This lesson is for business leaders who want to check a striking claim before using it in a conversation, briefing or decision. You do not need to investigate a real account. Start with the invented material below.
The aim is to produce a conclusion someone can inspect. Your bot should show which statement it checked, the evidence it used and what it still cannot establish.
It follows Critical Thinking for Humans and Bots and Teach Your Bot to Check Its Work. The companion Writing article explains the real research behind this exercise.
Separate the claim from the presentation
A striking photograph or a large percentage may deserve attention. It does not tell you whether the evidence fits the statement.
SIFT encourages readers to pause, investigate the source, find better coverage and trace original context. Lateral reading means looking outside an unfamiliar page to investigate it. Here we practise the underlying distinctions using a closed fictional bundle, without browsing.
Watch for four different problems:
| Problem | Question to ask |
|---|---|
| Inflated scope | Does evidence about a small group support a statement about everyone? |
| Repeated sourcing | Are these independent observations or retellings of one source? |
| False context | Does an old image actually show the place and event claimed? |
| Missing evidence | Is the statement contradicted, or simply not established by what we have? |
A false claim does not, on its own, tell you whether its author intended to deceive. Keep the assessment of the statement separate from any claim about motive.
Read this invented source bundle
The labels are part of the exercise. They are not real websites or files the bot has opened.
Post being checked: “80% of UK shops lost customers. This photograph proves shops are closing now. Two independent reports confirm it. The Government has announced emergency grants.”
Source A — local survey note, 21 October 2026: Ten self-selected shop owners on one fictional high street answered a question. Eight said the previous Tuesday had been quieter than the Saturday before it. The note supplies no customer counts, national sample or evidence that a shop closed. It contains no information about Government grants.
Source B — trade blog, 22 October 2026: Repeats the survey's eight-out-of-ten result and links to Source A. It reports no new survey.
Source C — local newsletter, 22 October 2026: Summarises Source B, identifies Source A as the original and adds no separate observations.
Source D — photograph record: In this fictional exercise, the supplied record identifies the photograph as a temporary refurbishment in a different town in 2022. It is not a record of current closures. The exercise assumes this record is accurate; it has not verified a real image's metadata.
Before asking your bot, name the individual claims. Write one sentence explaining what eight out of ten actually tells you here. Then list the evidence you would still need for the grants claim.
Ask your AI to teach you
Use an approved learning chat with browsing and connected work accounts switched off. Paste this complete prompt; it includes the invented source material.
A prompt is a learning brief, not a permission control. Keep real tools and business records disconnected for this exercise. “Teach” means practising a procedure and checking its output, not retraining the underlying model.
Check your reasoning after you try
The defensible conclusion from Source A is that eight of these ten respondents said one Tuesday was quieter than the preceding Saturday. Eight divided by ten is 80%. It is not a measured national fall in customers, and the comparison does not establish a like-for-like daily trend.
B and C lead back to A. They do not provide two independent surveys. Source D contradicts the photograph's claimed current context within the fictional bundle. It does not establish what is happening to real UK shops.
The grants claim is not established by this bundle. There is no Government announcement here to inspect. That is different from proving that no grant exists. In a real check, you would need the relevant current official announcement and its conditions.
One possible briefing sentence is: “A small local survey reports a quieter Tuesday than the preceding Saturday; it does not establish a UK-wide decline, current closures or a grant announcement.” The author's motive remains unknown.
For the transfer problem, check the claimed saving against the product's previous price, then explain why one price does not establish the supplier's whole price list.
Check the checker
Look for visible failures. Did it claim to browse? Count B and C as independent evidence? Say the grants definitely do not exist? Accuse the author of lying? Accept your preferred conclusion without checking it? Any of these would require correction before you reuse the method.
Try a fresh fictional example after the correction. Keep a record of the cases and errors you observed; one successful exercise does not establish a reliable detection rate. NIST's Generative AI Profile explains why confidently stated answers and supporting citations can themselves be wrong.
For later real work, ask for approved public sources and an inspectable evidence table. If the assistant cannot access the underlying record, it should mark that limit. Reputational or other consequential conclusions need an appropriate human reviewer before use.
Sources and boundaries
The linked source-method and AI-risk references were opened on 22 October 2026. All shops, posts, source records and prices in the exercises are invented. This lesson proposes practice; it has not measured learner outcomes, model accuracy or a live misinformation-detection system.