AI will not replace reporting — being there, making the call, working the source, verifying the record, and taking responsibility for what's true. It will absolutely replace the hour transcribing an interview, the afternoon reading a 300-page document dump, and the blank stare at a headline that won't come. This handbook is the practical middle ground: where AI genuinely helps journalism, the three rules that protect your credibility and your sources, and five workflows you can use this week — with verification and source protection front and center, because those are the ones that define the work.
~18 min readno coding5 workflow recipes1 judgment exercise
This is a guide to working with AI tools, written by an engineer — it is not legal or ethics advice, and it is not a substitute for your newsroom's AI and ethics policies, journalistic standards (SPJ Code of Ethics and equivalents), or source-protection obligations. When they conflict with anything here, they win.
01
Where AI actually helps — and where it doesn't
The single most useful mental model: treat AI like a fast, tireless assistant who transcribes and drafts well, does no reporting, and will fabricate facts, quotes, and sources with total confidence. You'd happily hand that assistant an interview to transcribe or a document to summarize. You would never publish its "facts" unverified, and you'd never trust it with a confidential source.
Task
Verdict
Why
Transcribing interviews and audio you provide
Excellent
Huge time saver — you still check quotes against the audio
Summarizing long documents, reports, transcripts you provide
Excellent
Grounded in text you gave it — far less room to invent
Headline, subhead, and structure brainstorming
Strong
Options to react to; you choose and sharpen, and check accuracy
Research starting points & question lists
Strong
Good at "what should I look into / ask?"; every fact verified
Plain-language explanations of complex topics
Strong
Translation for readers — which you fact-check against sources
Facts, figures, quotes, sources — unverified
Never unverified
Models fabricate them convincingly — see Rule 1
Confidential source material, or AI text as original reporting
Never
Source protection and honesty — see Rules 2 & 3
02
The three non-negotiable rules
Rule 1 — Verify everything. Never publish an unverified fact, figure, or quote.
Language models generate plausible text, not verified truth — they will state wrong facts, invent statistics, and fabricate quotes and sources with total confidence. A model can produce a realistic quote attributed to a real person who never said it, or cite a study that doesn't exist. AI does no reporting — it can't attend the event, make the call, or check the record. Every fact, figure, quote, and source in a story gets verified against primary sources by you, exactly as it would without AI. Treat everything an AI tells you as an unconfirmed tip, never a fact — a single fabricated detail can end a story's credibility and yours.
Rule 2 — Protect your sources and confidential material.
Free and consumer AI tools may retain what you paste and train on it, which can expose confidential documents, unpublished reporting, and — most seriously — the identity of confidential sources. Protecting sources is a core ethical obligation, and a leak through a tool is as damaging as any other. The discipline: never put source-identifying information or sensitive material into a consumer tool; use only newsroom-approved tools with contractual no-training and retention commitments for anything sensitive; and treat the confidentiality of what you feed an AI as seriously as the accuracy of what comes out. When a source's safety is at stake, keep it out of the machine entirely.
Rule 3 — Be transparent, and never pass AI text off as reporting.
Follow your newsroom's disclosure policy — the direction is toward transparency. Using AI to transcribe or brainstorm generally needs no disclosure; presenting AI-generated text, images, or summaries to the audience often does. And passing off AI-written material as your own reporting is a form of deception and a plagiarism-adjacent problem — AI output can also reproduce another's phrasing without attribution, so any AI-assisted text is checked for unattributed passages. The principles: be honest about AI's role where it matters to readers, never present AI output as original human reporting or as verified when it isn't, and keep attribution clean.
03
Five workflows you can use this week
Each recipe: what to give the AI, what to ask, and what you must verify by hand. The prompts are starting points — keep the verification and source-protection discipline.
1 · Transcribe and pull quotes from an interview
Provide the audio/transcript (a newsroom-approved tool if the source is sensitive, per Rule 2):
Transcribe this interview, then list the 8–10 most newsworthy quotes with a timestamp for each. Quote the exact words — do not paraphrase or clean up grammar in the quotes. Flag anything inaudible or unclear rather than guessing at it.
You verify: every quote you'll use against the actual audio (transcription errors change meaning, and the model may "smooth" a quote), and confirm the timestamp — a misquote is a serious error even if the transcription was 99% right.
2 · Summarize a document dump or long report
Provide the documents and ask for structure and leads, not conclusions:
Summarize these documents: the key facts and figures stated, who the main actors are, timeline of events, and anything unusual or potentially newsworthy. Quote exact language for anything I might use, with a location reference. Do not add outside context or draw conclusions — just surface what's in the documents and flag what's worth digging into.
You verify: every fact and quote against the actual document (the model can misread or invent), and treat the "newsworthy" flags as leads to report out — the summary points you at the story, it doesn't write it.
