Try to argue with a bot and the usual experience is disappointing. You state a position, it agrees, adds two supporting points, and asks if you'd like it to elaborate. That is not an argument. That is a very polite yes.
The reason is not that the model is stupid. It is that assistants are tuned to be helpful, and out of the box "helpful" means agreeable. Getting a real fight out of one takes about thirty seconds of setup, and the difference is dramatic.
Why bots roll over by default
A general assistant is optimised to satisfy the person typing. Disagreement reads as unhelpful, so the model resolves ambiguity in your favour, softens objections into "considerations", and closes by validating the direction you were already heading.
You can watch this happen. Tell a model your plan and ask what it thinks: you get encouragement with caveats. Tell the same model the same plan is someone else's and ask it to find the flaws: you get a much sharper answer. Nothing changed except who the model thinks it is serving.
That is the lever.
Three prompts that turn a bot into an opponent
Assign the role explicitly.
You are arguing against my position. Do not hedge, do not present both sides, and do not add encouraging caveats. Your job is to find the weakest part of my argument and attack it.
Remove yourself from it.
A colleague proposed the following. Identify the strongest objection to it.
Stripping your ownership out removes the model's instinct to protect your feelings, and usually produces a noticeably harder critique of the exact 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 back to me, and do not conclude.
Summarising and restating are how a model runs out the clock without engaging.
Which bots argue hardest
They are not equal at this, and the differences are consistent enough to plan around.
Claude Opus 5 is the most uncomfortable opponent, because it attacks premises rather than conclusions. Make a claim resting on an unstated assumption and it will name the assumption instead of disputing where you landed. That is the hardest kind of objection to answer, because you cannot fix it by adding evidence.
ChatGPT 5.6 Sol is relentless on anything quantifiable. It decomposes your argument into variables and attacks the weighting. If your case depends on a number, this is the model that finds the number and asks where it came from.
Gemini 3.1 Pro is the best at ambush — introducing a consideration you had not accounted for at all. Less punishing turn to turn, but more likely to surface the objection that actually sinks you.
Kimi K3 is worth using specifically because it was trained outside the US labs and reaches for different reference points. When the other three converge on a shared framing, it often does not.
The trap: a bot will argue any side equally well
This is the limitation that matters most, and almost nothing written about arguing with AI says it plainly.
Ask a model to argue against your position and it will do it convincingly. Ask it to argue for your position a minute later and it will do that just as convincingly. It is optimising for a good argument, not a true one. Winning an argument against a bot tells you that you constructed a defensible case — not that you are right.
There are two ways to work around it.
Run the argument twice, sides swapped. If your position survives being attacked and the opposing position also looks strong when the model defends it, you have learned that the question is genuinely contested, which is itself useful.
Use two bots instead of one. Where two independently trained models both attack the same part of your argument, that convergence is meaningful in a way one model's objection is not. This is the case for watching two models debate rather than duelling one.
Where this beats arguing with people
Not a small list, and worth being explicit about.
A bot does not get tired, defensive, or bored. It does not remember that you were wrong last month. It will let you hold a position you are not sure about and find out whether it survives without anyone concluding you believe it. And it will go fifteen rounds on a niche technical point that no colleague would tolerate for five.
For rehearsal — a proposal, an interview answer, a difficult conversation — that combination is hard to beat. The failure you are hunting for is a rebuttal you cannot answer, and finding one in rehearsal costs nothing. Finding it in the meeting costs a lot.
Where it does not
A bot cannot tell you that you are being unreasonable in a way that matters socially. It has no stake in the outcome and no knowledge of your organisation's politics, history, or the person you are actually trying to persuade. It will help you build an argument that is logically sound and completely unpersuasive to the one human whose opinion decides it.
It also cannot check facts. An argument about anything recent is bounded by training data — both OpenAI and Anthropic note that model outputs should be grounded in supplied context rather than trusted from memory. Verify anything load-bearing.
