Gains:
- Being able to distinguish where in the journalism workflow (transcription, first draft, scanning, translation) artificial intelligence saves real time and where decisions such as accuracy judgment and source protection are left to the journalist, depending on the task risk level
- Ability to apply a discipline that verifies each AI output by linking it to the primary source, confirming it with the secondary source, and manually checking the number.
- Ability to recognize the limits of hallucination and bias and acquire the anonymization and metadata cleaning reflex before giving sensitive documents and resources to artificial intelligence
The New Partner of Journalism: Roles, Limits and Verification Reflex of Artificial Intelligence
In a newsroom, time is the scarcest resource. The reporter returns from the field, carrying two hours of audio recording, a forty-page balance sheet, and a fifteen-tab browser history; The editor says, "I need a spot in half an hour." This pressure is the main reason why artificial intelligence (AI) — computer programs that can understand human language and produce output such as text, summaries, translations and classifications — is rapidly gaining ground in journalism. AI reduces repetitive and time-consuming tasks (transcription, first draft, scanning long document, translation) to minutes. But journalism is a profession where the margin of error is very small: a false name, a fabricated quote, an unsubstantiated claim, when published, can harm both a person's life and the publication's reputation. That's why the very first sentence of the module is: The AI does not decide for the reporter; drafts, speeds, scans—but it is always the journalist's job to verify the truth, protect the source, and decide what goes on air.
This first unit establishes three foundations: distinguishing where AI saves real time in the journalism workflow and where it should stop; reflex to verify every output; protection of source and sensitive data. These three are the underlying security basis for all subsequent units.
Where does AI accelerate, where does the decision belong to the journalist?
It helps to think of reporting tasks on a risk axis. Risk is how much damage it will cause to the public, the source and the publication if something is done wrong.
Low-risk tasks that AI speeds up a lot: extracting a raw transcript of an audio recording, scanning the main headlines of a long report, writing the first draft of a news story, generating ten alternatives for a headline, repackaging a text for different platforms, getting a rough translation of a foreign document, listing salient patterns in a large data table. Here AI solves the “blank page” and “stack sifting” problem; The journalist corrects, selects, enriches.
High-risk tasks where the decision is left to the journalist: Declaring a claim as "true", publishing a quote, making a statement accusing someone, judging the authenticity of a document, processing information containing the identity of a source, establishing the public interest-privacy balance. These decisions affect people's reputation and safety; The responsibility and final say always belongs to the journalist. Here, AI at most suggests a hypothesis or checklist, it cannot make a judgment.
Attention: "AI summarized like this" is not a news reason. The source of a claim, the truth of a quote, or the authenticity of a document cannot be published without the confirmation of a competent journalist. An unconfirmed AI output is like a single-source rumor — it may be true, but it needs a second source before it can go live.
Why is verification necessary? Hallucination and bias
AI produces text not by "knowing" but by "predicting the possible word". That's why sometimes it gives extremely fluent but actually incorrect information. This is called a hallucination: a court decision that doesn't exist, an unspoken quote, a false date, an academic paper that doesn't exist, a case number that appears real but is made up. This is fatal for journalism, because a fabricated "source" collapses the news and leads to a denial.
The second problem is bias: AI can repeat social patterns in the huge body of text on which it has been trained; can systematically frame a group, region, or opinion negatively or positively. If a reporter unknowingly brings this framework into the news, objectivity is damaged.
Because of these two limits, we propose a four-step verification discipline to be applied to each output:
- Go down to the source. Go back to the original document, record, or primary source for every fact, name, date, number, and quote that AI provides. AI is the epitome; It is not evidence.
- Look for a second source. The rule of reporting does not change with AI: the critical claim is confirmed by at least two independent sources.
- Check the number manually. Recalculate every numerical output such as percentage, total, ratio etc. from raw data.
- Take responsibility. The moment you approve the publication, that sentence is yours; The error is also in your signature.
Resource protection: privacy from the start
In journalism, source protection is the principle of keeping the identity of the informant confidential and a legal right in most countries; For a reporter, it is a matter of honor in the profession. Writing an informant's name, location, identity traces of the document he sent, or details that will allow him to be reached into a public AI tool instantly breaches this protection. Many free AI tools can store the text you enter or use it in model training; So you put the data out of your control.
The rule is simple: Anonymize first and extract sensitive data. The source's name, phone number, place of employment, and metadata in the document (author, date, location information embedded in a file) are cleared without going to the AI. Remove identifying lines before hashing a document; generalize people like "Source A", "a public official", etc. For sensitive files, if possible, choose offline or corporate tools with data processing assurance. We will deepen this principle in unit 10; Make it a reflex for now.
Tip: Before uploading a document to AI, clear the metadata and search the text for name, phone, address, institution. Even the smallest detail that would give away the source (a date, a description of a room) must be omitted.
three mini cases
Case 1 — Fabricated case record. A courthouse reporter asked YZ for a precedent summary. YZ gave it in convincing language along with three "Supreme Court decisions" numbers. The reporter ran the numbers through the official decision search system: two of the three were not available at all. If he had used it without checking, a news story would have appeared based on a fabricated precedent. The verification step ensured that it only used the only existing decision; The news came out 20 minutes late but solid.
