AI tools for journalists: a newsroom workflow from pitch to publish

This guide is part of our series on choosing an AI writing assistant you can trust with sources. It walks through each stage of a story and names the type of tool that helps, the risk to watch, and where AI should stay out entirely.

Short answer: the 6 stages and one tool type for each

Stage 1: research (studies, filings and data, plus chat with long PDFs)

Most research time goes into documents: a 300-page budget, a court filing, a peer-reviewed study. Two kinds of tools help.

Scholarly search. For science, health and policy stories, start with the research itself, not a press release about it. Search an index of scholarly papers by topic, then read the abstract, methods and limitations of the two or three most relevant ones. Check whether a paper is peer-reviewed or a preprint, and whether it has been retracted. The Retraction Watch Database is free to search.[1]

Chat with PDFs. Upload the document and ask questions: "Where does this filing describe the 2024 payments?" "What does the methods section say about sample size?" The tool is useful only if it shows where in the document the answer came from. An answer with a page number takes ten seconds to verify. An answer without one is a rumor about the document.

Good research prompts:

List every dollar figure in this budget document with its page number.
Which pages describe [named program]? Quote the first sentence of each.
What does this study say about its own limitations? Give page numbers.
Does this filing mention [person or company]? If not, say so.

The last prompt matters. Ask the tool to say when something is not in the document, and check that too.

Stage 2: interviews (transcription tools and their accuracy limits)

Interview transcription software turns recorded audio into text. Common options include Otter, Rev, Descript, Trint and apps built on OpenAI's open-source Whisper model. They differ on accuracy with accents and crosstalk, speaker labeling, price, and where the audio is processed and stored.

Whatever you use:

• Check every quote against the audio, not the transcript. Transcription errors cluster in exactly the places that matter: names, numbers, technical terms and negatives ("can" vs. "can't").
• Know your recording laws. Some U.S. states require every party's consent to record a call. Tell sources you're recording.
• Protect sensitive audio. For confidential sources, find out whether the tool uploads files to a cloud service and how long it keeps them.

After transcription comes the real work: finding the quotes and checking them. We cover that in interview transcription formats and tools and from transcript to story.

Stage 3: drafting (autocomplete, structure and the inverted pyramid)

AI is most tempting, and most dangerous, at the drafting stage. Fluent text arrives instantly. Some of it may be wrong.

Uses that work:

• Structure. Paste your notes and ask: "Order these facts by news value for an inverted-pyramid story. Don't add anything." Then write it yourself. Our guide to writing a news article step by step covers the structure.
• Getting unstuck. Ask for three possible ledes from your own notes, then write a fourth that's better.
• Context paragraphs. Autocomplete can draft a background sentence, if every fact in it links to a source you can check.

Uses that don't:

• Asking a chatbot for facts, numbers or background without a source.
• Asking for quotes, even "representative" ones. A quote in a news story means a real person said those words.
• Publishing AI-drafted text without a full edit and fact-check. AP's standards say staff do not use generative AI to create publishable content.[2]

Stage 4: verification (claim checks, source quality and reverse image search)

Verification is where AI tools are most useful if you treat them as assistants, not judges. Three tool types:

TinEye is free for non-commercial searches.[3] The InVID-WeVerify plugin, maintained by AFP's Medialab, splits video into keyframes and runs them through several reverse image search engines.[4]

Source quality matters as much as the claim. A real study can still be retracted, a preprint, or cited with the wrong DOI. The full method is in our guide to checking every claim with an AI fact checker, and the editor's version is in how to fact-check an article.

Stage 5: editing (AP style, tone and tracked changes with editors)

Editing is a conversation between reporter and editor, and the tool should keep a record of it.

• Style. Most U.S. newsrooms use AP style. AI proofreading catches many slips (numbers, dates, titles), but no general AI tool applies AP rules reliably. Learn the rules that trip writers up most from our AP style guide, and keep the Stylebook as the final word.
• Tone. Ask a reviewer to flag loaded words, unattributed characterizations and editorializing in a news story.
• Tracked changes. Every edit, human or AI, should be a suggestion the reporter can see and accept. If an AI "tightens" a paragraph, read the before and after: shortening is where meaning drifts.

Stage 6: publishing (headlines and SEO)

Chat tools are good at variations. Ask for ten headlines from your lede, then pick one and rewrite it. Check the winner against the story: a headline is a claim too, and it's the one most readers see. Ask for a 155-character meta description and a search-friendly version of the headline for the web, separate from the print one.

Don't let AI choose the angle. A tool optimized for clicks will pick the most dramatic reading of your story.

Ethics: newsroom AI policies at AP, Reuters and the Guardian

Three of the most cited newsroom policies agree on the basics.

