That justification is true and, I think, almost beside the point. The reason to steelman is not that you owe it to the other side. It is that you owe it to yourself, because the strawman version tells you nothing about whether you are right.

What refutation is for

Consider what you are actually doing when you argue against a position. You are running a test. The hypothesis is my view is correct and theirs is not, and the test consists of confronting your view with the best available reason to abandon it.

A test is only informative if it could have come out the other way. If you construct an opponent too weak to threaten your position, the outcome was fixed before you began, and passing tells you exactly nothing you didn't already believe. You have not learned that you are right. You have learned that you are capable of inventing someone who is wrong.

A refutation that could not have failed is not a refutation. It is a ceremony.

This is why the practice cannot be reduced to politeness. Politeness is about how the exchange feels to the other person. Steelmanning is about whether the exchange contains any information at all.

Three ways it goes wrong

The advice is easy to state and hard to follow, and it fails in recognisable ways.

Substituting the weakest actual holder for the strongest available argument. It is always possible to find someone defending your opponent's conclusion for terrible reasons. Quoting them is not engagement; the question is whether the position survives, not whether its worst advocate does.

Reconstructing the argument in your own vocabulary. Positions are often carried by their terminology. Translate a claim into the terms of the framework you are about to judge it by, and you have frequently assumed the conclusion before the argument starts. The test is whether a thoughtful adherent would recognise your restatement — not whether it sounds reasonable to you.

Steelmanning the conclusion instead of the reasoning. This is the subtlest failure. You find a good argument for their conclusion — one they never made — refute that, and declare the matter closed. But the reasoning people actually use is what determines what else they believe and what would change their minds. Defeating a substitute argument leaves the real one untouched.

The cost, which is real

Done properly, this is expensive. It requires reading people you find tiresome, and it requires holding a position in mind sympathetically for long enough to see what makes it attractive — which is uncomfortable, because sympathetic attention is how minds get changed.

That discomfort is not a side effect. It is the mechanism. A procedure that could never move you is not a procedure for finding out anything.

There is also a limit worth naming. Not every position has a strong version, and the demand to find one can become its own bad-faith move — an infinite obligation to keep improving your opponent's case until you have manufactured a view no one holds. The honest form of the practice includes a stopping point: here is the best version I can construct; here is why it fails; if there is a better one, I have not found it.

That last clause is the whole thing. It is what separates an argument from an announcement.

Why it matters more in politics

In most domains, believing false things has a cost that eventually finds you. Political beliefs are unusual in that almost none of them are ever tested against reality by the person holding them. Your vote does not decide the election. Your opinion on a policy will never be checked against what that policy does.

What that means is that the ordinary corrective — being wrong and finding out — is largely unavailable. The only correction left is argument. And argument only corrects anything if it is conducted against the real thing.

Which is to say: the reason to take the other side seriously is not that they deserve it. Sometimes they don't. It is that a habit of arguing against the weakest available opponent will, over years, leave you confidently holding a set of views you have never actually examined — and no way to tell the difference between that and knowledge.

That metaphor is doing work, and it should be made to earn it. What does a joint failure actually look like? If legitimacy can thin without any procedure changing, how would anyone know?

The measurement problem

Start with what makes this hard. The obvious indicators are the ones that move last.

Elections still happen. Courts still sit. Laws still pass and are mostly obeyed. If you were assessing the system on its formal properties, you would find nothing wrong right up until something is very wrong, because formal properties are exactly what a thinning legitimacy leaves intact.11 This is the standing embarrassment of institutional-quality indices, which reliably score states highly until shortly before they stop functioning. The measures aren't wrong; they're measuring the part that fails last.

The things that move first are attitudinal and therefore easy to dismiss as mood. People stop expecting institutions to work and start routing around them. Compliance continues but its character changes — from this is how things are done to this is what I have to do. That distinction is invisible in any statistic about compliance rates, and it is the whole difference between an institution that has authority and one that merely has power.

Three joints

If legitimacy rests on the three supports from the first part — consent, performance, procedure — then it has three characteristic ways of coming apart, and each looks different from inside.

