This chapter is about the grand simulations: Model UN, moot court, Reacting to the Past, the Yalta Conference re-enactment, deliberative mini-publics. It opens with a warning, because this is the one use case where enthusiastic adoption goes wrong: do not replace these with AI. Their value is thirty humans negotiating under rules, reading the room, forming coalitions, performing in public. Recall Chapter 2’s distinction: AI is spectacular at role play and a supporting player in simulations. The opportunity here is the supporting role, and it is bigger than it sounds.
What these simulations do, and where they leak
The formats are pedagogically excellent. Duke Kunshan’s catalog shows the range: a COP15 biodiversity negotiation with 30 students across 10 role categories holding collective and secret information packages; a Yalta simulation where students played the US, UK, and USSR with shared and secret documents across two class meetings; moot courts where students argue for clients before student judges; mini-publics where students deliberate policy as citizens and experts; scholarly dialogues where students embody academic authors and argue as them. Reacting to the Past has built an entire pedagogy movement on semester-long historical games.
But every one of them leaks value at the same three joints:
Preparation is unrehearsed. A student assigned “delegate from Brazil” does research, writes a position paper, and then speaks extemporaneously for the first time at the event itself. The event is their first rep, which is exactly the failure mode this playbook exists to fix. The instructor cannot spar with 30 delegates individually.
The event is one-shot. The Yalta simulation takes two class meetings; a student who fumbles their opening speech at minute ten does not get a second Yalta. High preparation cost, single performance: the worst possible ratio for skill building, however good it is for engagement and empathy.
Roles outnumber students, or students outnumber good roles. Someone must play the recalcitrant judge, the hostile journalist at the press conference, the fifth family member. Either a student burns their learning experience on a functionary role, or the role goes unplayed and the simulation loses texture.
Pattern 1: The sparring partner
The highest-leverage move is upstream of the event. Give every participant a private AI counterpart matched to their role, for the week before:
- The Model UN delegate negotiates bilaterals against an AI playing the delegations they will face, testing their position paper against real pushback before it costs them anything at the podium.
- The moot court advocate argues before an AI judge configured to interrupt with the exact style of cold-bench questioning the professor wants drilled: “Counsel, that’s a policy argument. Where in the statute?”
- The Reacting to the Past student rehearses their faction speech against an in-character opponent who knows the game’s source documents.
- The mini-public participant practices articulating a policy position against a citizen persona who disagrees from a coherent, researched perspective rather than a strawman.
This is the batting cage before the game, and it changes the event itself: students arrive having already survived contact with opposition, so the live simulation starts at a higher level of play. Instructors who run this report the quietest structural benefit: the students who used to freeze — the ones for whom one-shot public performance was a fairness problem — arrive with reps behind them.
Pattern 2: The missing stakeholder
During the event, AI fills the roles humans shouldn’t burn a seat on. A projected persona on the classroom screen (the group-shape choice from Chapter 3) can be the expert witness the whole court cross-examines, the head of state who takes questions at the summit press conference, the community member testifying before the student-run planning commission. The class interacts with it collectively; the persona holds its brief, resists leading questions, and never breaks character in hour two.
The scholarly-dialogue format inverts elegantly: instead of two students emailing as Berger and Gross, a student argues against the AI playing the author they are studying, which forces them to know the text well enough to survive its author’s rebuttals.
Pattern 3: The asymmetric briefing
The secret-dossier mechanic that makes these simulations work (Chapter 10 steals it for scenario design) has an AI-era upgrade: instead of a static information packet, each role gets a briefing conversation. Before the COP15 negotiation, the student playing an industry lobbyist is briefed by an AI playing their CEO, who answers questions, sets red lines, and refuses to reveal what the board hasn’t decided. The student extracts their constraints through questioning rather than reading them off a card, which is itself the skill, and no two students enter the simulation with identically processed information.
What to keep human
Everything else. The floor debate, the coalition-building in the hallway, the deliberation, the vote, the collective debrief where the room compares what each faction knew: these are the simulation. If a proposed AI use replaces a moment where students read other humans in real time, decline it. If it adds reps before that moment or texture inside it, take it.
Exercise
Take a simulation you run (or borrow the Yalta format). Pick the three roles whose players would benefit most from pre-event sparring, and for each, write a two-line spec for their AI counterpart: who it plays, and the one behavior it must exhibit (the judge interrupts; the rival delegate refuses to discuss article 4; the CEO will not approve overspend). Assign three practice rounds in the week before the event, then compare the first ten minutes of this year’s simulation with last year’s. That comparison is your evidence for expanding.