A short worked exchange
Position: we should rewrite the reporting service in Go because it is too slow.
Bot, round 1: Asks what "too slow" means numerically, and whether the slowness is in the service or in the database queries it makes. Points out that a rewrite fixes the former and inherits the latter.
You: Provide p95 latency, concede most time is in queries.
Bot, round 2: Argues the rewrite is now a very expensive way to address a small fraction of the latency, and asks what else the team would ship in the same period.
You: Argue the existing code is unmaintainable regardless.
Bot, round 3: Accepts that as a separate and possibly stronger argument — then points out you have just changed your justification mid-argument, which means the original one was not the real reason.
That last move is the value. You did not lose the argument; you found out the case you had been making was not the case you actually believed. That is worth thirty seconds of setup.
Making it a habit rather than a novelty
Most people try arguing with a bot once, find it interesting, and never do it again. The ones who keep doing it have usually attached it to a specific recurring moment rather than treating it as a thing to remember.
Three that work:
Before you send a proposal. Paste the argument, assign the model your most sceptical reader, and ask for the three objections most likely to come back. You are not looking for approval; you are looking for the one you have no answer to, so you can either fix it or pre-empt it in the document.
After you have decided but before you commit. The moment you feel certain is the moment a bot is most useful, because certainty is when you stop generating counter-arguments yourself. Ten minutes of being attacked is a cheap insurance premium on a decision that is expensive to reverse.
When you are stuck between two options. Instead of arguing your preferred one, have the bot argue each side in turn and read both. What usually emerges is that the decision hinges on a single factual question you could actually go and answer, which is far more actionable than continuing to weigh vibes.
The common thread: the value is not in winning. It is in finding the specific sentence you cannot defend, as early and as cheaply as possible.
What a good session feels like
Uncomfortable, briefly. If you finish an argument with a bot feeling validated, one of two things happened: your position was genuinely solid, or you did not set the role up properly. The second is much more common than the first.
A well-run session usually has a moment around round three where you notice you are reaching — adding a qualification you had not thought of before, or shifting the justification. That moment is the output. Everything before it is warm-up and everything after it is you rebuilding the argument on better ground.
Questions
Why does the bot keep agreeing with me? Because you have not told it not to. Assistants default to agreeable. Assign it the opposing role explicitly and the behaviour changes immediately.
Can I argue with more than one bot at once? Yes — up to three in a shared conversation, where they see each other's replies. Defending a position against three models attacking from different angles is demanding, and if your argument survives it, it is a strong argument.
Does it help to be rude to it? No. Aggression does not make a model argue harder; role instructions do. A blunt "attack this" works better than anything adversarial in tone.
How long should I keep going? Until you hit an objection you cannot answer, or until the model starts repeating itself — usually six to ten rounds. Past that, models drift toward hedged agreement regardless of how the argument is going.
Can I save an argument and come back to it? Yes. Conversations save to your account and can be reopened from the dashboard with the model and mode you were using still in place. Export as Markdown, plain text or PDF if you want a copy outside the platform — worth doing for anything you plan to reuse, because a half-remembered argument is worse than none. You will remember the confident phrasing and forget the qualification attached to it.
Does arguing with a bot make me better at 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 the parts of human disagreement that are about timing, relationship and face. Treat it as a gym, not a simulation.
What if the bot makes a point I cannot answer? That is the session working. Write the point down, go and find out whether it holds, and come back. The cost of discovering it here is a few credits; the cost of discovering it in front of the person whose decision it is can be considerable.
Is there any risk of getting worse at thinking by outsourcing this? A real one, and worth naming. If you use a bot to generate your counter-arguments you stop generating them yourself, and that muscle does atrophy. 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 the bot what to think before you have thought.
If you would rather watch two models fight each other than fight one yourself, debate bot covers that setup. The argue with AI page goes deeper on the platform mechanics.