Think Like An AI (Unplugged)

Overview

Today’s learners talk to chatbots that produce fluent, confident language, and it is easy to assume the machine understands or knows things the way a person does. In this unplugged activity, students experience the surprisingly simple mechanic underneath a chatbot: predicting a likely next word from patterns in text. Working in small groups, they build a tiny “language model” by hand from a short training text, generate new sentences from it, and then “turn up” the amount of context to see the output become more coherent.

By imitating mechanical thinking on purpose, learners become able to articulate what it is and what it is not, and gain the confidence to choose the style of thinking that fits the problem. 

Career Connections: AI/Machine Learning Engineer, Data Scientist, Data Analyst, Data Architect 

NB Curricular Connections

What You’ll Need

• • Printed training corpus (See Files) — one per small group.
  • Tally sheet (See Files) or blank paper, and pencils.
  • Optional: one six-sided die or a coin per group, for sampling.
  • Board or chart paper for the whole-group reflection.

Instructions

Engage – Auto-complete Simulation (10 min)

• Write on the board: “The cat sat on the ___.”

• Invite students to call out the next word (most say “mat,” “couch,” or “floor”). Ask how they knew, draw out that they used patterns from all the language they have ever heard.

• Then try one that forces a choice: “I poured milk into my ___.” When students disagree, name it: “There were several likely words, and you each picked one. A chatbot does exactly this to write a paragraph.”

 Explore – Building a Large Language Model (Generative AI) (15 min)

• Tell students they will discover how generative AI learns to “think”. Give each group the training corpus (See Files). Explain: “This tiny text is your model’s entire experience of language. It knows nothing else.”
• Groups build a table: for each word, list every word that comes immediately after it in the corpus, with a tally for how often. Model one entry on the board first (for example, “the” is followed by “robot,” “data,” and “pattern”).
• Remind students: You are not thinking about what the sentence means, whether it is true, or what would be wise to say. You are only counting which words tend to sit next to which. That counting is the ‘mechanical thinking’ we are imitating.

Generating and Sampling (5–10 min)

• Pick a starting word (give every group the same one, e.g. “The”) and look it up in the table. Students share what the word after it was.
• If there is more than one option, assign the options to die rolls (or flip a coin for two), roll, and take the winner, the model “not always choosing the most likely word.”
• Write the chosen word, then repeat the lookup for the next word. For example, if “The” was followed by “robot”, the students should write down “The robot”. For the next word, they should check their table for what words commonly followed “robot”.  
• Continue for 12–15 words or until a natural stop. Groups read their sentences aloud. Expect a mix of plausible phrases and nonsense.
• Ask students: “Why did every group start from the same word but end up with different sentences?” (Different samples, randomness) This is why asking a chatbot the same question twice can give different answers.

Turning Up Context (5 min) – Optional

• Change one rule: to pick the next word, look at the last TWO words together and find what tends to follow that pair.
• Generate a new sentence and compare it to the one-word-back version. Draw out the pattern: more training data usually means more coherent output. Real chatbots consider a larger set of words and are trained on far more than our paragraph, which is why they sound so fluent. However, the underlying move is still predicting the next word from patterns.

Closing Reflection (10–15 min)

• After the work, capture two columns on chart paper: what the “model” DID (matched patterns, used probability, sounded fluent, copied its training text) versus what it did NOT do (understand meaning, know what is true, hold a goal, check anything against the world). Then pose the choosing questions:

  • When is mechanical thinking like this genuinely useful? (Drafting, brainstorming, autocompleting familiar text, getting unstuck.)
  • When would you refuse to rely on it? (Deciding what is true, a safety or medical question, a moral choice — anywhere being confidently wrong is costly.)
  • Where do you run on ‘autopilot’ in your own thinking? When is that fine, and when is it a trap?
  • If a chatbot always sounds confident, how should that change how much you trust its fluency?
• Exit ticket:

  Each student completes: “A chatbot writes by ___. It does not ___. One time I would trust this kind of thinking is ___, and one time I would not is ___.”

Extension Ideas (Apply):

• English Language Arts: rewrite the corpus in a distinctive style (fairy tale, sports commentary) and notice how the generated “voice” changes with the training text.
• Mathematics: turn the follower tallies into fractions or percentages and discuss expected versus actual outcomes when sampling.
• Bias: if the corpus over-represents an idea, the model repeats it. What happens if the training text is one-sided? Connect to real-world data bias.