Performance failure with procedure intact

The most legible case. Things stop working: the trains, the permits, the courts' backlogs, the time it takes to get a decision from anyone about anything. Nobody suspends the constitution. The constitution is fine. What erodes is the everyday experience of the state as competent.

What makes this dangerous is that it is nearly always locally explicable. Every individual failure has a proximate cause — an unlucky year, a bad appointment, an underfunded department — and so at no point does anyone have to conclude anything general. The reasoning that dissolves a railway in the previous essay applies here too: each explanation is sound, and the aggregate goes unexamined because nobody is charged with examining it.

Procedural failure with performance intact

The reverse case, and the one most likely to be defended. Decisions get made faster and better by narrowing who makes them. Emergency powers that work. Consultations skipped because the consultation was going nowhere.

Each step is defensible on its own terms and frequently produces better outcomes than the alternative. That is precisely what makes it hard to argue against without sounding like a person who prefers worse outcomes for procedural reasons — which, stated baldly, sounds absurd.

But the proceduralist has an answer, and it is worth stating plainly. Procedure isn't a tax paid for the sake of fairness. It's what generates a reason to comply that doesn't depend on agreeing with the decision. Strip it out and you haven't made governance more efficient; you've made every future decision rest on whether people happen to like it. That's a much thinner foundation, and its thinness shows only under load.

An institution that works well and answers to no one is borrowing against a reserve it is not replenishing.

Consent failure with both intact

The subtlest. Procedures work, outcomes are acceptable, and a growing number of people simply do not regard the arrangement as theirs.

This is where the vocabulary of nationhood does most of its work, in both directions. A shared national story is one of the few things that can sustain the sense of ours through a period when neither performance nor procedure is doing it. That is a real function and it is why appeals to it intensify exactly when the other supports are under strain — the appeals are a symptom before they are a cause.

The same mechanism runs the other way. A story that defines the nation narrowly enough to exclude people who are, in fact, here does not repair consent; it converts a diffuse thinning into a specific one, and concentrates it on the people the story excludes. Both moves feel like restoration to the people making them. Only one of them adds anything.

Why it gets misread

The characteristic mistake in all three cases is to read a structural problem as a moral one about individuals.

Performance failure gets read as laziness or incompetence in particular officials. Procedural erosion gets read as the ambition of particular leaders. Consent failure gets read as ingratitude, or as extremism, or as the influence of whoever is currently blamed for the influence.

Sometimes those readings are correct. Individuals do behave badly, and pretending otherwise is its own evasion. But the individual reading is always available — there is always someone to blame — and its availability is the tell. An explanation that would have been produced regardless of the facts is not evidence about the facts.

The structural reading is harder because it doesn't terminate in anyone's guilt, and so doesn't discharge the feeling that something ought to be done about someone. It terminates instead in a question about which support is under strain and what would relieve it, which is duller and considerably more useful.

What follows

Not much that is comfortable. Three things, though, seem to me to survive the argument.

The first is that repair is specific. "Restoring trust" is not an action; strengthening a particular joint is. Which one is failing is an empirical question with a discoverable answer, and the answer is often not the one being loudly asserted.

The second is that the supports are not interchangeable. A state that has lost procedural legitimacy cannot compensate with performance indefinitely, and one that has lost performance cannot compensate with procedure. Each buys time. None substitutes.

The third is the one I find hardest to hold on to. Because the failure is gradual and every stage is survivable, there is no moment at which the situation announces itself — which means the judgement about whether things are getting worse has to be made without the evidence that would settle it. That is an uncomfortable epistemic position, and I don't think there is a way out of it. There is only the discipline of asking the question on a schedule rather than when it feels urgent, since by the time it feels urgent the answer has usually stopped being useful.

Three answers dominate. Each is compelling in isolation. Each fails in a way that points toward the others.

The answer from consent

The oldest and most intuitive: a government is legitimate because the governed agreed to it.

