50 of 81
ContentsPart IV · Chapter 12

Chapter 12Check before you trust

Match the check to the claim

What counts as an authoritative source depends on the claim. A total is checked against the figures, a quote against the document it came from, a date against whoever sets it. So the first step is to read an answer as a set of separate claims, then give each claim the check that fits. Figure 12.1 sets out five kinds of claim and the check for each.

If the answer states …… check it this wayA number or a totala quarter’s spending, a percentRedo the mathfrom your own figuresA quote or a named sourcea section, a study, a court caseOpen the sourceand find the words yourselfThat something is missing“there is no such clause”Search the original yourselffor every word it might usehardestA date, a name, a plain facta deadline, a season’s startLook it up elsewheresomewhere you already trustA judgment about a person“the strongest applicant”A person who knows decidesit is not a fact to look up
Figure 12.1. Five kinds of claim, five checks. Read down the left for the kind of claim, then across for the check that settles it. The claim that something is missing is the hardest, because the answer can’t show you what isn’t there. The last row is not a fact at all: no lookup settles it, so a person decides.

Try it: pick a claim, then pick how you would check it.

How would you check it?

Illustration

Interactive Pick a claim, then pick how you would check it.Try it on the web: aifromthepartsup.com/check-before-you-trust/3/

Different claims need different checks.

1. Pick a claim

2. How would you check it?

Now pick an answer.

  • “Travel spending was $48,210 this quarter, up 12% from last quarter.” Redo the math. You have the spreadsheet, so recompute it. A language model predicts the next token, digits included; a total that reads well can still be off.
  • “Section 4.2 of the employee handbook gives new hires 15 days of paid time off.” Open the source it names. It names its source, so open section 4.2 and read it. A real section number doesn’t make the sentence in it real.
  • “A 2019 decision, Smith v. Riverside Co., settled this question.” Open the source it names. Find the decision itself. Courts have sanctioned lawyers for filing decisions that never existed.1
  • “The 40-page vendor contract has no termination clause.” Search the original document yourself. Claims that something is missing are the hardest to check. Search the contract yourself for “terminate” and “termination.”
  • “The spring soccer season starts April 4.” Look it up somewhere you trust. Check the league’s own schedule. Dates are easy to state and easy to get wrong.
  • “This applicant is the strongest fit for the HR role.” A person who knows decides. That is a judgment about a person, not a fact to look up. A person who knows the job decides (Chapter 14).

Examples written for this page, not things an AI tool said. The names, numbers and the court case are made up.

A number. If you have the figures, redo the math. A model predicts digits the way it predicts words, so a total can read perfectly and still be off. A spreadsheet or a calculator settles it in a minute.

A quote or a named source. Open it and find the words yourself. A real-looking reference can still point to the wrong thing. In Ayinde, one invented decision came with a genuine citation number, but the number belonged to an unrelated case about a charity’s business rates.2 In the same way, a real section number in a handbook doesn’t make the sentence attributed to it real.

A date, a name or a plain fact. Look it up somewhere you already trust, such as the organization’s own site or schedule, rather than asking the same tool again. Professional fact-checkers do this by habit. A study compared 10 fact-checkers, 10 historians with Ph.D.s and 25 Stanford undergraduates as they judged websites, and the fact-checkers did best.3 They “read laterally, leaving a site after a quick scan and opening up new browser tabs” to see what others said about it. They reached better-supported conclusions in a fraction of the time.3 With an AI answer, reading laterally means the same thing: leave the answer and see what other sources say.

A judgment about a person. “This applicant is the strongest fit” is not a fact you can look up. It is a judgment, and a person who knows the job should make it and be able to explain it. Chapter 14 returns to decisions like this one.

The nurse in the survey did all of this in one move. She checked medical claims against the source her hospital trusts for medical claims.

The hardest claim: that something is missing

One kind of claim needs its own warning. “The contract has no termination clause.” “Nothing in the policy covers remote work.” “There are no other studies on this.” A claim that something exists can be checked by finding it. A claim that something is missing can’t be checked from the answer at all, because the thing you would need to see isn’t there. You have to go to the original and search it yourself, for every word the missing thing might use: terminate, termination, cancel, notice. Only then can you say it is absent.

Omissions hide inside summaries too. A summary of a forty-page document is a set of choices about what to leave out, and a fluent summary gives no sign of what it dropped. If something would matter to you, such as a deadline, a penalty or an exception, search the original for it. Don’t assume the summary would have mentioned it.

Check more when more is at stake

Not every answer needs every check. A list of dinner ideas needs none. A number going to your board, a quote in a published report, or anything that touches a person’s health, money or job needs the full treatment. A useful question is who would be hurt if this were wrong, and how badly.

Many workers in the survey said that the chance of a costly mistake was what made them stop and think, as it was for the pharmacist facing a review.4 The researchers add a caution. A skill used only in emergencies gets rusty, they note, so “it is risky for users to only apply critical thinking in high-stakes situations.”4 Small, everyday checks keep the habit ready for the day it matters.

This book holds itself to the same rule. Every fact in it carries a numbered note naming its source, so you can check the book the way this chapter suggests checking anything else.

Notes (4)
  1. 1U.S. District Court for the Southern District of New York (P. Kevin Castel, U.S. District Judge), “Mata v. Avianca, Inc., No. 22-cv-1461 (PKC): Opinion and Order on Sanctions (ECF No. 54),” No. 22-cv-1461 (PKC), ECF No. 54, June 22, 2023, https://www.courtlistener.com/docket/63107798/mata-v-avianca-inc/ (accessed October 27, 2026). Cited at p. 1.
  2. 2High Court of Justice, King's Bench Division, Divisional Court (Dame Victoria Sharp P and Johnson J), “R (Ayinde) v London Borough of Haringey; Al-Haroun v Qatar National Bank QPSC: Approved Judgment,” [2025] EWHC 1383 (Admin), June 6, 2025, https://caselaw.nationalarchives.gov.uk/ewhc/admin/2025/1383 (accessed October 27, 2026). Cited at para. 36.
  3. 3Sam Wineburg and Sarah McGrew, “Lateral Reading and the Nature of Expertise: Reading Less and Learning More When Evaluating Digital Information,” Teachers College Record 121, no. 11 (November 2019): 1–40, https://doi.org/10.1177/016146811912101102. Cited at abstract.
  4. 4Hao-Ping Lee, Advait Sarkar, Lev Tankelevitch, et al., “The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers,” in Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI '25) (New York: ACM, April 25, 2025), 1–22, https://doi.org/10.1145/3706598.3713778. Cited at pages 10, 11.

Chapter 12 · page 3 of 4

Turn pages with the ← and → keys.

Cite this page

Using this book

© 2026 Steve Sparks. All rights reserved. Teachers and students may assign, read, quote and excerpt chapters of this book for classes and study, with credit to Steve Sparks and a link to aifromthepartsup.com. Anyone may quote short passages and cite any page, with a link back. To repost, reproduce or sell the book or larger parts of it, ask through GitHub.

Quoted material and figures from other sources remain their owners'.