Gains:
- Ability to produce inverted pyramid skeleton, spot and header alternatives based on verified material with artificial intelligence support
- Ability to control and correct the title beyond the body, unattributed claims, and additions not included in the material.
- Understanding that artificial intelligence does not produce material, it edits verified material and the final reading belongs to the journalist
With verified facts and quotes in hand, it's time for the most visible task: writing the news. Journalism writing has its own unique architecture. The inverted pyramid — a structure that puts the most important information at the beginning and the details at the end — allows the reader to understand the event in the first sentence. Spot (lede/lid) — the first paragraph of the news, a summary answer to the questions "who, what, where, when, why, how" — is the lifeblood of the news. Artificial intelligence (AI) is a quick help in generating outlines of these structures, coming up with title alternatives, simplifying a text, and catching language errors. But let's keep the emphasis: AI gives the first draft and language suggestions; The accuracy of the facts, the appropriateness of the references, the impartiality and the final sentence published belong to the journalist. The difference between having the AI "make up news" and "make it edit verified material" is the limit of the profession.
Step by step: AI-powered news draft
Step 1 — Prepare the ingredients. Have verified facts, verbatim quotes (time-stamped), sources and context. The AI is given “stuff”; AI is not asked to "make stuff up".
Step 2 — Build the skeleton. Request inverted pyramid skeleton from AI: spotlight, main body, background, side views, result. Have them mark which verified fact each section is based on.
Step 3 — Generate the spot with several options. Write 3-4 different spots from AI; choose the strongest one and recheck your facts. Spot is the place with the most errors because it is the most intense compression.
Step 4 — Title alternatives. Let the AI produce 10 titles; but address exaggeration, clickbait — a headline that arouses curiosity but does not match the content — and claims that go beyond fact. The headline can't say what the body doesn't.
Step 5 — Objectivity and attribution check. Ask the AI to mark the attribution (who/what it is based on) of each factual claim in the text; Close the sentences with blank references.
Step 6 — The final reading is human. Proofreading for language, rhythm, fairness (all sides have a say), sensitive expressions and legal risk is the job of the journalist and editor.
Tip: When getting a draft from AI, always include the instruction "do not add any information that is not in the document/material, leave blank as missing [must be verified]". This single sentence keeps the hallucination from seeping into the manuscript.
Language, tone and adaptation
Same news; It can be presented in different formats as a short newsletter, a long analysis, a social media card and a newsletter introduction. AI is very useful at repackaging a text into different lengths and tones. Again the rule: facts remain constant, only the form changes; The title and spot are checked again with each new format.
three mini cases
Case 1 — The headboard extends beyond the body. A reporter asked the AI for a headline; YZ suggested the title "The Ministry withdrew the raise decision". However, the text said "considering withdrawal". The reporter caught the difference and corrected it. If it had been published, a certainty that was not stated in the body would have appeared in the title and a disclaimer would have been required.
Case 2 — Time savings, multiple formats. A reporter was manually transcribing a fact-checked story into three separate formats for web, newsletter, and social media (~50 minutes). Give the AI the verified text and have it produce three format drafts (~10 minutes); He checked the titles and spots of each format one by one. The time decreased to ~20 minutes, the facts did not change.
Case 3 — Added “context” trap. An intern asked the AI for a news draft; The AI added a "past event" context that wasn't in the material and made it look fluent. When the editor asked the source of this sentence, it turned out that it had no basis. If the instruction to "add what is not in the material" had been given from the beginning, this risk would not have arisen.
Copiable templates
1) Inverted pyramid skeleton:
Below are verified facts and quotes. The terspiramide skeleton for a news story emerges: spot (5W1H), main body, background, side views, conclusion. Mark the fact on which each sentence is based, like [O2]. Do not add any information that is NOT in the material; leave omissions [must be verified].Material: [facts + quotations]
2) Spot alternatives:
Write 4 different spots (first paragraphs) based on the following verified synopsis: one straight-news, one outcome-oriented, one human-story introduction, and one punch-line. Let none of it exceed the fact; Do not use exaggeration or adjectives. Summary: [here]
3) Title generation (canonical):
Suggest 10 headlines that fit the news body below. Rules: don't say what the body doesn't say, don't exaggerate, don't use clickbait, don't let the accuracy exceed the certainty in the body. Write under each heading the sentence it is based on. Body: [here]
4) Attribution and impartiality control:
In the news draft below: mark (1) factual sentences with ambiguous attribution, (2) places where fact and interpretation are confused, (3) sections where only one side speaks. Rewriting the text; Only one inspection report is produced. Draft: [here]
Weak prompt / Strong prompt
Weak prompt: "Write a story on: [topic]."