The appeal is obvious. It grounds authority in something we already think generates obligations — a promise. And it explains the intuition that conquest, however stable, doesn't produce a right to rule.

The problem is equally obvious, and Hume made it in 1748: nobody agreed to anything. You were born here. You never signed. The standard repair is tacit consent — by remaining, by voting, by using the roads, you have consented in effect.

Hume's reply has never really been answered.11 Locke saw the difficulty and tried to meet it with the doctrine of tacit consent; Hume's essay is a direct response. The modern repair is usually to abandon consent as a historical fact and reconstruct it as hypothetical — what a reasonable person would agree to — which is a different argument wearing the same word. He asks us to consider a man carried aboard a ship while asleep, who is told he consents to the captain's authority by remaining on board, when leaping into the ocean is his only alternative. Consent that cannot be withheld at tolerable cost is not consent. It is a description of the situation dressed up as a justification of it.

And notice the trouble runs deeper than emigration costs. Even actual, explicit consent would only bind the people who gave it — not their children, not the people who voted against, not anyone who arrives later.

The answer from performance

So try a different tack. A government is legitimate because of what it does: it keeps order, secures rights, provides what individuals cannot provide alone. On this view, legitimacy is earned by results, and a state that systematically fails to deliver them forfeits its claim.

This handles the cases consent theory chokes on. It explains why a functioning state you never agreed to still has a claim on you, and why a state that murders its citizens has none.

But it proves too much.22 This is sometimes called the "performance legitimacy" trap, and it is not merely academic — it is the explicit self-justification of several contemporary states. Most people who find performance theory attractive do not actually believe this. They believe there is something about how decisions get made that matters independently of how they turn out — which is to say they believe something performance theory cannot express. If legitimacy is a matter of results, then a sufficiently competent authoritarian is legitimate, and a democracy going through a bad decade is not.

There is a second problem. Results have to be evaluated by some standard, and which standard you pick is itself the contested political question. "Delivers good outcomes" is not neutral ground; it is the disagreement, relocated.

The answer from procedure

Hence the third answer, which is the modern one: what legitimates a government is not what people agreed to or what it delivers, but how it decides — through public rules, applied equally, revisable by those subject to them, with genuine opportunity for dissent and reversal.

The move here is subtle and worth naming. Procedure doesn't try to justify the content of political decisions. It justifies the authority of the process that produces them, so that losing a vote leaves you with a reason to comply that doesn't depend on your thinking the decision was correct.

That is a real achievement. It is the only one of the three that explains how a society containing genuine, permanent moral disagreement can be governed at all without one faction's answer simply being imposed.

But procedure alone is hollow, and the hollowness shows up in two places. First, a scrupulously fair process can produce monstrous results; proceduralism has to smuggle in some substantive limit — rights, usually — to avoid endorsing them, and where that limit comes from is exactly what procedure was supposed to avoid answering. Second, a process only confers authority on people who accept that process, which relocates the consent problem rather than solving it.

What survives

Notice the pattern. Consent explains why authority must be answerable to the governed but cannot show that it ever was. Performance explains why a state must actually do something but licenses too much. Procedure explains how disagreement can be managed but needs limits it can't itself supply.

Each theory fails where the others are strong. That is not a coincidence, and I don't think it means one of them is secretly correct.

Legitimacy is not a property a state has. It is a claim a state makes, which is more or less well supported, and never fully.

What follows from taking that seriously is a change in the question. The useful question is almost never is this government legitimate? — as though it were a threshold to clear. It is which of legitimacy's supports is under strain here, and what would it take to repair it?

A state that no longer delivers is in trouble even if its procedures are impeccable. A state with excellent outcomes and no mechanism for dissent is in trouble too, and its trouble is invisible right up until it isn't. Legitimacy is a standing achievement, maintained the way a building is maintained, and it fails the way buildings fail: slowly, at the joints, and then all at once.

That is a less satisfying answer than any of the three theories. I think it is the honest one, and I think it is more useful — because it tells you where to look.

