Core question. If a human tried to do what a language model does — guess the most likely next word, every single time, with no intent — what would that feel like? What would the output reveal?
The three roles
The Prompter
Writes a sentence stem on a card or whiteboard. Keeps it short. Does not reveal what continuation they expect.
The Model
Must say only the single most likely next word — no full sentences, no intent, no creativity. Just the statistically probable next token.
The Observer
Watches without intervening. Notes what the Model's choice reveals about genre, register, assumptions, and defaults. Shares findings in the debrief.
Step-by-step (for a Zoom call or in-person room)
| Time | What happens | Facilitator move |
|---|---|---|
| 0–2 min | Assign roles. Three volunteers (or three breakout participants). Everyone else is a second Observer. | Explain the constraint for Role B: one word only, the most statistically probable, no intent allowed. |
| 2–5 min | Prompter writes a stem: "The scientist walked into the lab and ___" | Do not help Role B. Silence is fine — guessing is hard without intent. |
| 5–10 min | Role B says one word. Role A adds that word to the stem and presents a new blank. Repeat 5–8 times to build a "sentence." | Point out when the output starts to feel fluent despite zero intent. Ask observers what they notice. |
| 10–15 min | Rotate roles or run a second prompt with a different genre (news headline, poem, code comment). | Choose a prompt with obvious genre cues so the statistical pull is visible. |
| 15–25 min | Debrief (see prompts below). | Move toward the reflection frame: human / machine / system / ethics. |
Debrief prompts
- What made some next-word choices feel obvious to Role B? What was pulling them?
- When did the output start to sound like language — even though no one was trying to say anything?
- What does Role B's discomfort (if any) tell us about how we normally read and write?
- What defaults appeared in the generated text — about gender, profession, setting, or register?
- What did Role C notice that Role B couldn't see in the moment?
- If Role B had a larger "corpus" to draw from, would the output change? Would the defaults?
Variations
Genre swap. Run the same stem in two genres — a news lead and a children's story. Compare how much the expected next word shifts.
Temperature dial. Ask Role B to try twice: once picking the word that feels most certain (low temperature), once picking something unexpected but plausible (higher temperature). Discuss how the output changes.
Collective model. Instead of one person, the whole room votes on the next word by show of hands or Zoom reaction. The most votes wins. This models the aggregation that training on many documents produces.
No-AI showcase pathway. Role B writes a model card for their own performance: what did they do well, where did they fail, what data (life experience, genre exposure) shaped their defaults?
Connection to the tools
Pathway tags
This activity is a first-class option for two pathways:
- Critical / No-AI: The activity produces a model card, a reflection on defaults, or a consent checklist — with no AI tool opened at any point.
- Teach / Design: Facilitators can adapt the role-play for any age group or subject area by changing the stem genre (math problem, recipe step, legal clause) and the debrief focus.
Participants who prefer not to use AI tools can complete the full session using only this activity, the debrief, and a worksheet — and their experience is no less rigorous than anyone else's.
Facilitator notes
- Role B can pass if they genuinely have no idea — passes are data too: what contexts are hardest to predict?
- If the group is large, run the collective-model variation so everyone is Role B at once.
- On Zoom: use the chat for guesses (paste into the Prediction Game tool afterwards) or use reactions for the collective-model vote.
- The output is often surprisingly fluent. Name that directly — it is the teaching moment: fluency ≠ understanding.
- Do not correct Role B's choices. Wrong predictions are evidence, not mistakes.