When you check a news claim, first write the sentence you want to verify on its own, then find where it first came from and what the original material actually proves. When you cannot find enough evidence, keeping “cannot confirm at present” is more reliable than writing a guess as fact.
The library story, records, and numbers below are entirely fictional material written to demonstrate the method. They do not correspond to any real institution or event. They have no real “original announcement” URL and should not be forwarded as news.
Step 1: Break one sentence into claims you can check
Suppose a screenshot arrives in your group chat with the headline: “A certain library doubled its weekend loans, proving the whole city reads more now.”
That sentence contains at least four different questions: which library, which weekend, whether loans actually doubled, and whether that change can say anything about reading habits across the city. The first few can be checked against records; the last one needs broader and longer-term evidence.
Write down the claims you actually need to check. If you only want to confirm whether the event happened, you do not also need to research city-wide reading trends. If you plan to quote it in an article, check it item by item.
Numbers and motives should be separated too. An institution may want to publicize its own event, but “has a promotional motive” does not directly imply the data is fake. The same distinction is explained in the fact-checking exercise inside the incentive analysis article.
Step 2: Follow the source back to the original material
For this teaching example, two pieces of fictional material are supplied: A is the forwarded screenshot; B is a supposed event record reading “this library lent out 240 volumes on event day, and the single-day record for the same weekday the previous week was 120 volumes.” The record gives no participant count, no data for other libraries in the city, and no actual reading completion.
With a real story, you can use a distinctive phrase from the headline, the organization name, and the date to find the original text, then check the body, attachments, statistical tables, and any correction notices. Record the publishing organization and the author, not just the avatar, bio, or verification badge of the account that reposted it.
Once you reach the first piece of material, you also need to judge what kind of thing it is: an official statistic, a press release, an opinion piece, a personal recollection, or another person’s retelling? The earliest appearance does not guarantee accuracy, but it tells you where the current conclusion came from.
This approach is close to Mike Caulfield’s four SIFT moves: stop, understand the source, find better coverage, and trace to the original context. The walkthrough below is an application example written for this site.
Step 3: Getting the arithmetic right solves only part of the problem
Within the stated limits of material B, (240 − 120) ÷ 120 = 100%, so the calculation “compared across these two days, loans rose by 100%” holds; 240 is also twice 120.
But you still need to check what the statistic counts. 240 volumes borrowed is not 240 readers, and it is certainly not 240 books finished. Even if one library’s two-day data is accurate, it cannot represent the whole city or a long-term trend.
Next, complete the scope: which two days are being compared, whether it is the same branch, whether opening hours were the same, whether there was an event that day, whether renewals are included, and whether the data has been corrected. Anything the material does not state should be listed as still to be checked, rather than filling in the author’s answer for them.
Publication date and event date also have to be separated. A screenshot reposted today may describe an event from long ago. An image with no year, month, location, or source can only serve as a lead; it cannot fill in those details by itself.
Step 4: Ten retellings may still mean one source
Suppose you find ten more articles, all citing the same event record. That shows the story spread widely, but it does not add ten independent statistics. When comparing evidence, look at where each report ultimately points.
If you find another independent record, check whether its time, scope, and method match the first. Two different numbers may mean different definitions, or one side may be wrong; you need to investigate the difference rather than average them.
For more on who holds information and who only sees a summary, read Solving Information Asymmetry. Keeping the full source chain is more useful during verification than collecting many links that all repost each other.
Step 5: Give every item a bounded conclusion
Back to this fictional material, you could record it this way:
- “Loans on event day were double the comparison day”: supported by the given record, computed as 240 ÷ 120 = 2; real verification still requires confirming the record is authentic and checking the statistical scope.
- “More different people took part in borrowing”: insufficient evidence, no de-duplicated count of borrowing users.
- “The event changed reading habits”: cannot confirm at present; there are only two days of records, and no data on reading completion or sustained behavior.
- “Everyone in the city reads more now”: this goes beyond the given evidence; one library’s loan volume cannot support a city-wide conclusion.
If you want to write it into an article, you could change it to: “In the event record provided, this library lent out 240 volumes on event day, against 120 the same weekday the previous week, a 100% rise in single-day loans. The available records are not enough to judge whether reading habits across the city have changed.”
This wording states the scope of the source, what is being compared, and the limitation. It does not hide an unproven inference inside a confident headline.
If you hand the verification process to AI, what still has to be done by you?
You can use the fact-checking prompt to have a model break out the claims, list the evidence gaps, check the arithmetic, and organize a verification table. Give it the original text and the full URL where you can, not just a screenshot taken out of context.
A person still has to open the source and confirm that the page really exists, that the original text matches the summary, and that the cited date and scope were not swapped out. A model saying “verified” is not evidence, and a title that looks like a real paper does not replace a paper you can look up.
If you do not have network access, limit the conclusion to “based on the material provided.” Failing to find evidence can neither automatically prove the story false nor let you relax and write it as confirmed.
A verification record you can save as is
Each time you only need to keep six items: the original sentence to check; the publisher and date; the original source and the relevant passage; the numbers, times, and scope; independent corroboration or conflict; the current conclusion and the questions still open. Once the record is done, then decide whether to repost or quote it.
When the material needs to become an article, you can go on to use the article outline prompt. Let the evidence decide how far the conclusion can go before you arrange the paragraphs, and it is easier to avoid one loud headline burying the facts.