Kahn's point was that the market had answered a question it was never asked. Each traveller chose between driving today and taking the train today. Nobody was offered the choice between a world with rail service and a world without it. Had that choice been on the table, many of the same people would have chosen differently — including people who had been driving all along.

He called it the tyranny of small decisions, and once you have the concept you find it everywhere.

The structure

The pattern requires three things:

1. The aggregate outcome is not what anyone wanted. This isn't a case of some people winning and others losing. It's a case where nearly everyone, shown the endpoint in advance, would have preferred a different one.
2. Every individual decision was reasonable given the others. No mistakes, no irrationality. Each person optimised correctly against the world as they found it.
3. The aggregate choice was never actually put to anyone. There was no moment at which the real question appeared on a ballot, a menu, or a contract.

That third condition is what separates this from the more familiar collective action problems. In a prisoner's dilemma or a tragedy of the commons, the participants can see the trap; the difficulty is that defecting is individually better even when everyone understands where it leads. Small decisions are worse in a specific way: the trap is invisible from the position of the chooser. There is nothing to see, because at the scale of a single decision nothing is happening.

Why the usual fixes underperform

Because the mechanism is different, the standard remedies land badly.

Better information doesn't help much. Telling the driver that rail service depends on ridership doesn't change the calculation for today's trip, because today's trip genuinely doesn't decide anything. The individual's reasoning was never mistaken.

Appeals to conscience don't scale. They ask a person to bear a real cost for an imperceptible benefit. Some will. Not enough, and the ones who do are penalised for it.

Pigouvian pricing works only when you can price the thing. Sometimes you can. But often what's being lost is the option itself — the continued existence of an alternative — and options are notoriously hard to price precisely because their value is contingent on futures that haven't happened.

What tends to work, when anything does, is changing the choice architecture so the aggregate question is actually asked: a franchise obligation, a common carrier rule, a subsidy that makes the alternative's survival independent of marginal demand. In each case the intervention isn't "make people choose better." It is "put the real question somewhere a decision can be made about it."

The political version

The reason this matters beyond transport economics is that political life is thick with the same structure, and our vocabulary for it is thin.

Consider the slow disappearance of an option. Cash payment. A non-algorithmic route to a human being. Local newspapers. Third places that cost nothing to sit in. In each case, no one decided. Each institution's decline was the sum of reasonable individual choices, and at no point was anyone asked whether they wanted the option to persist.

The political difficulty is that our normal grammar for these disputes is someone did this to us, and here nobody did. That leaves two unattractive responses: invent a villain, or conclude that because the outcome emerged from free choices it must be what people wanted.

The second conclusion is the one I want to reject, because it doesn't follow. Revealed preference tells you what someone chose from the options in front of them at the moment of choosing. It does not tell you what they would choose about the structure of the options, because that question was never put. Reading the first as though it answered the second isn't a defence of markets; it's a category error, and one that happens to have political consequences.

"Nobody decided this" is a description of a mechanism, not a defence of an outcome.

What it does not license

I want to be careful, because this argument is easy to over-extend, and over-extended it becomes a warrant for overriding any choice you dislike on the grounds that people would have chosen otherwise if only they'd been asked properly.

Three constraints keep it honest.

The first is that the counterfactual has to be genuine. It is not enough that you would prefer the alternative world. The claim is that the people who made the small decisions would prefer it — and that is an empirical claim, which can be false, and which you should be prepared to be wrong about.

The second is that plenty of things disappear because people stopped wanting them. Not every loss is a coordination failure; sometimes the aggregate outcome is exactly what the aggregate wanted, and calling it tyranny is nostalgia with a footnote.

The third is that the appropriate response is almost always to restore the question, not to impose the answer. Requiring that an option remain available is a different act from requiring that people take it. The first respects the choosers; the second assumes they were wrong.

Kahn's railway is a useful case precisely because it's small. Nobody's liberty was at stake. But the shape of the problem is the same at every scale, and the shape is what's worth carrying: when an outcome nobody wanted arrives with nobody's fingerprints on it, the failure is usually not in the choosers. It is in what they were given to choose between.

