C2: The Steelman Protocol

Which side understands the other?

In a 2011 interview, the economist Paul Krugman said liberals can describe conservative views and conservatives cannot return the favor: "A liberal can talk coherently about what the conservative view is because people like me actually do listen.… The reverse is not true." In June that year the libertarian economist Bryan Caplan replied. He granted that "at the meta-level, he's on to something," but doubted the claim, at least among academics: "I've known many liberal Ph.D.s from top-10 social science programs – and even those who know me best can't articulate my views well." Then he proposed a way to settle it, which he called the ideological Turing test, after Alan Turing's test of whether a machine can pass for human: can you describe the other side's views well enough that its own members cannot tell you from one of them? He wrote that "the beauty of the notion of the ideological Turing Test is that it's a test." Krugman was sure his side was the one that listened. Caplan doubted it and wanted to find out.

Researchers have since measured something close to the question. In a 2012 study, Jesse Graham, Brian Nosek and Jonathan Haidt had 2,212 Americans answer a moral-values questionnaire as themselves or as a "typical" liberal or conservative. Both groups "exaggerated the ideological extremity of moral concerns for the ingroup as well as the outgroup," and in that sample, which the authors note was not representative, "Liberals were least accurate about both groups." So the one claim in this exchange that was stated with confidence, that one's own side listens, is the one this study, as far as it reaches, ran against. The lesson is not about liberals. Being sure that your side understands the other is not evidence that it does, whichever side you are on.

A 2025 study ran Caplan's test for real, though not on left and right. Charlotte Brand, Daniel Brady and Tom Stafford took 1,200 people in the UK on opposite sides of three disputes, Brexit, COVID-19 vaccines and veganism, had each side write the other side's reasons, and had the other side rate them. Between 54 and 71 per cent of the arguments passed, depending on the topic, and in none of the three disputes was either camp more likely to pass. The authors called this "high levels of mutual understanding," while noting that it "remains open to interpretation" whether the figures show "a surprising failure" or "a surprising success."

Now notice what you did with those results. If the first thing you registered was which side came off worse, and only second that both sides misjudged their own side too, you have met the reflex this tool exists to catch. It runs in miniature whenever you have your reply ready before you reach the end of a message you disagree with. When its author writes back with something like that isn't what I said, the likeliest explanation is not that they are backpedaling. You answered a version of their argument that was easier to answer, and you built it yourself while you were still reading.

The move

The Steelman Protocol is a tool, which on this site means a mental operation you can run on a claim in real time. The name plays on the familiar "straw man," the flimsy version of an opposing view that is set up to be knocked down. A steelman is the opposite: the strongest version of the view you disagree with, built carefully enough that it can take a hit.

The idea is older than the name. In On Liberty (1859), John Stuart Mill wrote that "He who knows only his own side of the case, knows little of that," and that a person must hear opposing arguments "from persons who actually believe them." The philosopher Daniel Dennett, in a 2013 book extract, asked "Just how charitable are you supposed to be when criticising the views of an opponent?" As "the best antidote I know for this tendency to caricature one's opponent," he offered four rules he credited to the psychologist and game theorist Anatol Rapoport. The first: "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.'" The second and third are to list any points of agreement and to mention anything you have learned from your target. The fourth: "Only then are you permitted to say so much as a word of rebuttal or criticism." Caplan's ideological Turing test asks the same thing in other words: whether you can "state opposing views as clearly and persuasively as their proponents." The word "steel man" was already in use in 2011 on LessWrong, an online forum about reasoning, where one post defined it as "an improvement of someone's position or argument that is harder to defeat than their originally stated position or argument."

Those are two different standards. The forum's is about improving the other side's argument; Rapoport's rule and Caplan's test are about getting it right. This site's version follows the second. The tagline does not ask whether your version is stronger; it asks whether a thoughtful person who holds the view would read it and say something like yes, that is what I think, and that is why.

Running it takes four steps.

1. Before you reply to a position you disagree with, write it out in a few sentences as its best defender would put it, including the reasons that make it persuasive to the people who hold it.
2. Check your version against the sign-off test. Would a thoughtful holder of the view put their name to it? Do not settle this by imagining their answer. Find out what such people actually say, in their own words.
3. Look for the strongest point in it, the one you find hardest to answer, and say plainly what it is.
4. Only then make your argument, and aim it at that strongest version, starting with the hardest point.

The order matters. If the rebuttal comes first, the summary will be shaped to fit it.

Why it works

The tool belongs to the family the site calls Testing and Calibration, which is about proportioning how strongly you believe something to the evidence for it. So it is not mainly a courtesy to opponents; it is a test of your own position. Rebutting the weakest version of an opposing argument wins conversations and loses accuracy. If you have only ever beaten the weak version of a view, you do not know whether your own view survives contact with the strong one.

The Steelman Protocol is one of five tools this site treats as most important across every subject, on the ground that it is "the one tool that reliably changes the person using it." That is the site's judgment, and the research supports a narrower version of it.