3 · Headlines, structure, and angle brainstorming
Give your verified story facts and the outlet's style:
Here are the verified key facts of my story [paste, no unverified claims]. Suggest 10 headline options (accurate, not clickbait, no overstatement), 3 possible structures, and 2 alternative angles I might be missing. Every headline must be supportable by the facts I gave you — don't imply anything beyond them.
You verify: that no headline overstates or implies something the reporting doesn't support (the model tends to sensationalize), and that the facts you fed it are the verified ones — a great headline on a wrong fact is a correction waiting to happen.
4 · Research starting point and interview prep
Ask for the map and the questions, then report it out yourself:
I'm reporting on [topic]. What are the key questions to investigate, the kinds of sources and records I should seek, and the tough questions to ask [type of interviewee]? Frame everything as questions and avenues — do not assert facts, statistics, or conclusions, which I'll gather and verify myself.
You verify: everything — this is a reporting plan, not reporting. The "questions only, no facts" instruction steers the model away from fabricating claims (Rule 1) toward genuine usefulness: helping you not miss an angle or a question.
5 · Explain a complex topic for readers
Write (or verify) the accurate explanation first, then have AI simplify:
Rewrite this accurate explanation for a general audience: [your verified explanation]. Plain language, no unexplained jargon, keep every fact exactly right. Don't add any new claim, number, or example I didn't include. Preserve the nuance — don't oversimplify to the point of being wrong.
You verify: that simplification didn't introduce an error or drop a crucial caveat (the model trades accuracy for readability if unchecked), and that nothing new slipped in. Read it against your sources one more time.
04
The judgment exercise: spot the danger
Three scenarios from real practice patterns. Pick what you'd do — the point is calibrating when the rules bite.
1. Summarizing a report, the AI produces a perfect, quotable line: "'This is the worst crisis in a decade,' the agency's director said." It's exactly the quote you needed. Use it?
Rule 1. Models fabricate realistic, attributed quotes — a "perfect" quote is exactly what a hallucination looks like, and "reportedly" doesn't cover a quote that was never said. Publishing a fabricated quote attributed to a real person is a catastrophic journalistic error. Every quote is verified against the primary source (the document, the audio) before it runs.
2. You're working a sensitive story and want AI to help organize documents from a confidential source. The fastest option is a free chatbot. Do you use it?
Rule 2. A consumer tool may retain and train on the documents, and redacting obvious names rarely removes everything that could identify a source (metadata, phrasing, specifics). Protecting a confidential source is a core ethical duty, and a leak through a tool is as damaging as any other. Newsroom-approved, no-training tools only — or keep it out of the machine.
3. On deadline, an AI writes a clean, publishable news brief from your notes. It reads like your work. File it under your byline as-is?
Rules 1 & 3. "Reads fine" says nothing about whether the facts and quotes are right (verify all of them), and AI text can reproduce another's phrasing without attribution (a plagiarism risk). Filing AI-written text as your own reporting without verification or the required disclosure is deceptive. Fact-check, check attribution, and follow policy — the byline means you stand behind it.
05
Choosing tools: the questions that matter
You don't need to understand the technology to procure it well. You (and your newsroom) need answers, in writing, to five questions:
Question
Answer you want
Is our data / source material used to train your models?
No, contractually — not "you can opt out somewhere in settings"
How long are prompts and uploads retained?
Defined, short, and deletable on request
Where is data processed and stored?
A stated jurisdiction you can trust for source protection
Security posture?
SOC 2 Type II or equivalent, encryption in transit and at rest
Admin controls and access management?
Yes — the newsroom should be able to manage and audit use
General-purpose enterprise AI (Claude, ChatGPT's business tiers) covers the transcription (via approved transcription tools), summarizing, and brainstorming recipes here — never the reporting or verification. For sensitive material, the bar is a vetted, no-training tool, or no tool at all. And your newsroom's AI and ethics policy precedes everything here; if there isn't one yet, these three rules are a reasonable seed.
06
Quick answers
Can I trust AI to fact-check for me?
No — AI is a source of leads, not truth; it fabricates confidently. Use it to surface what to check, then verify every fact against primary sources yourself. AI does not do verification.
Is AI transcription reliable enough to quote from?
It's a huge time saver, but always check any quote you'll publish against the original audio — transcription errors change meaning, and a misquote is a serious error.
Do I have to disclose AI use?
Follow your newsroom's policy; the trend is toward transparency. Assistive uses (transcription, brainstorming) usually don't require it; AI-generated content shown to the audience usually does — and never pass AI text off as original reporting.
Where should a skeptical journalist start?
Recipe 1 — transcribing an interview you conducted, then checking quotes against the audio. Immediately useful, low-risk, and it keeps you in the "verify against the source" habit the rest depends on.
Finished this one? 0 / 208 Handbooks done
Explore the topic
See this alongside everything else on the same subject — handbooks, system designs, challenges and tools, in one place.