Next Word, By Hand

Overview

A generative AI chatbot writes by guessing what word is likely to come next, based on patterns in the writing it learned from. In this game, teams become that machine. They tally which word follows which in eight short sentence cards, then "write" new sentences by drawing slips from cups. The sentences come out grammatical and often silly. Then, in the lake round, the machine produces a sentence that sounds like a fact and is false, even though every piece of it came from true cards. Students test whether a longer prompt fixes the problem and leave able to say, in their own words, why a chatbot can sound sure and still be wrong.

Objectives

• Students will build a next-word model by tallying which word follows which in a set of sentence cards and use it to generate new sentences.
• Students will explain, in their own words, that a chatbot writes by predicting likely next words from patterns, not by looking up facts.
• Students will explain how their model produced a false "fact" out of true pieces, and name one way to check a chatbot's answer.
• Students will compare output from a short and a longer prompt and describe what the longer prompt fixed and what it could not fix.

Step by step

1. 10–4 min

   Hook

   Finish my sentence

   Pause before the last word each time: "Peanut butter and…" (the class says jelly). "Once upon a…" (time). "Twinkle, twinkle, little…" (star). Ask: "Did anyone look that up? How did you know?" Take answers, then say: "You predicted. You've heard those words together so many times that your brain guessed what comes next. A generative AI chatbot does something like that, one word after another, with a huge amount of writing. Today your team is going to build a tiny one out of paper."

   Facilitator noteSome students will shout silly endings ("peanut butter and pickles!"). Welcome them: they are proof that more than one next word is possible, which matters when the class starts drawing slips.

2. 24–9 min

   Model

   Tally one card together

   Show Story Card 1 under the document camera or write it on the board: the dog ran to the park . Point to each word and ask, "What word came right after it?" Put a tally on Handout B: after the, a mark for dog; after dog, a mark for ran; after ran, a mark for to; after to, a mark for the; after the again, a mark for park. Explain the one special rule: "When you draw park, bone, door, mat, bus, bench, or apple, the sentence is over. Put a period."

   Facilitator noteStudents often forget that "the" shows up twice in one card, so it gets two tallies. Catch it here, out loud, so every team starts right.

3. 39–18 min

   Explore

   Train the machine

   Teams split Story Cards 1–8 so each person tallies two cards on the shared Handout B. Then they make one slip for each tally mark, write the next word on it, and drop it in a cup labeled with the first word. For example, the ran cup gets three slips that all say to. Say: "Your cups ARE your machine. This is what we call training: the machine only knows what was in these cards."

   Facilitator noteSuggested jobs: Reader (reads each word pair aloud), Tallier, Slip Maker, Cup Keeper. Look-for: the "the" cup should have 16 slips with lots of different words. Ask a team, "Why is this cup so much fuller and messier than the ran cup?"

4. 418–25 min

   Practice

   Generate!

   Every sentence starts with the. Draw a slip from the the cup, write the word down, put the slip back, then draw from that word's cup. Keep going until the sentence ends. Each team writes five sentences and copies their best and silliest onto the "Our machine said…" chart. Ask the room three questions: "Do these sound like real sentences? Are they true? Did the machine ever see these exact sentences?"

   Facilitator noteExpect sentences like "the cat sat on the bus." Grammar is fine because the word patterns are real; the meaning is off because the machine keeps track of word order, not the real world. Faster option: close your eyes and tap a pencil on the tally marks in that word's row instead of drawing slips.

5. 525–33 min

   Apply

   The lake round

   Open the envelope. Read Lake Cards 9–12 aloud. They are the only facts the machine has about two made-up lakes. This time the machine looks at the last two words to guess the next one. Most two-word pairs have only one possible next word, so teams just follow the card. There are only three kinds of forks: after Wrenmere is or Mossmere is it could be in or the; after is in it could be Harlow or Pike; and after is the it could be deepest or largest. At each fork, flip the coin. Each team generates three sentences starting "Mossmere is" and three starting "Wrenmere is." Before long a team gets "Mossmere is the deepest lake in the state." Stop the class: "No card says that. Card 11 says Wrenmere is the deepest. Where did this come from? Is the machine lying?"

   Facilitator noteThis is the big idea. The false sentence is built entirely out of true pieces. The machine isn't lying or broken; it has no list of facts to check against, only patterns of which words go together. Name it: when an AI tool says something false in a confident voice, people call that a hallucination.

6. 633–38 min

   Practice

   Does a longer prompt fix it?

   Now give the machine three words to look at. After "Wrenmere is in," the cards only ever have Harlow. After "Mossmere is the," they only have largest. Teams regenerate and notice the false sentences disappear. Ask: "So does a longer, clearer prompt make a chatbot always right?" Guide students to two ideas: a clearer prompt usually helps, AND it can't fix what the machine never learned or learned wrong. If Card 11 had a mistake in it, no prompt could make the machine know better.

   Facilitator noteAdd the safety piece while prompts are on everyone's mind: "A longer prompt means you're telling the machine more. Names, addresses, and private stuff about you or your friends never go in a prompt."

7. 738–45 min

   Reflect

   Paper vs. real chatbots, honestly

   Be honest about the difference: "Real chatbots learned from far, far more writing than eight cards. They use pieces of words, not just whole words. They look back at much more than two or three words, and people trained them more so they answer helpfully. But the main move is the one you just did: guess what's likely to come next. That's why a chatbot can be helpful, sound sure, and still be wrong. So what do we do? We check with a trusted source: a book, a website a grown-up trusts, an expert." Students complete Handout C. Collect it as evidence.

   Facilitator noteListen for students who say the chatbot "lied" or "is dumb." Push back gently: "Did our paper machine lie? What was it actually doing?" You want explanations built on patterns and guessing.

Adaptations

Grade 3

Build one class machine at the front with student helpers, and run the lake round as a whole class with you flipping the coin. Use the sentence frame: "The machine said that because it saw ___ and ___."

Grades 6–8

After the three-word round, ask teams to write two new true lake cards that would make a hallucination more likely, then trade with another team to test them. Discuss why more varied training text makes more varied, and sometimes wrong, output.

Emergent bilinguals

Make a second set of Story Cards in a student's home language and compare which words follow which. Word order that differs between languages makes the "next word" idea vivid.

Students with fine-motor needs

Skip slips: generate by pointing to tally marks or rolling a die and counting along the row. A partner can write the sentence.

Pairs well with

• Gen AI Literacy Breakouts (K–8)Teacher-projected breakouts that extend how AI works and when to double-check it.