Teaching AI literacy with paper and dice
The Australian Framework for Generative AI in Schools calls for students to learn how generative AI works, including its limitations and biases. It also covers privacy, fairness and student wellbeing. I agree with those goals, but teachers need time and support to put them into practice.
As Lucinda McKnight and Leon Furze argued in The Conversation, the framework places an “extraordinary onus” on teachers. Teachers already have full workloads. If you’re trying to explain how a language model generates text, you can start with an activity that uses paper and dice.
In LLMs Unplugged, students build their own language model from a short text, such as a few sentences from a picture book. They count how often each word follows another, record the counts in a table, then roll dice to select the next word. Repeating this produces new text, sometimes funny or nonsensical and occasionally plausible.
Students can inspect the counts and follow how each dice roll selects the next word. They can also see how changing the training text changes the available words and their probabilities.
This demonstrates next-token prediction and sampling, two ideas used in modern LLMs. The paper model estimates probabilities from counts of neighbouring words; modern LLMs use trained neural networks and much longer contexts. The activity doesn’t teach transformer architectures, but it gives students a model they can build and examine themselves.
Students can do the activity without creating accounts or sending their work to an online service. Teachers choose the training text and supervise what students generate. That still requires judgement: rearranging words from a text can produce something inappropriate, even when the original seemed suitable.
Paper and dice also keep equipment costs low. Schools don’t need a paid AI subscription to run the core activity, though teachers may need to adapt the materials for their students.
The framework recognises teacher expertise, and I’d like AI professional learning to give teachers activities they can use and understand. I’ve designed these lessons for teachers without a technical background, with instructions for counting words, completing the table and generating text. Teachers can learn alongside their students, with connections to existing curriculum outcomes in maths, digital technologies and English.
I use Claude Code all the time for software development, and I think schools should discuss how to use AI tools. Building a small model by hand gives teachers and students something concrete to refer to in those discussions.
If you’d like to try LLMs Unplugged in your school, get in touch.