Gains:
- Ability to separate a claim into independently verifiable components and identify and cross-check a primary source for each
- Ability to produce and apply visual and reverse image search, metadata and internal consistency testing checklists with artificial intelligence support
- Understanding that artificial intelligence cannot be judged on accuracy, that the evidence comes from the primary source, and that the confirmation process must be written transparently.
In journalism, no sentence is true until it is "shown to be true." Fact-checking—the act of testing the truth of a claim, image, number, or document with independent evidence—is the backbone of the profession. A reporter in the flow of information accelerated by social media; He has to test within minutes whether a photo was actually taken at the claimed location, whether an account is real or not, and the source of a number. Artificial intelligence (AI) is an organization and ideation tool in this process: it enumerates verification steps, breaks a claim down into its components, deduces which questions to ask, lists all claims in a text that need to be checked. But here the most critical warning becomes even harsher: AI is not a source of fact-checking; You can't find out if a claim is true by "asking" the AI. AI streamlines the process, the journalist finds the evidence from the primary source. Ask the AI "is this true?" To ask is to invite a hallucination.
Step by step: confirming a claim
Step 1 — Break the claim into its components. The sentence "That institution laid off 500 people last year" is actually several claims: which institution, which year, how many people, what the source. AI does this parsing quickly; Each component is verified separately.
Step 2 — Identify the primary source. For each component, “where is the most reliable evidence?” Ask the question: official record, original document, first-hand witness, auditor's report. AI can suggest where to look; But you find the evidence and open it.
Step 3 — Test the image and location. For a photo/video: date first published, location shot, weather/shade consistency, signs in the background. Reverse image search — a method of finding where an image has previously appeared on the Internet — shows whether an old image is presented as new. The AI lists these checks; observation and search make the decision.
Step 4 — Check the identity and history of the resource. Opening date, past posts, consistency of a social media account. AI reminds us what to look for; You do the verification.
Step 5 — Cross-confirmation. The critical claim is verified by at least two independent sources. Two sites that copy each other are not "two sources"; Independence is essential.
Step 6 — Rate the result and write transparently. Provide a clear verdict and the evidence leading to it, such as "true," "misleading," "partially true," "insufficient evidence." The reader should see how you verify it.
Attention: Ask the AI "is this claim true?" Asking " is the most dangerous confirmation mistake. AI can produce a convincing “yes/no” or even a made-up “source.” Use AI as a process orchestrator; Primary evidence gives the verdict of truth.
AI-generated content and upside
AI has also become the subject of verification: texts, sounds and images can now be artificially produced (the subject of the next unit). So the fact-checker asks “could this content be an AI product?” He adds the question to his process. AI detection tools provide guidance but are not conclusive evidence; The real verification again comes down to the source, metadata and context.
three mini cases
Case 1 — Old image, new claim. A reporter received a photo of a fire that was said to have been "taken today in that city." He had the AI produce the verification steps; The first item was "reverse image search". The search showed that the same photo was published 3 years ago in another country. The visual did not make it to air; The reporter turned this into a fact-check.
Case 2 — Speed of parsing the claim. During one election season, a reporter had to fact-check 12 numerical claims in a 1-minute conversation. YZ divided the conversation into 12 separate claims and created a list of "which official source should be looked for" for each. The reporter went to 12 sources; 4 demonstrated with independent evidence that the claims were misleading. Having the AI do the parsing instead of doing it manually saved about an hour, but the reporter opened every piece of evidence.
Case 3 — Fabricated source. To verify a claim, an intern asks the AI directly, “is this true, what is the source?” he asked. The AI said “yes” and provided a report name and link. The editor opened the link: there was no such report. This turned into a training case that showed the team why the “Don't ask AI for accuracy” rule is absolute.
