Gains:
- Ability to explore a data set with artificial intelligence, generate newsworthy questions and extract sensitive data
- Instead of having artificial intelligence do the calculations, having the method described and run in an auditable tool and gaining the habit of verifying each finding from raw data.
- Understanding that it is the responsibility of the journalist to draft outliers and document archive relationships and confirm them in the original source.
Investigative journalism often starts with a pile: a tender spreadsheet of thousands of lines, a balance sheet of hundreds of pages, a leaked email archive, an open-source public expenditure set. Data journalism — the practice of finding patterns, anomalies, and relationships in numerical and documentary data and turning them into news — is exactly the job of making sense of this mass. In these piles, which would take a human lifetime to look through manually, artificial intelligence (AI) is a powerful first-screening and idea-generating tool: it can summarize a table, extract duplicate names in a text archive, suggest what questions to ask for an analysis, even describe data cleaning steps. But let's put one sentence from the beginning: AI is a partner in analyzing data; It is the journalist who decides the accuracy and news value of the result and re-examines the number from the raw data. The risk of hallucination is at its most dangerous here, in numbers.
Step by step: exploring a data set with AI
Step 1 — Get to know the data. First understand the columns, number of rows, date range, units. Know for yourself what it is before you load the raw file into the AI; Otherwise, AI's "findings" remain in the air.
Step 2 — Extract sensitive data. If it contains personal data (name, TR ID, address, health), anonymize it without giving it to the public AI tool or send only the columns to be analyzed. Traces of leaked documents that reveal the source are also cleared (Units 1 and 10).
Step 3 — Generate discovery questions with AI. “What news could this data set be hiding?” Start by saying. The AI comes up with a list of questions to ask: top 10 items, repeat suppliers, unusual jumps, areas left blank.
Step 4 — DON'T HAVE AI DO THE ANALYSIS, DESCRIBE IT. This is the critical point. The average or percentage that the AI calculates in head is often wrong. Instead, ask the AI for steps that you would run in a reliable tool (spreadsheet formula, query, script). Thus making calculus an auditable tool, AI just gives the method.
Step 5 — Verify the finding. See every anomaly that the AI points out by going back to the raw line. Filter the table and count the claim "This company won 31 out of 40 tenders". Any unverified numbers will not make it into the news.
Step 6 — Establish the context. The number alone is not news. Comparison (last year, similar institutions, ratio to population), context and expert opinion are the job of the journalist.
Attention: Having the AI say "calculate the average of this table" and putting the result directly into the news is the most common disclaimer-producing mistake of data journalism. AI is a language model, not a calculator. Have the tool do the math, just ask the AI for the method and interpretation.
Metadata and document analysis
Besides numerical tables, investigative journalism also deals with large text archives: emails, minutes, contracts. Here, AI can draft relationship maps such as "who talked to whom when", "which topic is repeated", "which names are mentioned together". Again, the rule is the same: AI gives an initial map; Every link is confirmed in the original document, because AI can convincingly explain a link that does not exist.
three mini cases
Case 1 — Concealed condensation. A reporter had a public procurement spreadsheet with 2,400 rows. He had the AI produce exploratory questions; The question "how many tenders did the same supplier receive?" came to the fore. The reporter didn't have the AI calculate this; He ran the spreadsheet formula suggested by YZ himself and found that a single firm received 380 (15.8%) of 2,400 items. He checked it manually and the number was correct. The AI's initial "estimate" said 22% — so if the AI were to be trusted, the news would be wrong.
Case 2 — Time saver, email archive. It would take days to manually search "what topics are going on" in an archive of 6,000 emails. The AI produced an outline of topic clusters in 15 minutes; From these, the reporter selected 3 promising clusters and read the original emails. Total discovery time was reduced to ~3 days, but every quote that went live was verified verbatim from the original email.