Writing out the other side's arguments does seem to change people. Rhia Catapano, Zakary Tormala and Derek Rucker open their 2019 paper with the sentence "Counterattitudinal-argument generation is a powerful tool for opening people up to alternative views," where the long term means writing arguments for a view you oppose. In their own experiments, simply writing the other side's arguments "seemed to increase participants' receptiveness," according to the journal publisher's press release. Brand's group found that people who passed their test were less likely to call opponents "ignorant, immoral, or irrational," though the link could run either way. A 2025 preprint, not yet reviewed by other scientists, assigned about half of its 203 student participants to argue the side they opposed, in writing or in a live debate. They moved toward that side or toward the middle, and that movement was still there at a follow-up two to six weeks later. The numbers are small (only 25 people wrote from the other side), only 58 per cent of participants came back for the follow-up, and most were liberal. Writing from the other side also softened feelings toward opponents at first, but by the follow-up that effect was no longer detectable.

How you do it seems to matter. Catapano's group compared two ways of writing the other side's arguments, in three preregistered experiments (the predictions and analysis were fixed in advance) with 2,734 participants. People who first tried to imagine a particular opponent's life, interests and intentions, and then wrote his arguments, came away less receptive to his view than people who simply wrote the arguments. The authors' explanation is that arguments written from inside the opponent's perspective are "incongruent with their own values," and the effect was weaker when the imagined opponent shared the writer's overall values. (The paper itself is paywalled; this account comes from its published abstract and the press release.) That finding bears on step 1 above, which asks you to put the view as its best defender would, including the reasons that persuade the people who hold it. In this study, doing that from inside the opponent's perspective left people less open to the view than writing the same kind of arguments as themselves.

And a widely cited finding about a related idea, that explaining how a policy works (rather than stating the other side's case) makes people more moderate, failed three preregistered replications by Jarret Crawford and John Ruscio in 2021; what did replicate is that writing out justifications for beliefs you already hold makes them stronger.

So doing the work often changes how you see the other side, and sometimes your view, but not reliably and not always toward them. Since arguing a side can move you toward it whether or not it is right, a change of mind is not by itself evidence. The tool's purpose is accuracy, not conversion.

Using it well

The legitimate use is building the strongest version of a position you disagree with and arguing against that version specifically. No real case found for this page shows the whole move cleanly, someone stating the other side in words its holders accepted and then arguing against that statement. Two real cases show parts of it.

Jerry Taylor spent 1991 to 2014 at the libertarian Cato Institute "arguing against climate action." In a 2019 essay he explained why he changed his mind. It first ran in The Bulwark, a site relaunched that January, by its own description, as "a voice for non-Trumpian conservatives," and was reposted by the Niskanen Center, the policy advocacy group he then led. In Taylor's account, both sides were arguing about how much warming was most likely, "even though the truth is that neither side can know with certainty which variant will come to pass." He quotes Representative Alexandria Ocasio-Cortez, "The world is going to end in 12 years" (meant, he notes, "figuratively"), as an example of the environmentalists' case. What moved him was a different argument for the same conclusion: "Risk management is not about discerning the optimal response to the most likely outcome. It is about discerning the appropriate response to the most likely distribution of possible outcomes." Several of the reasons he gives for it come, he says, from the economist Richard Tol. He then states an objection from his own former side, that this might "strike some conservatives as a suspicious attempt to smuggle in the 'precautionary principle'", and answers it. On this site's reading, two parts of the tool are visible here: Taylor found a stronger case for the other side's conclusion than its loudest voices were making, and he answered his old side's objection directly rather than waving it away. But he did not rebut the strong case. It changed his mind.

In his 2009 commencement speech at the University of Notre Dame, a visit that had drawn controversy, as he acknowledged, President Barack Obama told of an email from a doctor who called himself "a Christian who was strongly pro-life." The doctor objected to a line on Obama's Senate campaign website promising to fight "right-wing ideologues who want to take away a woman's right to choose." That summary failed the sign-off test, and a holder of the view said so: "I do not ask at this point that you oppose abortion, only that you speak about this issue in fair-minded words." Here the holder of the view did the check that step two asks you to do yourself, and Obama, in his own account, accepted it: "I didn't change my underlying position, but I did tell my staff to change the words on my website." The story is Obama's; the doctor's side of the exchange has not been found.

Taylor changed his position and Obama kept his. The tool asks you to aim at the real target; it does not dictate where you end up.

When it becomes a weapon

The weaponized form is the decorative steelman: a token summary of the other side followed by the identical rebuttal you had already written.

On this site's view, the decorative steelman can be harder to spot than a straw man, because it borrows the look of fairness. It opens with something like to be fair, critics make a reasonable point, offers a sentence or two, and then proceeds exactly as it would have anyway. Mill described something close to it in its classroom form, a student who hears "the arguments of adversaries from his own teachers, presented as they state them, and accompanied by what they offer as refutations. That is not the way to do justice to the arguments, or bring them into real contact with his own mind." It can do more damage than an honest attack, because readers come away believing the other side has been heard and answered when it has not.

Example still being found. a real, quoted, dated instance of a decorative steelman, a brief, formulaic summary of the opposing side followed by a rebuttal that never engages it, preferably one the summarized side publicly said misrepresented them; find instances from both the political left and the political right. Searched in two rounds without finding a case that clears that bar; reported to Karl in the reading record.

