Gains:
- Ability to use artificial intelligence in limited and material-dependent tasks in the eight stages of a news story (clue, data, decipherment, confirmation, writing, packaging, ethics/source, publication).
- At the output of each stage, a human verification gate operates and asks 'what is unverified in this output?' Being able to prevent the hallucination from growing with the question of transition
- Ability to understand that the chain of responsibility consists of people, that post-release errors are resolved with a transparent correction, and that the process is transformed into a repeatable checklist.
End-to-End Workflow: The Complete AI-Powered Journey of a News
In the previous ten units, we learned the pieces individually: the verification reflex, data analysis, transcription, fact-checking, disinformation defense, writing, summarization, translation, ethics, and source protection. In this final unit we combine the parts into a single workflow. The goal is to produce real news from start to finish — from tip to publication to editing — using AI in the right places, with the right boundaries. In a newsroom, this flow operates within an editorial process—the chain in which news moves from reporter to editor to fact-checking desk to publication. The integrative principle is this: AI accelerates at each stage, but there is a human verification gate at the exit of each stage; No output goes to the next stage without being validated. This unit turns what you've learned into a checklist and repeatable process.
The eight stages of a news story and the place of AI
Stage 1 — Hint and preliminary research. AI sketches the background of a topic, generates questions to ask, describes relevant public data. Door: The journalist decides which claim is truly worth investigating.
Phase 2 — Document and data collection/analysis. AI scans document stack, extracts exploratory questions, describes analysis method (Unit 2). Gate: Each number is re-polled from raw data; The calculation is not made by AI.
Stage 3 — Interview and transcription. AI transcribes the recording, extracting quote candidates with timestamp (Unit 3). Door: Each quote is verified verbatim from the record.
Stage 4 — Confirmation. AI breaks down claims into their components and produces checklists (Unit 4-5). Door: Judgment of accuracy is made from the primary source, with two independent confirmations; AI does not make judgments.
Stage 5 — Writing. AI produces inverted pyramid skeleton, spot and cap alternatives (Unit 6). Door: What is not included in the material is not added; the title does not exceed the body; references are closed.
Stage 6 — Summary/translation/packaging. AI produces drafts for different formats and languages (Unit 7-8). Gate: Context is preserved; quote translation is human confirmed; Number/title is checked in each format.
Stage 7 — Ethics and resource protection audit. AI conducts ethical risk and source-compromise screening (Units 9-10). Door: Added transparency statement; synthetic ingredient is labeled; weld traces are cleaned.
Stage 8 — Publication, correction, archive. If an error occurs after publication, a quick and transparent correction is made. Door: Responsibility lies with the person; There is no "the vehicle did it" excuse.
Tip: Concrete each stage gate with a “transition question”: “What is in this output that is unverified, and how do I confirm it before moving it to the next stage?” This single question prevents the hallucination from growing along the chain.
Role distribution in editing
The introduction of AI does not change the distribution of roles, it strengthens them: the reporter collects and verifies the material, the fact-checking desk cross-checks, the editor observes integrity and ethics, and editorial management assumes responsibility. AI adds speed to each of these roles; but it does not replace any of them. The chain of responsibility consists of people.
three mini cases
Case 1 — Solid, end-to-end coverage. A team ran a public procurement investigation with this flow: AI crawled 3,000 lines of data (analysis verified by hand), transcribed 6 hours of interviews (quotes confirmed from the recording), expedited the manuscript (citations closed), conducted ethics and source scanning. Normally a 3 week job was done in ~9 days; no issue or quote was released without verification. The news received no denial.
Case 2 — The door caught a bug. During the writing phase, YZ added a "past penalty" information to the draft that was not in the material. The “seize what is not in the material” inspection at the Stage 5 gate caught this. If there were no door, a sentence that was unsupported and at risk of slander would be published.
Case 3 — Correction culture. After publication, a reader reported that a percentage was incorrect. The team probed the data again, saw the error (a conversion error), issued a transparent fix in 20 minutes, and added a “second-person verification of digital output” step to their flow. Accountability preserved trust.
Copiable templates
1) Stage gate access control:
I'm currently at [stage name] and the output is: [output]. List each element I need to verify before moving on to the next step: unconfirmed fact/number/quote, added information, source risk, loss of context. Write one sentence "how do I confirm" for each. It's up to me.
