Model UN prep with AI is genuinely useful for about two thirds of the work and actively misleading for the rest, and knowing which is which is most of the value.
You get your country and committee in late September. The conference is in November. In between sits a position paper, a policy you have to defend whether or not you agree with it, and a moderated caucus where thirty delegates compete for the chair's attention. This is a six-week countdown through all of it.
Week 6 — Country and committee
The first job is not research. It is working out what your country's actual position is, which is frequently not what you would expect and occasionally not what your country says publicly.
Start broad and adversarial:
I am representing [country] in [committee] on the topic of [topic]. Lay out this country's real policy position, including where its stated position and its voting record diverge. Do not soften anything.
The divergence question is the one that matters. A delegation that knows its country votes against a resolution it rhetorically supports has something to work with in caucus that nobody else has.
Then verify all of it. Voting records live at the UN Digital Library, and a model will state a voting record confidently whether or not it has one. Treat the AI output as a list of things to check, not as the answer.
Week 5 — The position paper
Position papers are formulaic by design, which makes this the phase where AI helps most and where the temptation to misuse it is strongest.
What works: giving it your draft and asking it to attack. "You are a chair reading four hundred position papers. What in this one is generic enough that you would skim it?" The answer is usually the second paragraph, and it is usually right.
What also works: asking for the standard structure of a position paper for your specific committee, which varies more than most delegates realise between a General Assembly committee and a crisis or specialised body.
What does not work, and will be noticed: having it write the paper. Chairs read hundreds of these and the tell is not vocabulary — it is that an AI-written paper argues the position without ever committing to a specific proposal your delegation would actually table. If your paper could be submitted by four other countries with the name swapped, it is not finished.
Week 4 — Bloc mapping
This is where multi-model work earns its place, and it is the phase most delegates skip.
Ask one model to build the case for your bloc's position. Ask a second — from a different provider — to build the case for the opposing bloc. Read both, and pay attention to where they identify the same pressure point from opposite sides. That point is where the resolution will actually be negotiated.
| Task | Single model | Two models arguing |
|---|---|---|
| Learning your own position | Fine | Unnecessary |
| Finding the opposition's best argument | Weak — it hedges | Strong |
| Locating the real clash | Poor | This is the use case |
| Predicting who will merge blocs | Speculative either way | Slightly better |
The reason the two-model version works is not that the models are smarter in pairs. It is that a single model asked for "both sides" produces a balanced summary in which nothing is at stake, whereas two models each told to win produce an argument with a visible fault line.
Week 3 — Moderated caucus drills
Moderated caucus is the part of MUN that most resembles competitive debate, and it is drillable.
Set up an opponent explicitly: "You are the delegate of [opposing country]. You have sixty seconds to respond to my speech. Attack my proposal's funding mechanism. Stay in character and do not break to give me feedback."
Run it against three or four different opposing countries. What you are training is not argument quality — it is the reflex of hearing an objection and having a response ready before the chair moves on, which is a timing skill and only comes from repetition.
A second drill worth more than it looks: give a model your one-minute speech and ask it to summarise what you proposed in one sentence. If the summary is vague, your speech was vague, and thirty delegates half-listening will have got even less than the model did.
Week 2 — Resolution drafting
Operative clauses have a grammar of their own, and models handle the form well and the substance poorly.
Useful: checking that your clauses use correct operative phrasing, that you have not duplicated a clause in different words, and that the resolution actually addresses the topic in the committee's mandate rather than drifting into another body's remit.
Not useful: generating clauses. AI-generated operative clauses tend toward the unobjectionable — "urges member states to cooperate" — which is exactly the kind of clause that survives to the final draft while accomplishing nothing and impressing nobody.
The productive version is writing your clauses yourself, then asking: "Which of these clauses would a sceptical delegation object to, and on what grounds?" Anything nobody would object to is probably not worth a clause.
Week 1 — Crisis readiness and the delegate you have to be
If you are in a crisis committee, the preparation is different: you cannot research the crisis, because it has not happened yet. What you can do is build the reflex.
Feed a model an invented crisis in your committee's domain and give yourself ninety seconds to write a directive. Then have it play the crisis staff and tell you what went wrong. Repeat with a different scenario. Five rounds of this is worth more than any amount of background reading, because crisis is entirely about response speed under incomplete information.
For standard committees, week one is speech rehearsal, which AI cannot help with at all, and re-reading your own position paper so that the specifics are actually in your head rather than in a document.
Committee-specific notes
The countdown above is written for a standard General Assembly committee. Three variants change the emphasis enough to be worth calling out.