Rules that apply in every newsroom, whatever its policy says:

1. Never let AI invent quotes. Not as placeholders, not as paraphrases presented in quote marks.
2. Never publish a fact you haven't traced to a source. The 2023 CNET case, where an audit found errors in 41 of 77 AI-assisted explainers, shows how fluent errors survive editing.[7]
3. Don't paste confidential material into consumer AI tools unless your newsroom has approved them and you know the data terms.
4. Disclose material AI use in the way your outlet's policy sets out.

How WriteOrbit fits a reporter's workflow

WriteOrbit is a web editor that covers stages 1, 3, 4 and 5 in one place:

• Ask a document, with page numbers. Upload court filings, budgets or studies to the Library and use Ask on one PDF at a time. Every answer carries page numbers, so you can open the page and check the passage before you use it. For questions across sources, AI Chat beside your draft can use the web, your Library or the document, and cites real papers.
• Search 250M+ scholarly works from inside the editor and type @ to cite one, for science, health and policy stories.
• Claim checks and source-quality checks. Run AI review on your draft: it checks claims and flags retracted papers, preprints and wrong DOIs. Results appear as comments in the text.
• Tracked changes for reporter and editor. Real-time co-editing, comments, suggestions and version history, so every change has an author.
• A full editor on your phone. File from the field in a phone browser: it is the full editor, not a viewer.
• Word in and out, for desks that still run on .docx.

What it doesn't do: transcribe audio, check AP style rules, generate headlines with a dedicated tool, or run reverse image searches. Use the tools above for those. See features and pricing.

If you also research documentaries or scripted projects, the same research workflow applies; we compare dedicated screenwriting software separately.

FAQ

Can journalists use ChatGPT?

Usually yes, within limits set by their newsroom. Common rules: use it for research leads, summaries and brainstorming; never publish its text unchecked; never let it supply quotes or facts you haven't verified; don't paste confidential source material into consumer tools. AP, for example, treats generative AI output as unvetted source material. Check your outlet's written policy first.

Do newsrooms allow AI?

Most large newsrooms allow limited AI use under written rules. AP lets staff experiment but not create publishable content with generative AI. Reuters makes the journalist whose name is on a story responsible for it, and promises transparency. The Guardian requires a clear reader benefit, an editor's sign-off and disclosure. The common thread is human checking and accountability.

Which AI tool checks facts?

Several tools help, none fully replace a human check. Claim-checking reviewers flag statements and match them to sources. Google Fact Check Explorer finds published fact checks of a claim. Reverse image tools such as TinEye and the InVID-WeVerify plugin help trace images and video. Every result still needs a journalist to open the source and judge it.

Is it OK to use AI to transcribe interviews?

Generally yes, and it is one of the most common newsroom uses. Check three things: accuracy, because names, numbers and accents trip up transcription; consent and recording laws where you and your source are; and where the audio is stored. Always check any quote you publish against the recording itself, not just the transcript.

Should journalists disclose AI use?

Follow your newsroom's policy, which increasingly requires disclosure when AI materially shaped published content. Reuters and the Guardian both commit to transparency about AI use. Routine uses such as spell-check or transcription are rarely disclosed. AI-generated text or images in a published story generally should be, if they are allowed at all.

Where to go next

For the full comparison of assistants and the six criteria behind it, see the best AI writing assistant guide. The next step in the workflow is verification: read AI fact checkers and how to check every claim in a draft, then chat with PDF and how to check the answers.

Sources

1. Retraction Watch. The Retraction Watch Database. https://retractionwatch.com/the-retraction-watch-database/
2. Nieman Lab (2023). "Not a replacement of journalists in any way": AP clarifies standards around generative AI. https://www.niemanlab.org/2023/08/not-a-replacement-of-journalists-in-any-way-ap-clarifies-standards-around-generative-ai/
3. TinEye. Reverse image search. https://tineye.com/ ; see also Bellingcat Online Investigations Toolkit: TinEye. https://bellingcat.gitbook.io/toolkit/more/all-tools/tineye
4. WITNESS Media Lab (2019). How to use InVID-WeVerify. https://citizenevidence.org/2019/12/11/how-to-use-invid-the-swiss-army-knife-of-digital-verification/
5. Talking Biz News (2023). What Reuters is telling its journalists about using artificial intelligence. https://talkingbiznews.com/media-news/what-reuters-is-telling-its-journalists-about-using-artificial-intelligence/
6. Viner, K., & Bateson, A. (2023). The Guardian's approach to generative AI. https://www.theguardian.com/help/insideguardian/2023/jun/16/the-guardians-approach-to-generative-ai
7. Engadget (2023). CNET had to correct most of its AI-written articles. https://www.engadget.com/cnet-corrected-41-of-its-77-ai-written-articles-201519489.html
8. Google Fact Check Tools. Fact Check Explorer. https://toolbox.google.com/factcheck/explorer