2) End-to-end news checklist:
I will give you a news draft ready for publication. Mark the following 8 headings as missing/risky: (1) unverified numbers, (2) unverified quotes from the record, (3) single source claims, (4) title exceeding the body, (5) unattributed sentences, (6) loss of context, (7) risk of giving away the source, (8) lack of ethics/transparency. Just produce inspection report. Draft: [here]
3) Editorial note:
I am transferring this news to the confirmation desk. Write a brief draft of the memorandum that includes: which facts have been verified and how, which still[need to be verified], which quotes have been confirmed from the record, precautions taken to protect the source. Don't add information I didn't provide. Status: [here]
4) Post-publication correction draft:
The following error appeared in the published news: [error]. Draft a correction that is honest, clear, and nondefensive to the reader; State what was wrong, what was correct, and when it was corrected. Don't blame the vehicle; Use responsible language.
Weak prompt / Strong prompt
Weak prompt: "Produce a complete news story on this topic."
Result: AI fills all stages with imagination; there are no verification gates, the hallucination grows along the chain, the output cannot be published.
Powerful prompt: Split the flow into pieces: run the AI on a limited, material-dependent task at each stage and put human control at the stage gate. "At this stage, just do [your] task; do not add what is not in the material; mark what needs to be verified [must be verified] before moving on to the next stage."
Difference: The strong approach limits the AI in stages, puts a verification gate at each output, and prevents hallucination from accumulating.
comparison chart
Stage
AI's mission
verification gate
Tip/preliminary research
background, questions
Newsworthiness decision
Data analysis
screening, method
Check the number from raw data
decipher
Text, citation candidate
Confirm the quote from the record
Confirmation
Parse, list
Two independent primary sources
spelling
skeleton, hood
No extra material, full attribution
Summary/translation
Format, language draft
Context + human translation confirmation
Ethics/source
risk screening
Declaration, label, trace cleaning
Publication/correction
Revision draft
Responsibility lies with people
Common mistakes
- Bypassing stage gates. Moving an output from validation to the next stage; the error grows along the chain.
- Letting the AI do all the flow. End-to-end "produce news" means; It produces a mass of unverifiable text.
- Losing context during handover. Delegating without specifying what is/has not been verified.
- Delaying correction or being defensive. Undermining accountability and losing trust.
- Not documenting the process. Making mistakes from the beginning on every news story without a repeatable checklist.
Making the process stick: checklist and team culture
Producing a news with this flow once is success; but the real value is in turning the flow into a repeatable system. Experienced newsrooms add each lesson they learn to a checklist: when they receive a rebuttal, they write down the missing step that led to that mistake as a permanent item on the list. Thus, the information does not remain in the head of a single reporter, but passes into the collective memory of the institution. AI can help draft these checklists and workflow templates; But it is the team that keeps the list alive, refers to it in every news and updates it.
A second element of permanence is team culture. AI-powered streaming only works safely in a culture where everyone shares the same verification discipline. An intern must apply the "AI cannot be asked for accuracy" principle, a senior editor must apply the "synthetic image is tagged" rule, and a fact-checking desk must apply the "two independent sources" requirement with the same seriousness. This culture is established not by punishing mistakes, but by an environment where reporting mistakes early and honestly is rewarded; Because while a concealed mistake goes to publication, a shared suspicion catches it at the door. AI tools will change rapidly, new capabilities and new risks will come; But the principle that "every output passes through a human door and the responsibility lies with the human being" will remain the unchanging core of journalism, no matter what tool is used.
In summary
End-to-end AI-powered journalism is an eight-step chain: tip, data, transcription, fact-checking, writing, packaging, ethics/source control, publication/editing. AI accelerates at every stage; But there is a human verification gate at the exit of each stage, and no output goes to the next stage without being verified. “What is unverified in this output?” The transition question prevents the hallucination from growing. The chain of responsibility consists of people; Post-publication, the error is fixed with a transparent and accountable correction. Turning this flow into a checklist ensures both speed and reliability.
Application task
Choose a real or fictional news topic and write the eight stages in a table. For each phase: (1) determine what limited task you will use the AI for, (2) determine the transition question at that phase's verification gate. Then apply the “End-to-end news checklist” prompt to a draft and actually address the resulting risks.
checklist
- [ ] I operate a human verification gate at the exit of each stage.
