Loading
Two neighbours argue about a new parking rule. One says the rule is greedy. The other says the rule is fair. They talk for an hour and neither moves. A week later, one of them finds out what the other actually meant. It was not the thing she spent the hour beating.
She won. She also never met the argument.
A straw man is a weak copy of someone's view. It is easy to knock down. You have met that idea before. This module goes the other way. You build the strongest version of their view first. Then you argue with that one.
Two words carry the whole move. Strongest means the best case there is. Honest means the person would recognise it as theirs. Drop either word and the move stops working.
That is the trap in one sentence. The version that convinces you is built from your reasons. Theirs may rest on something you do not weigh at all. Build it your way and you answer yourself in their voice.
There is one check, and it is short. You say their case back to them. Then you watch their face. A yes means you have it. Almost, or not quite, means you do not.
Who decides whether your version of their view is good enough?
The practice is much older than the word. Anatol Rapoport was a game theorist. In 1960 he wrote a book called Fights, Games, and Debates. He treated those three as different things. In a fight you remove your opponent. In a debate you have to reach them.
On his account, a debate opens by stating the other side's case.
The philosopher Daniel Dennett turned Rapoport's idea into four steps. He put them in a 2013 book about thinking tools. Here is the first one.
“You should attempt to re-express your target's position so clearly, vividly, and fairly that your target says, “Thanks, I wish I'd thought of putting it that way.””— Daniel Dennett, Intuition Pumps and Other Tools for Thinking (2013), giving the first of Rapoport's rules
Three steps follow the first one, and the order is the point.
Put Dennett's four steps back in order.
Put these in order
Naming your agreements feels like handing over ground. It is the step people skip most often.
It feels that way because agreement sounds like retreat. But the points you share were never in dispute anyway. Naming them shrinks the argument to the part that is live. That is the only part worth an hour.
The word itself is young, and it came from the internet. In 2003 someone on a forum made a joke. This straw man is so big, he said. Soon it will turn into a steel man. The joke spread on its own. Nobody set out to name a technique.
Nine years later, a blogger gave the joke a definition.
“the art of addressing the best form of the other person's argument, even if it's not the one they presented”— Chana Messinger, “Knocking Down a Steel Man: How to Argue Better” (2012)
One clause there does all the damage. Even if it is not the one they presented. That clause is the freedom in this move. It is also the trap.
The freedom is real. People state their own views badly all the time. They pick the wrong example. They reach for the loud reason instead of the true one. Cleaning that up is a kindness.
The line to watch is simple. You may fix how they said it. You may not fix what they meant.
A parent says phones should be allowed in class. Here are six ways of saying her view back to her.
Sort each restatement of the parent's view.
Tap an item, then a bucket
Almost everyone believes they understand the other side. That belief is free. It costs nothing to hold and nothing to check. So most of the time nobody checks it.
A real check needs a judge who is not you.
In 2011 the economist Bryan Caplan proposed one. He borrowed the shape of the Turing test. In that test a machine tries to pass for a person.
“Put me and five random liberal social science Ph.D.s in a chat room. Let liberal readers ask questions for an hour, then vote on who isn't really a liberal.”— Bryan Caplan, “The Ideological Turing Test”, EconLog, 20 June 2011
He was answering a claim that one side understood the other better. His answer turned the claim into something a person could lose.
You pass by writing their case so well they claim it as theirs. Being fair to them is not enough. Being polite about them is not enough. You pass by being mistaken for one of them.
Which is harder: attacking a view, or writing it so they claim it?
For years this stayed a thought experiment. Then a group of British researchers actually ran it. They published the results in 2025. They picked three arguments people in Britain were really having. Veganism. Covid vaccines. The vote to leave the European Union.
Six hundred people wrote the arguments. That was a hundred from each side of each topic. Each person wrote three reasons for their own side. Then they wrote three more for the other side.
A different twelve hundred people did the rating. They saw the arguments one at a time. They did not know who wrote what. An argument passed on one condition. Opponents had to agree with it as much as their own side's.
A person writing for the other side has to fake it. So you might expect the raters to catch them.
Of every 100 arguments written for the other side, how many passed?
These were three arguments the country really did fight about. Most of the borrowed arguments still passed. The researchers read that as good news, and it is.
It also means the misses are worth a look. Between three and five arguments in every ten did not pass.
People often claim their own side understands the other one better. That claim is what started this whole test in 2011.
On the main measure, both sides passed at about the same rate. The belief survives because we only ever watch our own side try. Their trying happens somewhere we are not.
The people who passed rated their opponents differently. They were less likely to call them ignorant, immoral or irrational.
That is a link between two things, not a cause. Kinder people may find the test easier. Or doing the work may soften how you see them. This study cannot tell those two apart.