Critics of steelmanning have described a related danger: the steelman that replaces the view it claims to strengthen. The writer Ozy Brennan argued in 2016 that recasting an opponent's view in your own terms ends up "transforming everyone into stupider versions of yourself." Holden Karnofsky, then of the grant-making foundation Open Philanthropy, wrote in 2017, from the receiving end: "I find it very unsatisfying to engage with 'steel-man' versions of my arguments, which rarely resemble my actual views." Rob Bensinger put it in a line in 2018, which he repeated in 2022: "you shouldn't 'steelman' that belief by replacing it with a different one." The economics blogger Noah Smith, crediting Brennan, made the same case in 2024: "a critic of an idea is inherently an unreliable guide to the best version of an idea."

Karnofsky's and Bensinger's complaints land on the forum definition, improving someone's argument whether or not they made it. (Defenders, such as a commenter on Bensinger's post, reply that when the question is whether a conclusion is true, rather than what a particular person believes, improving the argument is the point. They are right about that question.) When you are answering a person or a side, the sign-off test meets those two complaints: a version the view's holders would reject is not a steelman, however generous it looks to its author.

Brennan's critique reaches further, to the sign-off test itself. Brennan anticipates the reply "Real steelmanning is being able to put other people's viewpoints in words they themselves find more compelling than their own arguments!" and answers that "that is an extraordinarily rare and difficult skill," adding that "it's a bit arrogant to declare that you're definitely doing it. Even when you think you are, you usually aren't." That is a fair warning against this page as much as against anyone. The sign-off test is a standard, not something you can confirm by reading your own summary and approving it. The honest way to meet it is the one Brennan offers as an alternative to steelmanning: "seek out intelligent and well-informed advocates of viewpoints you disagree with," and quote them. That is how step two is meant to be run.

Some signs help tell the legitimate form from the weaponized one:

• The legitimate form passes the sign-off test: the view's holders would recognize it. The decorative form describes the view in terms its holders would not use.
• The legitimate form names the strongest point and answers it directly. The decorative form moves on to the weaker points.
• In the legitimate form, the rebuttal is shaped by the steelman; something in the argument changed because of it. In the decorative form, delete the summary and the rest of the piece reads exactly the same.
• Both forms may concede something. In the legitimate form, the concession changes the conclusion or how far it reaches. In the decorative form, it changes nothing.

The deletion test works as well on your own writing as on anyone else's.

Run it on your own side first

Every tool on this site carries this instruction. For this tool it means something specific: before you steelman the other side, check whether you can even state it. Try these:

• Pick a position you oppose on an issue you care about. Could you write a paragraph defending it that someone who holds it would sign? If you tried and could not, what does that tell you about how well you know the debate?
• Krugman was sure his side understood the other better; Caplan doubted it and wanted a test. Which side do you think understands the other? What test would you accept, and would the other side accept the same one?
• Think of the last time you summarized an opposing view. Would its holders have recognized it, or was the summary there to set up your next line? Did any of them actually tell you, or did you decide for them?
• When did you last change any part of your own view because of the other side's best argument? If you cannot remember, is that because your view has been right every time, or because you have not been meeting their best argument?

If that paragraph is hard to write, that is the most useful thing the tool can show you.

Grammars it works against

A grammar, on this site, is a recurring pattern of argument that turns up across many different subjects. The Steelman Protocol is one of four tools that work against every grammar (a different list from the five most important tools mentioned above), because every grammar has two poles, and each pole tends to argue against the weakest version of the other. It matters most in three places.

• G4: Loyalty, Heresy, and Betrayal. Groups punish the dissenter inside harder than the opponent outside. The Steelman Protocol is a first-line tool here. It asks the group to state the heretic's argument well enough that the heretic would sign it before deciding what the heretic deserves, and it asks the heretic to do the same for the group that expelled them, rather than treating expulsion as proof of being right.
• G10: Victimhood as Standing. Being harmed can grant real authority to speak, and can also be treated as settling factual questions the injury gives no special access to. A related idea is "steelmanning the injury while auditing the inference": take the strongest account of the harm seriously, then judge the factual claims built on it separately.
• G6: The Hidden Hand. Explanations built around a concealed coordinating actor are dangerous because they are sometimes right. Steelmanning such a claim means separating its best-documented core from what has been piled on top, which guards against both swallowing it whole and dismissing it whole.

Related tools

• F3: The Charity-with-Audit Protocol starts from the most reasonable reading of an ambiguous statement and revises only on a pattern of evidence; it is C2's partner against G4, and the site's version of the charity Dennett asks about.
• B5: The Fallacy-Fallacy Guard asks, after catching a bad argument, what the strongest non-fallacious version of the same conclusion would be. That question is a steelman in miniature.
• F1: The Crux Finder looks for the specific question a disagreement depends on. A good steelman often shows you where that question is.
• A4: The Mirror Check asks whether you held a claim to the same standard as its opposite, including whether your steelman of the other side is as strong as the version of your own side you carry around.