- [ ] I give the AI limited and material-dependent tasks at each stage, not the entire flow.
- [ ] "What is unverified in this output?" I ask the question of transition in every era.
- [ ] I clearly state what is/has not been verified during proofreading cycles.
- [ ] I fix the post-release bug with a transparent, accountable fix and document the process.
Module Exam
1. A reporter puts a 'Supreme Court decision number' given by artificial intelligence into the news without checking it, and the number belongs to a decision that does not actually exist. What is the main lesson of this situation?
- A) Publishing the artificial intelligence output without verifying it from the primary source; Ignoring the risk of hallucinations and that the responsibility lies with the journalist ✔
- B) It is strictly forbidden to use artificial intelligence in courthouse news
- C) Supreme Court decisions are never included in the news
- D) The decision is not presented in a table
Description: Artificial intelligence is fluent but can actually produce false information (fabricated source, fake number); this is called hallucination. Not every fact, name, number and source included in the news can be published without verification from the primary source; The responsibility lies with the journalist.
2. In data journalism, what is the safest approach to calculating percentages and averages in a table?
- A) Let artificial intelligence calculate the number directly and put the result in the news
- B) Have the artificial intelligence describe the method/formula, run it in an auditable tool, and verify the result from raw data ✔
- C) Do not use numbers at all and only write comments
- D) Counting the first guess given by artificial intelligence as two sources
Description: Artificial intelligence is a language model, not a calculator; Frequently calculates averages and percentages incorrectly. The correct method is not to have artificial intelligence do the calculation, but to have the method described, run it in an auditable tool (spreadsheet, query) and check the result from the raw data.
3. Which step is mandatory for a quote to be included in the news in an interview text extracted with automatic transcription (ASR)?
- A) Having the quote corrected by artificial intelligence to make it more fluid
- B) Assuming the sentence in the transcript is correct and publishing it directly
- C) Verify the quote by comparing it verbatim with the record from the timestamp ✔
- D) Relying only on speaker discrimination and not listening to the source
Description: Automatic transcription is a raw draft; Especially in proper names, numbers and jargon, a single word can change and reverse the meaning. Every quoted sentence in the news should be verified by going to the time stamp and listening to the recording verbatim.
4. To confirm a claim, an intern can directly ask the AI 'is this true, what is its source?' he asks. What is the main problem with this approach?
- A) The question is too long
- B) The claim is not based on a visual
- C) Artificial intelligence cannot be used for confirmation at all
- D) Artificial intelligence's ability to produce truth judgments and fabricated sources; ✔ evidence must come from a primary source
Disclosure: Artificial intelligence is not a source of fact-checking; 'is this true?' may give a convincing but wrong conclusion or even a fabricated source to the question. In verification, artificial intelligence streamlines the process (breaks down the claim into components, produces a checklist); Evidence of accuracy comes from the primary source, the hand of the journalist.
5. What is the most effective first step to test the authenticity of a photo of an event that is said on the Internet as 'taken today in that city'?
- A) Finding where/when the image was previously published by doing a reverse image search ✔
- B) Ask artificial intelligence 'is this photo real?' asking and trusting the answer
- C) Looking at the aesthetic quality of the photo
- D) Looking at the number of followers of the sharing account
Explanation: The most common visual disinformation is presenting an old image as a new event. Reverse image search often single-handedly reveals such fakes by showing where and when that image was previously published.
6. An artificial intelligence 'fake detection tool' says an image is '78% artificial'. How should this output be used in the editorial decision?
- A) If the percentage is high, directly declare the image as fake and withdraw it from publication.
- B) Take it as a clue and do the actual verification by going back to the source and reverse search ✔
- C) If the percentage is low, publish the image without checking it at all
- D) Writing the output of the tool into the news as a justification for publication
Explanation: Detection tools can be wrong in both directions: they can call the real fake and the fake real. This output is not definitive proof, but a clue. Provision; provided by source extraction, metadata, reverse search, and independent context verification.
7. While artificial intelligence produces a suitable headline for a news body, it titles a decision that is 'considering withdrawing' as 'withdrawn'. What is this a violation of?
