The Human Language Model

Overview

Generative AI chatbots don't look facts up the way a search engine or a database does. They generate text by repeatedly predicting a likely next piece of text based on patterns learned from enormous amounts of training data. In this unplugged session, teams "train" a ten-sentence language model by hand, generate new sentences by drawing slips from a cup, and then ask their model a factual question. The fluent, plausible, wrong answers that come out are the most memorable explanation of hallucination most educators will ever get, and one they can run with students the next day.

Objectives

• Participants will build a next-word frequency table from a small training text and use it to generate new sentences by chance draws.
• Participants will explain in their own words why a system that predicts likely next words can produce fluent text that is false (a "hallucination").
• Participants will distinguish accurately between what the paper model shows and how real large language models differ (tokens, longer context, learned weights, scale).
• Participants will draft a one-minute explanation of hallucination for their own students or colleagues.

Step by step

1. 10–5 min

   Hook

   Finish my sentence

   Say slowly: "Once upon a…" and let the room finish it ("time"). Then: "Peanut butter and…" ("jelly"). Then: "The capital of France is…" ("Paris"). Ask: "Did you look any of those up, or did you predict them?" Tell them: "Today you're going to become a language model. A very small one. By the end you'll be able to explain to anyone why AI sometimes makes things up."

   Facilitator noteThe third sentence is deliberate. Prediction and knowing a fact can feel identical from the inside. That's the whole session in one moment.

2. 25–15 min

   Explore

   Train your model

   Teams read Handout A, the ten-sentence "training library." On Handout B they tally, for each listed word, which word came right after it every time it appeared (ignore capital letters; count a period as the word END). Then they write one slip per tally mark, e.g., the word "the" is followed by "teacher" four times, so they write four "teacher" slips and file them under "the." Group slips by starting word in small piles or envelopes.

   Facilitator noteThis is "training": turning text into statistics about what tends to follow what. Teams that ask "what about meaning?" are asking exactly the right question. The model never gets meaning, only patterns.

3. 315–25 min

   Practice

   Generate: draw the next word

   Teams draw Card 1 from Handout C. Start with "The." Put all the slips that follow "the" into the cup, shake, draw one, write it down. Now empty the cup and load the slips that follow that word. Keep going until you draw END. Generate at least four sentences. Read the best (and worst) aloud and post them on the "Our model said…" chart. Expect gems like "The principal opened the award in May."

   Facilitator notePoint out that the draw is random but weighted: "teacher" comes up more often after "the" because it appeared more often in training. That's why the same prompt can give different answers on different tries.

4. 425–33 min

   Apply

   Ask your model a fact

   Teams draw Cards 4–6, the fact questions. For "Who won the spelling bee in June?" they generate an answer by starting "The" and drawing until END, re-drawing only if the sentence doesn't mention the spelling bee. Record answers. Most teams get something fluent and specific, e.g., "The teacher won the spelling bee in June," that appears nowhere in the training text. Ask: "Is your model lying?"

   Facilitator noteLand the key idea: the model isn't lying or "confused." It has no mechanism for checking truth. It produced a likely-sounding continuation, which is exactly what it was built to do. That's a hallucination.

5. 533–40 min

   Model

   Honest differences: paper vs. real models

   Walk through Cards 7–9 aloud. Real models (1) work with tokens, often word pieces, not whole words; (2) weigh a long stretch of prior context, not just one previous word; (3) learn billions of numerical weights in a neural network rather than keeping literal tally piles; and (4) are further tuned with human feedback to be helpful and follow instructions. Some tools also search the web or a document and feed what they find into the prompt, which helps but does not guarantee accuracy. Then ask: "What stays the same?" (Output is generated by predicting likely text from learned patterns. Fluency is not evidence of truth.)

   Facilitator noteResist the urge to say "it's just autocomplete." It's far more capable than that, and people who later see a strong answer will feel misled. Say instead: "It is prediction at enormous scale, which is why it is both impressive and unreliable in specific ways."

6. 640–46 min

   Create

   Your one-minute explanation

   Each participant draws Card 10 and writes a one-minute explanation of hallucination for their own audience: second graders, high school seniors, a parent, or a faculty senate. It must include (1) prediction from patterns, (2) why that can produce false statements, and (3) one thing to do about it (check a trusted source). Partners read each other's aloud and give one "clear" and one "accurate?" note.

   Facilitator noteListen for explanations that slip into "it looks things up" or "it thinks." Gently redirect to "it predicts likely words."

7. 746–50 min

   Reflect

   Where does AI belong?

   Close with: "Knowing how this works, where in your subject would an AI chatbot's answer be most likely to sound right and be wrong?" Collect answers on the chart. Optional: project your district-approved chatbot and ask your pre-planned local or obscure question. Compare its fluency with the paper model's.

   Facilitator noteTypical answers: local history, recent events, specific citations and page numbers, niche statistics, names of small-town people. These are thin-data zones where plausible invention is most likely.

Adaptations

K–2 and 3–5 teachers

With young students, use only sentences 1–5, tally just "the" and one other word as a class on the board, and have the teacher draw from the cup. Keep the takeaway to one line: "AI guesses the next word. Guesses can be wrong, so we check."

Higher Ed faculty

Add a round where teams remove sentence 8 from the library and retrain, then discuss what happens to facts that appear once, or never, in training data. Connect to why AI-generated citations and references must be checked individually.

Emergent bilinguals

Build a second library of five sentences in a student's home language. Compare how word order changes the tally, and discuss why a model trained mostly on English text may perform less reliably in other languages.

Pairs well with

• Gen AI Literacy Breakouts (K–8)Student-facing follow-up games on how generative AI works and why we verify.
• Augmented K–16 AI Learning FrameworkHelps place this concept lesson within a K–16 AI literacy progression.