Case 3 — Hallucination caught. An economic reporter received a summary from YZ that "personnel expenses increased by 60% in one year" from a balance sheet. When he returned to the raw table, he saw that the increase was 6% and that the AI had misread one digit. A single fact check blocked a false and pretentious headline like "60% explosion".
Copiable templates
1) Data set recognition and discovery questions:
I will give you the column headings and the first 20 rows of a data set. Generate 10 journalistic questions that CAN be done with this data: in terms of concentration, outlier, jump in time, missing/empty areas, recurring actors. No number calculation; just write which questions are worth asking and why they might make news.Data: [headlines + sample lines]
2) Not making calculations, but making them describe:
I want to do the following analysis: [e.g. find the total tender amount and share percentage of each supplier]. I want to run this myself in a spreadsheet. Write me step by step which formulas/pivot settings to install. Result of SEN calculation; just give the method.My columns: [column names]
3) Outlier hunting:
I will give summary statistics (min, max, average, median) below. Explain which values are unusual/suspicious and why I should check them for news. Do not make final judgments; A checklist appears in the format "the following lines should be checked manually". Summaries: [here]
4) Document archive relationship map (draft):
I will give you a set of document text. Output the following as a DRAFT: names that occur frequently together, recurring topics, notable dates. Mark the document from which each link comes, like [B12]. Do not add any relationships that are not explicitly written in the document. Documentation: [here]
Weak prompt / Strong prompt
Weak prompt: "Analyze this chart and state any significant findings."
Result: AI makes up percentages and averages head on, imagines anomalies, it's unclear which line it's based on. It cannot be verified.
Powerful prompt: "I give the columns and first 20 rows of this table. (1) List which analyzes might be newsworthy. (2) Suggest a spreadsheet formula for each analysis so I can run it. (3) Do not calculate any results. (4) Mark the rows to be checked, like [S45]."
Difference: Strong prompt leaves the calculation to the controllable tool, reducing AI to the role of method and direction; prevents hallucination from leaking into the number.
comparison chart
Stage
AI does
Journalist does
verification
Recognizing data
Column/pattern summary
Knows the unit and context
Source data dictionary
discovery questions
Generates a list of questions
Selects news value
—
calculation
Describes the method
Runs the formula in the vehicle
Manual provisioning
outlier
marks the candidates
Looks at raw line
Filter and count
Comment/context
draft frame
Comparison, expert
second source
Common mistakes
- Making AI calculate. Having AI calculate average, percentage, total etc. and use the result; AI makes mistakes frequently, let the vehicle do the math.
- Not connecting the finding to the raw line. Writing the claim "This company won most of the tenders" without filtering the table and counting.
- Publishing numbers without context. Reporting bare numbers without comparison and ratio is misleading.
- Uploading sensitive data as is. Giving personal data or a document revealing the source to the vehicle without cleaning it.
- Thinking that the fake relationship is true. Processing the link that the AI "sees" in the archive into the map without confirming it in the original document.
In summary
In data journalism, AI is a discovery and insight partner: it scans the pile, generates questions to ask, describes the method of analysis, flags candidates for outliers. But having the AI calculate the number itself is the riskiest mistake; Outsource the calculation to an auditable tool, verify each finding by going back to the raw data, and add context and comparison to the number. In document archives, AI gives a starting map; Each link is confirmed in the original document.
Application task
Take a data set you have (or is publicly available). First, have them generate 10 questions with the "Exploration questions" prompt. Select a question, get the formula from the AI with the "Describe the calculation" prompt, and run it yourself. Then ask the AI directly for the same calculation and compare the two results; Note the difference and its lesson.
checklist
- [ ] I understood the columns, units and sensitive fields before giving the data to the tool.
- [ ] I don't let the AI do the calculations; I describe the method and run it in the auditable tool.
- [ ] I manually verify each finding by going back to the raw row.
- [ ] I add comparison and context to the number; I'm not reporting the bare numbers.
- [ ] I confirm every relationship in the archive in the original document.