- A) The title exceeds the precision in the body; The title can't say what the body doesn't ✔
- B) The title is too short
- C) No number in the title
- D) The title is not in question form
Explanation: The headline cannot say more than the body of the news does; A title that exceeds the precision of the body misleads the reader and creates the risk of refutation. The titles produced by artificial intelligence must be inspected by comparing them with the body.
8. While summarizing a long court decision, artificial intelligence says 'the defendant was acquitted'; However, the decision is 'due to lack of evidence, the possibility of appeal is open' and another lawsuit is ongoing. What is this typical mistake?
- A) The summary is too long
- B) The summary is not put in a table
- C) Loss of context; A correct sentence becomes misleading by dropping condition and process information ✔
- D) The decision must be in a foreign language
Explanation: Summaries are most misleading by 'decontextualizing': a sentence is technically correct but creates the wrong perception because the condition, time or source condition is omitted. Circumstances surrounding critical provisions should not be omitted from the summary and information should be verified from the original.
9. What is the correct attitude towards a foreign quote extracted through machine translation in multilingual journalism?
- A) Publishing the machine translation as is, in quotes
- B) Just checking the length of the quote
- C) Translating the idioms verbatim and using them exactly
- D) Consider the translation as a draft, confirm it with the original with a person who knows that language, and keep the original on record ✔
Explanation: Machine translation may drop a negative (not/none), a conditional or irony and mistranslate phrases; This can make a quote mean the exact opposite. The quote to be included in the news is confirmed by comparing it with the original by a person who knows that language, and the original sentence is kept in the record.
10. What is the ethically correct way to use a 'crime scene' image produced by artificial intelligence in a news story?
- A) Using the image without tags as if it were a real photo
- B) Using the image clearly labeling it as 'representative, produced with artificial intelligence' and in a way that does not imitate reality ✔
- C) Putting the image without any tags but in a small size
- D) Substituting a visual of the real event without asking the reader
Explanation: A synthetic image presented as representing a real event amounts to disinformation. AI-generated imagery can only be used if it is clearly labeled (representational, produced by AI) and does not mimic reality; Synthetic media that give the impression of reality are not included in the news.
11. A sentence describing a source as 'the only disabled employee at the head office' is used when consulting artificial intelligence. Why is this risky?
- A) The sentence is too long
- B) The sentence is not taken from an official document
- C) Indirect identifier refers to a single person in a small group, revealing the identity as much as the name ✔
- D) Artificial intelligence cannot understand the issue of disability
Explanation: Resource protection is not just hiding the name; Indirect identifiers that refer to a single person in a small group are as revealing of identity as a name. 'How many people will this recipe suit?' If the test is one, the resource is not protected; Additionally, giving this data to a public tool causes loss of control.
12. What is the critical step before having AI summarize a sensitive leak document?
- A) Save time by quickly uploading the document as it is
- B) Scanning the document in color
- C) Just change the title of the document
- D) Cleaning the metadata of the document and converting it to plain text and anonymizing identifiers ✔
Description: Metadata embedded within a file (author name, edit history, creation date, location) can directly identify the source. Before the document is exported to a tool, metadata must be cleaned, converted to plain text, and all direct/indirect identifiers must be anonymized.
13. An incorrect number originating from artificial intelligence appears in a published news. What does the principle of accountability require?
- A) Take responsibility and issue a transparent and non-defensive correction ✔
- B) Attributing the mistake to artificial intelligence by saying 'the vehicle made it'
- C) Silently deleting the error and not making any explanation
- D) Change the vehicle and leave the news as it is.
Explanation: Liability cannot be transferred to the vehicle; 'The artificial intelligence summarized/calculated that way' is not a valid defence. A person stands behind everything published; The error is corrected through a transparent and accountable correction, without placing the blame on the vehicle.
14. What does the concept of 'stage gate' mean in the end-to-end AI-supported journalism flow?
- A) Artificial intelligence automatically switches between stages
- B) A human verifies the output at the exit of each stage; Do not proceed to the next stage without verification ✔
- C) Delivering the news to publication as quickly as possible
- D) Delegating the entire flow to a single artificial intelligence request
Description: Stage gate is a human verification check at the output of each stage: no output goes to the next stage without being verified. 'What is unverified in this output?' The transition question prevents the hallucination from growing along the chain.