Result: AI fills the topic with imagination, adds non-existent quotations and numbers, and produces unattributed claims. An unverifiable "news-looking text" appears.
Powerful prompt: "Write an inverted pyramid news story based ONLY the following VERIFIED facts and quotes. Link each sentence to the fact on which it is based with [O#]. Do not add any information that is not in the material; leave missing parts as [must be verified]. Material: [here]"
Difference: Strong prompt traps the AI in authenticated material, making each sentence traceable and leaving the hallucination visible.
comparison chart
item
Contribution of AI
Journalist's supervision
Spot (led)
3-4 alternatives
5W1H accuracy, no exaggeration
Title
10 candidates
No overreach, no clickbait
body/skeleton
Inverted pyramid outline
Attribution, fact-interpretation distinction
Quote placement
Candidate locations
Literacy and context
Format adaptation
Web/newsletter/social
Recheck in every format
Common mistakes
- Making AI make up news about the subject. Giving "subject" instead of material; The result is an unattributed, fabricated text.
- The headboard extends beyond the body. Taking the hyperbolic title of AI without checking; The most common reason for denial.
- Not reconfirming the spot. Omitting a factual error in the spot, which is the most compressed place.
- Not noticing the added context. Dropping the flimsy sentences that AI added for fluency.
- Writing one side. Ignoring the principle of justice and conveying only one side's opinion.
Separation of fact and interpretation and neutrality of language
The finest skill of journalistic writing is to separate fact from interpretation. The fact is a verifiable fact (“the organization laid off 500 people”); comment is an assessment of it ("this is a purge of scandalous proportions"). In plain news, comment only enters by being attributed to a source ("the opposition called it a 'scandal'"); not in the journalist's own voice. AI drafts often confuse the two: adding value-laden adjectives (“scandalous,” “embarrassing,” “historical”) for the sake of fluency. These adjectives unknowingly distort the impartiality of the news. When reviewing a draft, underline each adjective and adverb and ask, “is this a fact or my interpretation?” It's a good habit to ask.
The neutrality of language is hidden in the frame beyond the choice of words. Calling the same event "hike" or "price regulation", "demonstrator" or "activist" directs the reader's perception. AI can repeat the dominant framework in the data on which it was trained; Therefore, it is necessary to consciously question the nomenclature used in a draft. The safest way is to cite the parties' own designations and for the journalist to choose as neutral, descriptive terms as possible in his own text. Neutrality is not passivity; It means giving every party the right to speak, separating fact from interpretation, and choosing language consciously.
One final subtlety is the difference between balance and false balance. It is right to give every party the right to speak; But it misleads the reader to present a claim that has been clearly refuted by evidence with equal weight to a confirmed fact, just to be balanced. For example, giving the same weight to an unfounded denial as opposed to an established fact with official data actually gives credibility to the mistake. AI drafts sometimes strike this false balance with a "let's give it both views" reflex. The journalist's job is to announce the sides and make it clear to the reader where the evidence stands; balance is fidelity to evidence, not equal distance from truth.
In summary
AI in news writing; It quickly produces the inverted pyramid framework, spot and title alternatives, format adaptations and language suggestions. But the facts come from verified material, the citations are appropriate, the title does not extend beyond the body, and the final reading is human. The instruction to "add what is not in the material, leave what is missing blank" should be part of every draft request. AI does not produce materials; edits verified material.
Application task
For an event you are verifying, first pull out your list of facts and quotes. You draft with the "Inverted pyramid skeleton" prompt, then you produce 10 headlines with the "Headline generation" prompt. Mark how many of the headings extend beyond the body and apply the "Citation and neutrality check" prompt to the manuscript and correct any omissions you find.
checklist
- [ ] I give AI verified material, not subject matter.
- [ ] I put the "insert what is not in the material" instruction into every draft request.
- [ ] I check that the title does not exceed the body and is not clickbait.
- [ ] I reaffirm the spot and quotes; I'm closing the references.
- [ ] I observe the principle of justice and give every party the right to speak; The final reading is mine.