Most advice on how to argue with ai skips the only part that matters. The models are perfectly capable of arguing. They just will not do it with you unless you tell them to, because an assistant tuned to be helpful reads disagreement as failure.
So the first move is not a clever prompt. It is removing the model's incentive to agree with you.
Why it defaults to agreeing
An assistant is optimised to satisfy the person typing. That objective quietly shapes everything: it resolves ambiguity in your favour, converts objections into "considerations", and closes by endorsing the direction you were already going.
You can demonstrate this to yourself in thirty seconds. Describe a plan as yours and ask what the model thinks — you get encouragement with light caveats. Open a new conversation, describe the identical plan as a colleague's, and ask for the strongest objection. The critique gets noticeably sharper. Nothing changed except who the model believes it is serving.
That gap is the whole technique.
Setting up an actual opponent
Three instructions do most of the work.
Assign the role explicitly. "You are arguing against my position. Do not hedge, do not present both sides, and do not add encouraging caveats. Find the weakest part of my argument and attack it."
Take yourself out of it. "A colleague proposed the following. What is the strongest objection?" Removing your ownership removes the instinct to protect your feelings, and reliably produces a harder critique of the same idea.
Ban the retreat. "If you concede a point, say so in one sentence and move immediately to your next strongest objection. Do not summarise, do not restate my position, do not conclude." Summarising and restating are how a model runs down the clock while appearing to engage.
The four moves that keep it productive
Name the dodge. "You did not answer the cost objection. Answer it before anything else." Models are skilled at responding fluently while changing the subject, and the point that got quietly dropped is usually the one that was hardest.
Attack the premise, not the conclusion. Ask directly: "What am I assuming that I have not stated?" Arguments rarely fail at the conclusion. They fail at an unexamined premise, and a model asked to hunt for one is good at it.
Force a commitment. "One sentence, no qualifications: what would you actually do?" Use this at the end. It cuts through the hedging that models produce when the honest answer is "it depends".
Swap sides. Have it argue your position as hard as it argued against it. If both cases look strong, you have learned the question is genuinely contested — which is worth knowing and is not what you get from a single exchange.
A worked example, start to finish
Abstract advice is easy to nod at and hard to apply, so here is a full exchange in compressed form. The position under attack: we should drop our free tier because it costs more in support than it converts.
Opening instruction — role assigned, ownership removed, retreat banned:
A colleague argues their company should drop its free tier because support costs exceed conversion value. Argue against this as forcefully as you can. No hedging, no balanced overview, no encouraging caveats.
Round one comes back with the obvious objections: free users generate word of mouth, the funnel needs a top, removing it is hard to reverse. Predictable, and none of it lands hard. This is normal — the first response is almost always the argument you already anticipated.
Round two is where the work happens. Rather than answering each point, name the frame: "You are assuming the free tier is the only acquisition channel. Argue as if it is one of four." The response sharpens immediately, because the strongest version of an argument usually appears only after you strip away the version that depends on an easy assumption.
Round three — the premise hunt: "What am I assuming that I have not stated?" Here it surfaces something genuinely uncomfortable: the claim that support costs exceed conversion value assumes support cost is attributable per-tier, which most companies cannot actually measure. That is not a counter-argument. It is a hole in the evidence, and it is the single most valuable output of the session.
Round four — the commitment: "One sentence, no qualifications: what would you do?" The answer will be some form of "measure attributable support cost for one quarter before deciding", which is more useful than either original position.
Four rounds, maybe eight minutes. The result is not that the model won or lost — it is that a decision which felt settled turned out to rest on a number nobody had.
The trap almost nobody mentions
A model will argue any side equally well.
Ask it to demolish your position and it will. Ask it to defend the same position a minute later and it will do that just as fluently. It is optimising for a good argument, not a true one. Winning an argument against a model tells you that you constructed a defensible case. It does not tell you that you are right.
There are two honest ways around this.
Run the argument twice with sides swapped and read both. Or — better — put two models from different providers on it and watch where they independently attack the same joint. Two models trained by different labs converging on the same objection is meaningfully stronger evidence than one model raising it, because they do not share the same blind spots. That is the case for watching two models argue rather than duelling one.
