The useful part is unglamorous. Transcribe a 90-minute interview and search it. Feed 400 pages of a records release into a tool that answers only from those pages, with citations. Clean a spreadsheet of campaign filings and ask what changed year over year. Draft the nut graf three ways and pick one. Each of these used to cost an afternoon.
The dangerous part is that the same tools produce fluent, confident prose with fabricated quotes, invented studies, and experts who do not exist, and they do it most often when you ask for exactly what you need. In 2025 a syndicated summer reading list ran in major papers recommending books that were never written. Every incident like it started with someone trusting the output.
So the rule: AI can help you find, organize, check, and phrase. It cannot be a source, and nothing it produces goes into a story until you have traced it to a document, a dataset, or a person you spoke to. If your newsroom has no policy, AP's approach (treat output as unvetted source material) is a sound default.
Quick wins this week
- Transcribe your next interview with Otter or Descript and search the transcript for the exact quote instead of scrubbing audio; verify it against the recording before you use it.
- Load a public report or court filing into NotebookLM and ask what it says about your question, with page citations, so a 300-page document fits in the hour you have.
- Paste a public dataset into Claude or ChatGPT and ask what changed and what three questions a reporter should ask; then recompute any number you plan to print.
- Ask a chat assistant to explain a bond covenant, a zoning variance, or a clinical-trial phase at three levels, then confirm with an expert on the record.
What AI can do for journalists, task by task
Interview transcription and quote retrieval
Record with consent (some states require all parties to agree), transcribe with Otter or Descript, and search the text for the quote instead of scrubbing. Auto-transcripts drop negations, mangle names, and merge speakers, so verify every quote against the audio before it goes in. Check where the tool stores recordings and for how long, and keep confidential interviews out of tools your newsroom has not approved.
Records requests and document dumps
Upload the release into NotebookLM, a Claude Project, or Gemini and ask questions that can only be answered from the documents, with page citations. Scanned pages need OCR first; Gemini and ChatGPT read images of documents reasonably well. Open every cited page yourself; the summary is a finding aid, not a source. Ask for every name, dollar figure, and date the set contains as a starting index.
Data analysis for stories
Paste or upload the spreadsheet into ChatGPT, Claude, or Julius and ask for the cleaning steps, the calculation, and the code, so you can rerun it. Models misread ambiguous headers, silently drop rows, and get arithmetic wrong, so recompute anything you will print in a spreadsheet and ask a source who knows the data whether the pattern is real or a reporting change.
Background research and expert sourcing
Use Perplexity or a chat assistant with web search for background, and open every link it cites; a citation that does not say what the summary claims is common. Never ask for 'experts on X' by name; models invent professors and titles. Find people through their published work and confirm them in the institution's directory before you call.
Drafting and structure from your notes
Give the model your notes and the outlet's style and ask for a draft that uses only what is in the notes, with '[NEED]' markers where the reporting is thin. Ask for three ledes or a nut graf rather than the whole story if you want to keep your own voice. Every quote in the draft must appear verbatim in your notes; check each one.
Pre-publication fact check
Ask the model to list every claim, number, name, title, date, and quote in your draft and mark which your notes support. Verify the rest or cut them. Have it flag characterizations that would be defamatory if wrong and numbers that may have been rounded. It checks against your notes, not the world; it is a second pair of eyes, not the fact-checker.
Beat monitoring and agenda packets
Paste the 200-page council agenda packet or a regulator's docket into NotebookLM or Gemini and ask what on it touches your beat, with page numbers. A Zapier flow can send new filings from an RSS feed or inbox to a chat model for a one-paragraph summary each morning. Treat summaries as tips; read the item before you write about it.
Prompts for journalists
Replace the bracketed placeholders, paste into any chat assistant, and iterate on the result.
Interrogate a records release
You are a research assistant working only from the documents I have uploaded. Do not use outside knowledge. Answer these questions: [YOUR QUESTIONS]. For each answer quote the exact passage with the document name and page. If the documents do not answer a question, say so. Then list every named person with the role the documents give them, every dollar figure with its context, every date in order, and anything that contradicts something else in the set.
Tip: Works in NotebookLM, a Claude Project, or Gemini with files attached; the citations are the point.
Clean and question a dataset
Act as a data journalist's analyst. I am pasting [DESCRIBE DATASET, SOURCE, AND DATE RANGE]. First describe the columns and flag anything ambiguous or dirty (mixed date formats, blanks, duplicates, units), and ask me before assuming. Then compute [WHAT YOU WANT, E.G. YEAR-OVER-YEAR CHANGE BY CATEGORY] and show the steps or code so I can rerun them. Give me the five findings a reporter should check first and, for each, what would make it wrong. [PASTE DATA]
Tip: Recompute the headline number in a spreadsheet before it goes anywhere near a draft.
Draft from notes, no invention
You are a copy editor helping me draft. Write a [LENGTH]-word news story in [OUTLET] style from the notes below. Use only facts, quotes, and attributions in the notes; where a paragraph needs something the notes lack, write '[NEED: X]'. Never add a quote, statistic, or expert that is not in my notes. Attribute every claim the way my notes do. Lead with [THE NEWS OR ANGLE]. Short sentences, most important fact first. Notes: [PASTE REPORTING NOTES]
Tip: Count the NEED markers; that is your remaining reporting list.
