Lesson 2: When AI gets it wrong
Core question: Why are wrong answers so convincing? Objective: S2.1 — identify at least two claims that need checking and verify each with a reliable source. Handout: Fact-check worksheet (includes the worked example and answer key reference)
Last reviewed: 2026-10-25
Hook — 5 min
Teacher script prompt: read this aloud (do not fact-check it yet):
“The Great Fire of London started in 1666 in a bakery on Pudding Lane and destroyed 13,000 homes. The city was rebuilt in exactly five years, funded by the King’s personal fortune of two million pounds.”
Ask: “How many of you think all four facts in there are right?” (Show of hands.) “Now — how do you know you think that? You checked nothing. It sounded right. That feeling has a name: fluency. Today we build the habit that beats it.”
Main activity — 25 min
The fact-check challenge (pairs)
Each pair gets a chatbot-style answer (below) and the worksheet. Rules:
- Read it once, normally. Mark with a pencil: T (sounds true), ? (not sure), F (sounds false).
- Now find at least two claims that need checking. Claims to check = claims with a number, a date, a name, a source, or a superlative (“the first”, “the only”, “exactly”).
- Verify each claim with a reliable source (textbook, library site, official website — agree in class which sources count).
- Record on the worksheet: claim → what the source says → verdict (correct / wrong / can’t tell).
The sample answer to check (also printed on the worksheet)
“The Great Wall of China is the only man-made structure visible from space with the naked eye. It stretches exactly 21,196 kilometers. The wall was built mainly to keep out rabbits and deer, a fact documented in the 1998 book Walls of the Middle Kingdom by Professor Henry Caldwell.”
(Answer key for the teacher: the “visible from space” claim is a famous myth — false; the 21,196 km figure is roughly right for all sections of the wall across dynasties — needs the source named; the rabbit/deer claim is invented; the 1998 book does not exist. Students need two verified claims to meet the objective; the conversation about why the false ones sounded fine is the real lesson.)
Teacher moves during the activity
- Circulate and ask: “What kind of claim is that?” (number/date/name/source/superlative → check it)
- When a pair finds a false claim, ask: “Did your first pencil mark say T or F?” Most will have said T. That is the hook for the discussion.
Discussion — 10 min
- “Why are wrong answers so convincing?” → Because they are built from likely text — they sound exactly like true text. Tone carries no truth signal.
- “What types of claims must always be checked?” → Numbers, dates, names, sources, superlatives. Collect the class list on the board — this list is Lesson 4’s rule material.
- “What do you do when the source partially disagrees?” → Stretch pairs share; note that “mostly right with one wrong number” is the most dangerous case of all.
Exit ticket — 5 min
One sentence: “A chatbot answer sounds convincing. Name the two things you will check before you use it.”
No devices
No devices needed at any point — verification uses printed sources or the library. If sources are scarce, the teacher pre-prints 4–5 “source cards” (true reference pages) and pairs must match their claims to the right card.
Differentiation
Support: worksheet has sentence frames (“The claim says __. The source says __. So the claim is ___.”) and asks for two claims, not three. Stretch: find the partially true claim and write why partial truth is harder to spot than total nonsense; or bring your own chatbot answer from home and fact-check it (personal AI use discussed without judgment — see the teacher guide).