The full test needs strangers and ratings. The small one needs one person. You write their view in four sentences. You send it to them. You ask whether you got it right.
The answer comes back in a word or two. Anything short of yes is information.
The replies fall into a few shapes.
Match what they say back to what it tells you.
Tap one side, then its match
There is a version people can feel from across the room. You say you are steelmanning. Then you hand someone an argument they do not recognise. The label made it worse, not better.
A restatement they accept needs no label. A label cannot rescue one they reject.
The obvious fix is to imagine being them. Researchers tested that fix directly. They ran twenty-five experiments. Some asked people to read a face. Some asked them to guess what a husband or wife would like.
Told to take another person's view, what happened to accuracy?
In the same line of work, one thing did raise accuracy. People asked. They put a question to the other person and heard the answer. Their guesses got better.
Asking is slower than imagining. It needs another person and some nerve. It is also the only one of the two that worked.
It sits with the people who hold it. The best of them have written it down somewhere. That writing is what you are looking for.
Reading the loudest person in your feed is not reading the other side. It is reading whoever your feed rewards.
A neighbourhood is fighting about a new bike lane. You want the best case against it.
Which of these gives you their case?
Tap an item, then a bucket
A useful steelman is short. What they are trying to get. Their main reason. The thing they are afraid of. The evidence they would point at.
Which of these does not belong in a steelman?
Here is a version of an argument families really have.
Your brother will not put money into a pension. You think he is being reckless with his future.
The steelman changed the argument itself. It started as reckless against responsible. It ended as locked money against money he can reach.
The second argument has an answer somewhere. The first one only had a winner.
This is the point where a module usually promises results. So here is the state of the evidence instead. Steelmanning itself has barely been studied.
Nobody has run a proper trial with a comparison group. What has been tested are the things next door to it.
In 2013 researchers tried something close. They asked people to explain how a policy actually works, step by step. Not why they liked it. How it works.
What happened to how strongly people held their position?
That gap is the useful part. Listing reasons is rehearsing what you already feel. Explaining a mechanism makes you notice the steps you cannot fill in.
In 2021 two researchers ran that study again, three times over. They filed their plans in public before collecting any data. None of the main effects turned up.
One piece held. People still rated their own understanding lower after trying to explain the steps. Their positions simply did not move with it.
So the humbling was real. The softening was not, at least not there.
In 2022 one of the original authors tried again, with newer topics. This time the softening showed up. But only on some of the issues.
Which issues split, and how they split, turned out to be the finding.
On issues people treat as trade-offs, explaining the steps softened positions. On issues held as protected values, nothing shifted. Same exercise, same room, two different results.
The honest summary is short. Working out how a thing actually works can loosen a position. It seems to work on some kinds of position and not others. And it has failed once in careful hands.
Not necessarily, and not necessarily a true one either. A failed repeat means the effect is shakier than it first looked. Sometimes it is smaller than reported. Sometimes it only shows up under certain conditions. The 2022 result points at conditions.
None of this shows that steelmanning changes the other person. That was never what got measured. Every result here is about the person doing the explaining.
When the exercise moved anything, it moved the explainer. That is a smaller claim, and a more useful one.
Read that again, because it is the whole module. Building the other side's case is work aimed at you.
That reading survives because it does sound like a good debating move. But a version built to disarm someone is built for your purposes. It is not built from theirs. It is a straw man that took lessons.
Done honestly, the move has a cost. You might build their case and find that it holds. Then you have to say so out loud. The person was arguing with you.
A tool that can only ever confirm you is not a tool. It is a mirror.
You can do all of this and still not know whether it worked.
You steelman a view you disagree with. Which ending shows you actually did it?
Here is the limit, and it is not a small one. Some positions do not have a strong version. Building one for them anyway is its own mistake.
You have met false equivalence before. This is how a careful person walks into it politely.
A steelman is missing for one of three reasons, usually.
Sort each case by why the strong version is missing.
Tap an item, then a bucket
Some questions have already been measured. The strongest version of a wrong answer is still wrong. Dressing it up does not reopen the question. It only makes a settled one look open.
The fair move there is to say what was measured, and how.
Some positions are not arguments at all. They are wants with an argument tied on afterwards. Steelmanning the argument leaves the want completely untouched.
The question that helps is not what is their best case. It is what does this person get.
Sometimes you try honestly and find nothing. That result counts for something. It is not proof that the other side has no reasons.
A failed steelman produces exactly that feeling. It is the worst moment to trust yourself. You just spent an hour inside your own version of their head. Asking them is the cheap fix. It is also the step people skip.
So the practice is narrow, and it is honest about itself. It does not win arguments. It does not repair the other person. It has no trial behind it.
It gives you one thing. A version of their view good enough to catch you. If you are wrong about them, you find out.