Security Council. Fifteen delegations, veto politics, and far more speaking time each than a hundred-delegate GA committee. Preparation shifts from breadth to depth: you will be asked to defend specifics repeatedly, and vague policy knowledge is exposed within minutes. The most valuable drill here is not caucus speed but sustained cross-examination — have a model interrogate your delegation's position for ten straight exchanges without letting you change the subject. If you are a P5 delegation, spend an evening on the actual history of your veto use; models summarise this badly and delegates who know it precisely have an enormous advantage in negotiation.
Specialised agencies and ECOSOC bodies. The mandate constraint is the whole game. Half the resolutions that fail in these committees fail because they propose something the body has no authority to do. Before drafting anything, ask a model to state your committee's mandate and then verify it against the body's founding document — and treat any clause that strays outside it as dead on arrival, however good the idea.
Historical and crisis committees. The cutoff problem inverts here: the model knows the period well, and knows how it turned out, which is a specific hazard. A delegate arguing from hindsight sounds anachronistic and chairs notice. Instruct it explicitly to reason only from information available at the committee's start date, and expect to remind it more than once.
Non-English committees. If your conference runs a committee in a language other than English, model quality drops noticeably in both policy nuance and formal register. Draft in the committee language rather than translating from English — translation produces text that is grammatically correct and diplomatically wrong.
Where AI will get your country wrong
Four failure modes, in order of how often they cause a problem.
It defaults to the Western framing of the topic. Models trained predominantly on English-language sources will characterise a contested issue the way the anglophone press does. If you are representing a country whose position is minority or oppositional, the model will subtly argue against you while appearing to help. Counter it by naming the framing explicitly: "Present this topic as [country]'s foreign ministry would frame it, not as Western media frames it."
It confuses stated policy with actual behaviour. Governments say one thing at the UN and do another. Models tend to report the stated position because that is what is written down.
It invents specifics. Resolution numbers, treaty dates, vote tallies. These are exactly the details that make a speech credible and exactly the details a model is most likely to fabricate. Every one of them gets checked against a primary source.
It has a cutoff. For any topic that moved in the last year — and MUN topics are chosen because they are current — the model's picture is stale in ways it will not flag.
Things that go wrong at conference
- Reading the position paper aloud. It is a document, not a speech. Delegates who present arguments rather than recite paragraphs get called on more.
- Over-preparing the opening, under-preparing the third caucus. Everyone rehearses their opening speech. Almost nobody rehearses the point where the bloc has fractured and someone has to hold it together.
- Defending the AI's version of your position instead of your own. If your knowledge of your country came entirely from one conversation, the first delegate who knows the region properly will expose it.
- Ignoring the chair's actual instructions. No amount of preparation compensates.
Questions delegates ask
Is using AI for MUN prep against the rules? Preparation is not regulated at any conference we are aware of. In-session assistance is increasingly addressed explicitly in conference handbooks and the direction of travel is restrictive — check your specific conference, and check it again the week before, because these policies have been changing every season.
Will chairs be able to tell if my position paper was AI-written? Often, yes — not from style but from the absence of anything specific to your delegation. Awards go to papers that commit to a concrete proposal, and generated papers rarely do.
Can I practise moderated caucus alone with this? Yes, and it is the single best use here. Set a timer, put a model in character as an opposing delegation, and drill the sixty-second response. It is the closest thing to real caucus pressure that solo prep can produce.
My country's position is one I personally find indefensible. Does that change anything? It changes the prompting, not the ethics — representing a position is the exercise. The practical problem is that models are safety-tuned to resist arguing certain positions persuasively, so you may get hedged, weakened versions of your own delegation's case. Framing it as academic representation of a documented state position rather than as personal advocacy generally gets you the substantive answer, and where it does not, primary sources — the delegation's own statements in the UN record — are the better route anyway.
How much of this should a first-time delegate attempt? Weeks six, five and three. Skip bloc mapping and resolution drafting on a first conference; they matter, but a novice gains far more from knowing their country cold and being able to speak for sixty seconds without freezing. Adding four workflows at once usually means none of them get done properly.
Is it worth using more than one model for this? For research, marginally — a second model catches the occasional confident error. For bloc mapping, substantially, for the reason in week four: one model asked for both sides gives you a balanced summary with nothing at stake, while two models each told to win produce a visible fault line. That fault line is where the resolution gets negotiated.
The drills here assume a Model UN format specifically. For parliamentary, policy or Lincoln-Douglas debate, AI for debate practice covers the format-specific differences, and AI for debate research goes deeper on verifying what a model tells you before it reaches a round. For the prompts that make a model behave like an opposing delegation rather than an assistant, see AI debate prompts, and for the bloc-mapping technique in week four, multi AI debate covers what putting two models against each other actually produces.