Pre-publication claim audit
Act as a fact-checker at a rigorous magazine. Below are my draft and my notes. Build a table of every factual claim, number, name, title, date, and quote in the draft: where it appears, whether my notes support it (yes, no, partly), and what would verify it. Flag any contested claim stated as fact, any characterization of a person that could be defamatory if wrong, and any number that may have been rounded. Check only against my notes, not your own knowledge. Draft: [PASTE DRAFT] Notes: [PASTE NOTES]
Tip: Keep the table with the story file; it is your record if the piece is challenged.
Interview prep from the record
You are preparing me to interview [NAME AND ROLE]. I have pasted [WHAT YOU HAVE, E.G. PRIOR STATEMENTS, AN OFFICIAL BIO, A REPORT THEY AUTHORED]. Using only this material, summarize their positions, note where their statements changed over time, and draft twelve questions in order: three easy, six substantive, three that press on the inconsistencies. For each hard question quote the passage it rests on so I can cite it if they push back. Do not add facts about this person from your own knowledge. [PASTE MATERIAL]
Tip: The 'only this material' line matters; without it the model will invent a scandal.
Want a prompt for something else? Use the Prompt Builder.
Skills to build
Verification as reflex
Why: Every fabricated quote that reached print passed through someone in a hurry, and models are most wrong when they sound most sure.
How: Adopt one rule: nothing from a model enters a story without a document, dataset, or person behind it. Run the claim audit on every piece for a month until it is automatic.
Source-grounded research
Why: A model answering from training data is guessing; a model answering from your documents with citations is a finding aid you can check.
How: Do document work in NotebookLM, Claude Projects, or Gemini with files attached, and open every citation. Never ask an open chat for facts about a person or event.
Data literacy with a spreadsheet beside the chat
Why: Public data is where the stories nobody else has live, and the model makes the analysis approachable but not reliable.
How: Ask for the method and the code, rerun the key number yourself, and learn enough pivot tables to catch a wrong join.
Protecting sources while using cloud tools
Why: A source's identity in a chat log is a record a vendor could be compelled to produce, and shield laws may not reach it.
How: Keep identifying details out of prompts, use newsroom-approved accounts with retention controls, and know where your transcription audio lives.
Knowing your newsroom's AI policy
Why: Outlets differ on disclosure, bylines, and which tools are allowed, and a policy you have not read cannot protect you.
How: Read it, ask about the gaps, and if there is none, propose one built on AP's standard: unvetted source material, never publishable output.
Tools worth knowing
Otter.ai
Meeting transcription and summaries that follow you into Zoom, Teams, and Meet.
NotebookLM
A research notebook that only answers from the sources you give it, with citations.
Claude
A careful writing and analysis assistant that shines on long documents.
Perplexity
An answer engine that cites its sources, built for research rather than chat.
ChatGPT
The general-purpose AI assistant most of your coworkers already use.
Julius AI
Chat with your spreadsheets and data files and get charts, stats, and answers back
Cautions for journalists
Ask a chat assistant for a supporting quote or study and it will produce one, with a plausible name, title, and journal. The 2025 syndicated reading list with nonexistent books ran under real bylines in real papers. Every quote must come from a person you or a colleague spoke to or a document you read, and every study needs a link you opened.
Never paste a confidential source's name, unpublished documents, or identifying details into a consumer AI tool unless your newsroom has approved it. Consumer tools may retain conversations and use them for training, vendors can be subpoenaed, and shield laws vary by state and may not cover data on a third party's servers. Use newsroom accounts with training off and retention limits, and keep identities offline.
Models attach adjectives ('disgraced', 'embattled', 'controversial') and confuse people who share a name. Strip every characterization the draft did not get from your notes, verify identities with a second detail such as age or title, and remember the model is not a libel lawyer.
Follow your outlet's rule on disclosing AI assistance, and never publish AI-generated images or audio as if they were recorded. Readers who catch one undisclosed use stop trusting the rest, and platform labeling rules for synthetic media apply to news accounts too.
A model's training has a cutoff and it will state old facts (who holds an office, a company's latest results) as current. Check any 'latest' claim against a dated source before it appears in copy.
Your 30-day plan
- Week 1: Read your newsroom's AI policy or ask your editor what the rules are. Move interview transcription to an approved tool, check its retention settings, and keep source identities out of any prompt.
- Week 2: Put one records release into NotebookLM or a Claude Project and report from it with citations. Draft one story from notes only and count the NEED markers.
- Week 3: Do one data story with ChatGPT, Claude, or Julius, asking for the code and recomputing the key number yourself.
- Week 3: Run the claim audit on every story before filing and keep the table with the story.
- Week 4: Use the interview-prep prompt on your next hard interview and the agenda-packet workflow on your beat. Tally the hours saved and put them into reporting.
Frequently asked questions
Will AI replace journalists?
Can I use ChatGPT to write news articles?
Is it safe to use AI transcription for confidential interviews?
How do I know if an AI tool made up a source?
Terms used on this page
Related roles
- AI for Content WritersAI can research, outline, draft, and edit alongside you. The writers who thrive with it use it to reach the interesting part faster, keep their own voice, and are straight with clients about how they work.
- AI for PR ProfessionalsAI can profile a reporter from their last twenty stories, draft the pitch, turn a call into a briefing doc, and pull a coverage report in minutes. It cannot know who is real, what your client actually said, or what the FTC now counts as a fake testimonial. Those stay your job.
- AI for Social Media ManagersFeed the AI your voice guide and your best posts and it will draft a month of captions, cut a webinar into clips, and summarize what your audience is saying. You still decide what ships, and you still own the labels and the DMs.