Lesson 1: How AI makes text
Core question: Does a chatbot “know” things? Objective: S1.1 — describe how a chatbot produces text (prediction) and correct the misconception that it “knows” or “looks things up”. Handout: Next-word recording sheet · Teacher support: misconception table
Last reviewed: 2026-10-25
Hook — 5 min
Teacher script prompt: “I’m going to write a sentence on the board, but I’ll stop in the middle. You tell me the next word. ‘The goalkeeper caught the…’”
Take 4–5 answers (ball, ball, ball, ball, ball). Ask: “Why did almost everyone say ball?” → because it’s the word that usually comes next. “Congratulations — for one sentence, you are now a language model. Today we find out what chatbots actually do, and why ‘it just knows things’ is wrong.”
Main activity — 25 min
Part A: the human predictor game (10 min, works with or without devices)
In pairs, students play the game properly:
- Student A writes a sentence starter (5–8 words) on the recording sheet.
- Student B predicts the next word. A writes it and continues the sentence. Repeat for 5–8 steps.
- Round 2: same game, but B must sometimes pick a surprising but possible word (that is what the temperature slider does — see Part B).
- Both record: Which words were easy to predict? When did it get weird?
Point to land: B was never “thinking about” what the sentence means. B was predicting. Yet the sentences came out sensible — and sometimes nonsense. That is exactly how a chatbot works, just with vastly more patterns.
Part B: the next-word interactive (15 min, devices; no-devices version below)
Students open the interactive and, on the recording sheet:
- Generate a sentence word by word at low temperature. What do they notice? (Predictable, boring, safe.)
- Reset. Same starter, high temperature. (Surprising, creative, sometimes nonsense.)
- Watch the probability bars: the model doesn’t “choose” one word — it has probabilities for every next word; the slider changes how adventurous the pick is.
- Record answers to: Does the model know if a sentence is true? How can you tell from the interactive that it doesn’t?
Key observation to fish for: at no point does the interactive ask whether the sentence is true. Truth is not part of the machinery. Students should write this in their own words.
Discussion — 10 min
- “So: does a chatbot know things?” → It predicts likely text. It can produce true things, false things, and perfect-sounding nonsense — with the same confident tone.
- “If it doesn’t know anything, why do answers often look right?” → Because likely text is often right — and sometimes not. You can’t tell from the tone.
- “When could this help you, and when could it trick you?” → Collect on the board; Lesson 4 returns to this list.
Exit ticket — 5 min
Students write two sentences without looking at anything:
- How does a chatbot produce text?
- Why is “it looked it up” wrong?
Meet the objective = both sentences correct the misconception in their own words.
No devices
Skip Part B. Run two more rounds of Part A: Round 3 at “temperature 0.1” (B must pick the most predictable word every time) vs. Round 4 at “temperature 1.5” (B must surprise). Draw the probability bars on the board for “The goalkeeper caught the ___”: tall bar for ball, thin bars for rebound, cold and trophy (the same words the interactive offers). The lesson works identically on paper.
Differentiation
Support: sentence starters provided on the recording sheet; sentence frames for the exit ticket (“A chatbot doesn’t __; it __”). Stretch: “Why do companies add a temperature control for customers? Name one product where you’d want it low and one where you’d want it high.” Also: connect to Lesson 2 — if truth isn’t in the machinery, what has to be?