What a good session feels like
Uncomfortable, briefly, around round three.
If you finish feeling validated, one of two things happened: your position was genuinely solid, or the role setup did not take. The second is far more common. A well-run argument has a moment where you notice you are reaching — adding a qualification you had not thought of, or quietly shifting your justification.
That moment is the output. Everything before it is warm-up; everything after is you rebuilding on better ground.
When to stop
Six to ten rounds. Value is front-loaded and falls off sharply.
Past about ten exchanges, models drift toward hedged agreement because that is what they are tuned to produce. A conversation running to round fifteen is not deeper than one stopped at eight — it is longer and more polite. If it is still productive at twelve, that usually means you intervened well rather than that the model found new ground.
Three habits that make it stick
Most people try this once, find it interesting, and never repeat it. The ones who keep going attach it to a specific recurring moment.
Before sending a proposal. Paste the argument, assign the model your most sceptical reader, ask for the three objections most likely to come back. You want the one you cannot answer, so you can fix it or pre-empt it.
After deciding, before committing. The moment you feel certain is when you stop generating counter-arguments yourself. Ten minutes of being attacked is cheap insurance on a decision that is expensive to reverse.
When stuck between two options. Rather than arguing your preference, have the model argue each side in turn. What usually emerges is that the decision hinges on one factual question you could go and answer — far more actionable than continuing to weigh instincts.
Four ways people get this wrong
Arguing to win. The instinct is to defend your position until the model concedes. It will concede — models concede readily under sustained pressure — and you will have learned nothing except that you can out-stubborn a system tuned to be agreeable. The useful posture is trying to find the objection you cannot answer, which means welcoming the moment it lands rather than deflecting it.
Feeding it a summary instead of the argument. A three-line précis gets you a three-line critique of a strawman. Paste the actual reasoning, including the parts you are least sure about, because those are the parts worth attacking and the parts a summary quietly drops.
Accepting the first response as the model's real position. Round one is nearly always the conventional objection set. The model has not engaged with your specifics yet; it has pattern-matched to the topic. Everything useful is downstream of round two.
Letting it be polite. If the replies open with "That's a thoughtful point" or close with "Ultimately, both approaches have merit", the role instruction has decayed and you are back to talking to an assistant. Reassert it in one line — "You are still arguing against me. Continue." — rather than starting over.
What it cannot do for you
It has no stake in the outcome and no knowledge of the person you are actually trying to persuade. It will help you build something logically sound and completely unpersuasive to the one human whose opinion decides the matter.
It also cannot check facts. An argument about anything recent is bounded by training data. Both OpenAI and Anthropic advise grounding claims in supplied context rather than trusting model memory, and that applies with full force to anything load-bearing in your argument.
And there is a real risk of outsourcing the wrong half. If you use a model to generate your counter-arguments, you stop generating them yourself, and that muscle does weaken. The version that helps is arguing your own case as hard as you can first, then handing it over to be attacked. The version that hurts is asking what to think before you have thought.
Common questions
Does being blunt or rude make it argue harder? No. Aggression changes nothing; role instructions change everything. A flat "attack this" outperforms anything adversarial in tone.
Can I argue with more than one model at once? Yes — up to three in a shared conversation, each seeing the others' replies. Defending a position against three models attacking from different angles is demanding, and an argument that survives it is a strong argument.
What if it makes a point I cannot answer? That is the session working. Write the point down, go find out whether it holds, come back. Discovering it here costs a few credits; discovering it in front of the decision-maker costs considerably more.
Is arguing with a model good practice for arguing with people? Partly. It sharpens the structural side — spotting unstated assumptions, noticing when a justification shifts, holding a line under pressure. It does nothing for timing, relationship or face, which is most of what human disagreement actually involves. Treat it as a gym, not a simulation.
For the mechanics of making a model behave like an opponent rather than an assistant, argue with a bot goes further. If you are training for competition, AI for debate practice has structured drills.