag idioms or legal/technical terms a bilingual editor should confirm.
Fill in your details and the prompt updates live — then copy.
A FOIA release lands as 4,000 pages, or a public dataset has thousands of rows, and you need to find the handful that matter. This is where AI genuinely earns its place in investigations — ICIJ used machine learning to cut one review from 110,000 documents to 3,000. The discipline that makes it defensible: keep sensitive material on infrastructure you control, and confirm every hit by hand.
Prompt
You are helping a reporter triage a batch of PUBLIC records to decide what to read first. Everything you return is a candidate flag I will open and read myself.
Records (public documents, pasted or listed): {{records}}
What I am hunting for: {{search_criteria}}
Produce:
1. A ranked shortlist of the entries most likely to match my criteria, each with the document/row identifier and the specific text that triggered the flag.
2. Named entities that recur across the set (people, organizations, amounts, dates), with where they appear.
3. Obvious gaps — missing dates, redactions, or numbering breaks worth questioning.
Rules:
- Point to a document ID, page, or row for every flag. If you cannot cite where a match is, do not list it.
- Use only the records provided. Add no outside context, and do not infer motive or guilt — describe only what the text shows.
- Where a document is redacted or unclear, say so and flag it for manual review rather than filling the gap.
Fill in your details and the prompt updates live — then copy.
Common questions from journalists
Does my newsroom allow AI, and do I have to disclose it?
It depends on your outlet, so read your own policy first — 60% of UK journalists say their organization has AI guidelines. As a baseline, AP and Reuters permit AI for assistive tasks but forbid publishing AI-generated content without human verification, and Reuters discloses when a tool is material to the result (for example, labeling machine-translated stories). When in doubt, verify with an editor and disclose.
Is it safe to put interview recordings or source documents into ChatGPT?
Not confidential ones. Inputs to consumer AI tools can be stored and used to train models, and Reuters explicitly bars putting unpublished stories into open tools like ChatGPT. Public records are generally fine; recordings of off-the-record sources, leaked documents, and anything that could identify a confidential source should stay out of consumer tools and, if needed, be handled with a locally run model or infrastructure you control.
Can I trust AI to get facts and quotes right?
No — treat every AI output as unvetted source material, in AP's words. Models fabricate quotes, statistics, and biographical details with total confidence. Ars Technica fired a reporter in 2026 over AI-fabricated quotes, and The New York Times corrected a story that turned an AI summary into a direct quotation. Verify every fact against a primary source and every quote against the original recording before it publishes.
Will AI replace journalists?
The reporting says no, even as adoption climbs. Journalists themselves are wary — 62% call AI a large threat — but the tasks AI handles are transcription, translation, and copy-editing, not sourcing, verification, or accountability. AP frames AI as "not a replacement of journalists in any way." The judgment about what is true and worth publishing stays human, and it is exactly the part the tools cannot do.