Case 2 — Time savings in transcription. An investigation team would normally spend ~40 hours manually decoding 14 hours of meeting footage. Using AI-based transcription, the raw text was output in 2 hours; The team spent the remaining time correcting the 6 critical passages against the recording and establishing context. Total time went down from 40 hours to ~9 hours; Accuracy did not decrease because the journalist manually corrected the name and number errors in the raw transcript.
Case 3 — Resource-protecting reflex. A reporter, while trying to consult AI about the situation of a public employee who leaked a document to him, was about to paste the document as it was. He noticed without sending; cleaned the document's metadata, removed the signature and unit name, generalized the person as "an employee of an institution." His publishing-strategy ideas worked again; but no trace emerged to give away the source.
Copiable templates
1) Separating the task by risk level:
Your role: sub-editorial assistant to a journalist. Divide the following reporting tasks into two lists: (A) low-risk jobs where the AI can confidently produce an outline/summary, (B) high-risk jobs where the final decision must rest with the journalist. Write a one-sentence justification for each task. Tasks: [paste your own weekly tasks]
2) Hallucination hunt prompt:
List each fact, name, date, number, quote, and source item mentioned in the text below. Label each one "must be verified from primary source" and suggest which document/record it should be looked for. Do not say TRUE based on your own knowledge; It only infers what needs to be verified. Text: [here]
3) Anonymization pre-check:
Before publishing the text below, flag any elements that might give away the source: name, title, institution, location, date, recognizable detail. Suggest a safer generalization for each flag (e.g., "a public official"). Don't change the text, just report risks. Text: [here]
4) Editorial verification checklist:
You will evaluate a news draft. List the deficiencies under the following headings: (1) single-source claims, (2) unverified numbers, (3) quotes with unclear attribution, (4) places where the parties were not given the right to speak, (5) sentences where inference and fact are mixed. Just produce a checklist, do not rewrite the news. Draft: [here]
Weak prompt / Strong prompt
Weak prompt: "Write a news story from this document."
Result: AI injects holes, produces unattributed claims, adds context that is not in the document. A journalist cannot distinguish between what is a document and what is a fabrication.
Powerful prompt: "Make the skeleton of a news story based ONLY on the following document. Next to each sentence, mark which paragraph in the document it is based on, like [P3]. Do not add any information that is not in the document; leave the missing parts blank as '[must be verified]'. Document: [here]"
The difference: The powerful prompt links the AI to the document, making each claim traceable and making the hallucination visible. The journalist knows which sentence to confirm.
comparison chart
Quest
Role of AI
Decision/responsibility
Mandatory verification
Voice recording transcription
Generates raw text
Critical passages are in the journalist
Compare name/number with record
Long document scanning
Creates title and summary
What's news is up to the journalist
Go back to original paragraph
First draft writing
Suggests skeleton and language
Citations and facts are with the journalist
Link every claim to the source
Quote selection
Nominative sentences signs
Publication by journalist
Verbatim confirmation from the record
Source/document processing
(limited) assistance
Identity protection for journalists
Metadata and name cleaning
Common mistakes
- Mistaking the AI output for a source. AI is a draft generator, not a resource; Publishing the fact without confirming it from the primary document leads to denial.
- Pasting the sensitive document as is. No document should enter a public tool without purging metadata and names that give away the source.
- Bypassing the "two resources" rule with AI. The compendium of AI is a single source; The critical claim again requires two independent confirmations.
- Trusting the numbers. AI frequently miscalculates percentages and totals; retest each number from the raw data.
- Not recognizing bias. Taking the framework of AI without questioning and transferring it to the language of the news disrupts objectivity.
In summary
In this unit we position AI as an assistant in reporting: it speeds up the repetitive and mass-sifting tasks, but it is up to the journalist to decide the truth, publish the quote and protect the source. Hallucination and bias are the two main limits; Against these, the four-step verification discipline (get to the source, second source, check the number, assume responsibility) and the privacy reflex from the beginning (anonymize, clean metadata) are applied to every output. The next units will be installed on this ground.
Application task
Write down 8 reporting tasks you did in your last week. Label each as “low risk — AI makes it faster” or “high risk — it's up to me” and add a one-sentence justification. Then choose one and apply the “Hallucination hunt” prompt above to an actual text; Note how many claims the AI flags actually need to be verified.
checklist
- [ ] I can separate the task by risk level; I know jobs where the decision is left to the journalist.
- [ ] I attribute each AI output to the primary source; I don't bypass the "two resources" rule with AI.
- [ ] I know the risks of hallucination and bias, and I crunch the numbers by hand.
- [ ] I clean up any traces of metadata and identity before giving the sensitive document/resource to the AI.
- [ ] I have internalized that the "AI said" justification is invalid for broadcast.