Switch course ▾
Session 3. Fact-Checking AI Output
Objective. Find, correct and explain errors in AI-generated classroom content using a simple, repeatable checking routine.
AI can write a worksheet in seconds, but it cannot tell you which of its sentences are wrong, and neither can reading it again more slowly. You are the checker, and your time is short. This session gives you a faster method used by professional fact-checkers: stop reading the text line by line and go out to other sources instead. Ten focused minutes can save a whole class from learning a wrong date, a wrong river or a quotation nobody ever said.
Key ideas
Why errors look right
- AI errors are not random typos; they are plausible: a near-miss date, a neighbouring country, a round number.
- Confident, tidy formatting makes a wrong claim feel checked.
- Invented sources and quotations are a known failure: they look real and cannot be found.
Where errors hide
- Names, dates, numbers and units.
- Calculations and totals — AI often gets arithmetic wrong.
- Definitions that swap two similar terms.
- Local facts: your country, your curriculum, your region.
The five-step routine
- 1 Mark every checkable claim.
- 2 Rank by risk: what would hurt a student if wrong?
- 3 Check against a trusted source — textbook, official site, or two independent sources.
- 4 Correct it and note what was wrong.
- 5 Re-read the whole text once more.
Trusted sources, and what to do when none exists
- Prefer your approved textbook, official bodies and original documents over blogs and other AI outputs.
- Asking the same AI 'are you sure?' is not checking.
- If you cannot verify a claim, remove it. A missing fact is better than a wrong one.
Lateral reading of an AI claim
Lateral reading: leave the page
Researchers at the Stanford History Education Group (now the Digital Inquiry Group) found that professional fact-checkers do not stay on an unfamiliar page and judge it by how it looks. They open new tabs and ask what other sources say. This is called lateral reading.
Treat the AI answer as that unfamiliar page. Reading it a second time, however carefully, only makes it feel more familiar. Instead, pick one claim, search for it outside the chat, and see whether a source you trust says the same thing. Familiar is not the same as confirmed.
Choose what to check, then say what you found
You cannot check everything, so check what is specific, easy to test and costly if wrong: numbers, dates, names, definitions that students will memorise, and any quotation or source the AI mentions. A reference that you cannot find after a short search should be treated as invented.
Use two independent sources, not two pages that copy each other. If your textbook and a reference book disagree, follow the textbook your students are examined on, and note the difference. Finish by writing a short mark next to each claim: confirmed, corrected, or removed because it could not be verified.
Examples from the classroom
The river lengths
Scene. Dilnoza teaches Grade 7 geography to 32 students in Nukus. An AI-written worksheet says the Amu Darya and Syr Darya both feed the Aral Sea and gives a length for each. She underlines the two numbers, checks the atlas table in the textbook and one encyclopaedia, and finds one figure differs. In 10 minutes she corrects it and deletes a sentence she cannot confirm.
Insight. Numbers are the first place to look. When sources differ, name the one your curriculum follows.
The missing quotation
Scene. Aziz prepares a Grade 10 history lesson on Samarkand. The AI adds a stirring quotation attributed to Ulugh Beg, with no source. Aziz searches the exact words together with the name in two library catalogues and a trusted encyclopaedia and finds nothing. He removes the quotation and instead asks students to find one real, referenced statement.
Insight. A claim with no findable source should be removed. Invented quotations look real.
The answer key
Scene. Malika teaches Grade 5 mathematics in a Russian-medium class in Tashkent. The AI produced 12 problems with an answer key. Before copying it, she solves each problem on paper in 6 minutes. One total in the key is wrong. She also asks two students to solve it at the board, so the class sees checking as normal.
Insight. AI is not a calculator. Recompute every answer yourself, and make checking visible to students.
Common pitfalls
- Asking the same AI whether it is sure and treating a confident yes as a check.
- Checking only the last paragraph or only what looks suspicious, while a quiet wrong date passes.
- Using a second website that is itself copied from the same AI output.
- Keeping an unverified fact because deleting it leaves a gap in the lesson.
Find the Planted Errors
- Take the worksheet 'Central Asia: Rivers and the Aral Sea' from the Activity Kit. It is a sample AI-written worksheet with 13 statements.
- Alone (15 min): mark each statement True, Error or Cannot verify. Use the five-step routine and a reference source.
- In pairs (15 min): compare, resolve differences with evidence, and correct each error in writing.
- For each error write one line: why does this one sound plausible? This is the highest-level skill.
Your facilitator will guide you through this activity in the session.
Questions to discuss
- Which error was hardest to spot and why?
- What would have happened if you had handed this worksheet to a class unread?
- What is your personal rule for deciding a claim 'needs checking'?
Reflect
- Think of the last time you trusted a source because it sounded confident. What would checking have cost?
- Which subject you teach has the most local facts that AI is likely to get wrong?
- How can you teach your students to check, not only to trust?
Try it tomorrow
- Take one AI-written text for next week and mark the five riskiest claims before checking anything.
- For each marked claim, open a new tab and confirm it in your textbook and one other source.
- Show students one corrected error and ask them how they would have caught it.
Further reading and links
Plausible is not true. Check names, numbers and sources, and delete what you cannot verify.