Free lesson

The 5-Minute Check: How Fact-Checkers Do It

The 5-Minute Check: How Fact-Checkers Do It — AI for Research: Find Information Fast and Check What’s True. Why careful reading is not enough; SIFT: four practical moves; Fictional worked example: the lemon-water message; Applying the check to an AI answer; Depth of checking and a written verdict. Practical assignment: One check with a written trail.

Demo edition. The materials are undergoing editorial review.

Topics covered

  • Why careful reading is not enough
  • SIFT: four practical moves
  • Fictional worked example: the lemon-water message
  • Applying the check to an AI answer
  • Depth of checking and a written verdict

Why careful reading is not enough

A page can have polished design, a logo, a .org domain, an “About us” page, and a long reference list—and still be unreliable. The publisher controls those signals. A fact-checker’s first decision is therefore whether to keep reading the page or first establish who made it.

A study of lateral reading compared 10 PhD historians, 10 professional fact-checkers, and 25 Stanford undergraduates. The historians and students more often stayed on the site and evaluated its internal features. The fact-checkers quickly opened new tabs, searched for information about the organization, and compared independent sources. The researchers called this lateral reading.

This does not mean never reading the page. Establish its provenance and reputation first, then return to its content. The same principle applies to an AI answer: fluent prose is a presentation feature, not evidence that the answer is correct.

SIFT: four practical moves

SIFT is a short procedure developed by digital-literacy educator Mike Caulfield.

  1. Stop. Do not share immediately. State the checkable claim in one sentence and separate it from opinions. “This drink is healthy” is too vague; “this drink lowers blood pressure by 20% in two weeks” can be checked.
  2. Investigate the source. In a new tab, identify the author or owner, their relevant expertise, and whether they sell something connected to the claim. Their own “About” page supplies leads, not independent confirmation.
  3. Find better coverage. Search for the same claim in competent, independent sources. Several sites rewriting one company press release do not provide several independent confirmations.
  4. Trace to the original. Open the full study, official document, registry entry, full video, or transcript. Check the date and context.

SIFT does not produce truth automatically. It helps you reach better evidence quickly. If one move provides a decisive answer, you do not have to spend equal time mechanically completing the others.

Fictional worked example: the lemon-water message

Fictional learning scenario. Marta receives this message: “Japanese scientists PROVED that warm lemon water on an empty stomach lowers blood pressure by 20% in two weeks. Doctors are hiding it. Share with everyone over 50.”

Marta first records the exact claim: “Warm lemon water lowers blood pressure by 20% in two weeks.” Capital letters and a demand to share are reasons to slow down, but they do not themselves prove that the claim is false. She investigates the site: no authors are named, while the site sells “detox” drops. A search for independent coverage produces only copies of the same text; the national health service’s pages do not list this treatment. There is no study link, and older versions attribute the claim to scientists from different countries.

Verdict: not supported, rather than “conclusively disproved.” Failing to find a study does not always prove that none exists. It does mean this message is not a sound basis for action, especially for changing treatment. Decisions about prescribed medicine require discussion with a qualified health professional.

Applying the check to an AI answer

An AI answer often bundles claims of different importance. Begin with a claim map: list facts, numbers, dates, quotations, and names separately. For each item, record the proposed source, what checking would require, and the claim’s role in the conclusion.

A claim is load-bearing when its being false would change your decision or main conclusion. In an answer about returning a purchase, the deadline, jurisdiction, and exceptions are load-bearing; the history of the law may be minor. Run SIFT on the load-bearing claims first.

A useful assistant prompt is: “Split the previous answer into separate checkable claims. For each, give the source or ‘from memory,’ what would be needed to check it, and whether it is load-bearing. Do not add or verify anything yourself.”

A common mistake is asking the same AI, “Are you sure?” The model may repeat the answer or change it in response to the wording. That is another generation, not independent verification. Open any cited link, confirm that it exists and actually supports the claim, then find an outside source.

Depth of checking and a written verdict

The depth of a check depends on the cost of error. A quick glance may be enough for a recipe or a film summary. For a purchase, work summary, or travel plan, run a quick SIFT check on the load-bearing claims. For health, money, law, safety, or something you will publish, trace every load-bearing claim to a current primary source and check the applicable date and jurisdiction.

A low personal cost does not justify broadcasting an unchecked claim; sharing transfers possible consequences to other people. Finish with one of four verdicts: supported, partly supported, not supported, or couldn’t verify. The last does not mean “false.” It means the available evidence is insufficient and the claim should not yet carry your decision.

Record the exact claim, the decisive source, the access date, and a short reason for the verdict. Common mistakes include counting copies of one story as independent sources, stopping at a search snippet, treating absence of evidence as automatic disproof, or finding a document without checking whether it says what the claim attributes to it.

Practical assignment

One check with a written trail

Continue to learning

Course outline

  1. Confidently Wrong: Why AI Gets Facts Wrong
  2. The 5-Minute Check: How Fact-Checkers Do It
  3. Search With AI Better Than With a Search Engine
  4. Links and Quotes: Trust, but Open Them
  5. Numbers, Statistics and Studies
  6. Deep Research: A Report in 15 Minutes — and How to Check It
  7. Working From Your Own Sources: NotebookLM and “Documents Only”
  8. High Stakes: Health, Money, Law and News
  9. Your Research Process: From Question to Conclusion