Copiable templates
1) Separating the claim into components:
I will confirm the following sentence. Break it down into independently verifiable claims. For each claim: (1) what information is to be verified, (2) what type of primary source would be the most reliable, (3) at what point it would be misleading. Don't say TRUE/FALSE yourself. Sentence: [here]
2) Visual/video verification checklist:
A photo claims to have been "taken at such and such a place on such and such a date." List the verification steps I would need to do to test this claim: reverse image search, location clues (signage, language, architecture), shadow/air consistency, first publication date. Write how to follow each step.Image what I know about: [recipe]
3) Removing all claims from the text:
List EVERY factual claim in the text below that requires corroboration (numbers, dates, attributions, quotes, comparisons). Suggest "what independent source should be verified" for each. Separate interpretation and fact. Change text. Text: [here]
4) Independence test:
I found 3 sources for the following claim: [summarize sources]. What questions should I ask to evaluate whether these are truly INDEPENDENT or duplicate/based on a single original source? Decisive decision-making; control questions appear.
Weak prompt / Strong prompt
Weak prompt: "Is this claim true? Give your source."
Result: AI produces a convincing verdict and often a fabricated source. The journalist falls into false confidence.
Powerful prompt: "Break this claim into independently verifiable components. For each component, write down which PRIMARY source type I should look for and, if misleading, at what point it might be. You DON'T judge the claim; I will verify it from the source."
Difference: Strong prompt reduces the AI to a process editor, prevents it from producing fabricated "source/judgment", and keeps the responsibility of evidence on the journalist.
comparison chart
Claim type
Contribution of AI
primary evidence
trap
numeric data
Parses, recommends sources
Official record/report
% of AI making up
Quote/attribution
Divides into components
Original record/document
taken out of context
Image/video
checklist
Reverse search, metadata
Old visual-new claim
social media account
Lists what to look for
Account history, consistency
Imitation/bot account
document
Marks suspicious spots
Source, metadata, signature
Fake/fictional document
Common mistakes
- Asking AI about accuracy. "Is this true?" The question produces fabricated judgments and sources; AI does not provide evidence.
- Copy resources count as "two resources". Sites that transfer each other are not independent; It violates the principle of cross-confirmation.
- Do not search the image in reverse. Presenting an old image as new is the most common disinformation.
- Omitting context. Misleading by giving an accurate quote in the wrong context is also a fact-checking error.
- Not being transparent. A confirmation that does not show the reader how it is confirmed does not inspire trust.
Rating and transparent language of the confirmation
Confirmation often occurs in the gray area between “true” and “false”; A good fact-checker translates this gray into clear categories. A common scale is: True (claim supported by primary evidence), Partially true (partly true but missing or distorted), Misleading (creating a false impression with a technically correct element), False (disproved by evidence), and Insufficient evidence (can neither be confirmed nor refuted). The last category is critical: just because you can't refute a claim doesn't make it true; Saying "evidence is insufficient" is also honest fact-checking and is better than selling false certainty to the reader.
In addition to the rating, the fact that the confirmation news shows the evidence trail is the basis of trust. The reader should be able to see which document is being looked at, which source is speaking, what the reverse search reveals; It is not enough to say "we investigated and it turned out to be wrong". This is where AI can help you turn your fact-checking process into an evidence grid: each claim, the evidence on which it is based, and the rating given, all together. Such a table both makes the editor's job easier in internal editing and allows the reader to trust the confirmation when published. Remember: the purpose of fact-checking is not just to catch falsehood, but to make accurate information visible and traceable.
In summary
Verification is breaking down the claim into its components, finding primary evidence for each, testing the image and the source, conducting independent cross-checking, and writing the result transparently. In the process, AI parses the claim, creates checklists, suggests where to look — but never makes truth judgments or produces sources. Hardest rule: Ask the AI "is this true?" It is not asked; The evidence comes from the primary source, the hand of the journalist.
Application task
Choose a claim or image that is circulating on social media. Follow the “Break the claim into components” and (if visual) “Visual verification checklist” prompts. Search for each component the AI extracts in the actual primary source; Rate the result as "true / misleading / partially / insufficient evidence" and write how you verified it in 5 items.
checklist
- [ ] I break the claim into independent components and verify each separately.
- [ ] I'm not asking the AI to judge truth; I find the evidence from primary source.
- [ ] I test images with reverse search and location clues.
- [ ] I confirm the critical claim with at least two truly independent sources.
- [ ] I show transparently to the reader how I verify.