I think the metaphor is wrong, and I think defenders of free expression should be glad to be rid of it. The principle it's used to support is sound. The support is not.

What a market actually needs

Markets aggregate information well under conditions. The conditions matter.

A market needs participants who bear the costs of being wrong. It needs a feedback mechanism that transmits those costs back to the decision-maker. It needs the thing being traded to have a quality that reveals itself through use. Buy a bad car and you find out. That discovery is what makes the price system informative.

Now apply this to beliefs about, say, trade policy. What is the cost to me of holding a false one? For nearly everyone, essentially zero. My belief doesn't determine policy. It won't be tested against outcomes. What it does do is signal my affiliations and let me participate in conversations I want to be part of — and those returns are real, immediate, and entirely unrelated to whether the belief is true.

This isn't a market failure at the edges. The central mechanism is absent. There is no feedback loop connecting the truth of most publicly contested beliefs to any consequence borne by the people holding them.

Where being wrong is free, competition selects for what is appealing, not for what is true.

What the competition actually rewards

If the selection pressure isn't truth, what is it? Roughly: whatever spreads. Ideas that are easy to hold, that flatter the holder, that arrive with a ready-made community, that convert into identity, that resist disconfirmation by construction.

Some true ideas have these properties. Many don't — true ideas are frequently complicated, unflattering, provisional, and boring. Meanwhile falsehoods engineered for transmission have every advantage.

The marketplace metaphor predicts that error is self-correcting. What we observe is that some errors are self-correcting and others are self-reinforcing, and which is which depends on features of the belief that have nothing to do with its truth. That's not a market working slowly. That's a different mechanism than the one the metaphor describes.

Why this matters practically

Metaphors have consequences. This one has at least two bad ones.

It licenses complacency. If open competition tends toward truth, then any particular falsehood is a temporary state that the process will handle. So the argument doesn't need to be made, the correction doesn't need to be issued, the institution doesn't need defending. The market will sort it out. But the market, in this sense, isn't sorting anything — and the belief that it is removes exactly the individual effort that does the actual work.

It concedes the wrong ground. If the justification for free expression is that it produces true beliefs, then someone who shows it isn't producing true beliefs has produced an argument for restricting it. Defenders of open discourse have accepted a premise that makes their position hostage to an empirical claim that is, at best, shaky.

The better argument, which was there all along

Mill's actual case in On Liberty isn't the marketplace. Read carefully, it has three parts, and none requires believing that the best argument wins.

The first is fallibilism. Any belief we suppress might be true, and we have no reliable way of identifying in advance which ones are — because the mechanism we'd use to identify them is the same mechanism that produced our current confidence. This isn't a claim about outcomes; it's a claim about our epistemic position, and it holds regardless of how the competition goes.

The second is that even true beliefs decay without challenge. A belief held without ever encountering a serious objection is held as prejudice — the holder can't say why it's true, can't recognise its limits, can't apply it to a case it wasn't handed to them with. Mill's point is that the contest maintains understanding, whether or not it settles the question.

The third is about who does the suppressing. Any power to restrict expression is exercised by particular people with particular interests, and it will be used — over time, and predictably — to protect the powerful from criticism. This is an argument from institutional design, not from epistemology, and it is untouched by any finding about how discourse actually performs.

Notice what these have in common: not one of them claims that open discourse converges on truth. They claim that we cannot identify error in advance, that unchallenged belief degrades, and that the cure is worse than the disease. Those claims survive everything I've said above.

They also cost something the marketplace metaphor pretends is free. On Mill's account, an open society doesn't automatically produce good beliefs; it produces the conditions under which good beliefs remain possible, and the rest is work — argument that has to be made, error that has to be answered, institutions that have to be maintained by people who could do otherwise.

That's a harder case to make than the one about the market. It's also the one that's actually true, and it doesn't collapse the first time someone points at the internet.

What follows is a short list of the ones I find most load-bearing. The exercise in each case is the same: replace the term with what it's actually asserting, and see whether the argument still works.

"Natural"

Almost always doing evaluative work while claiming to be descriptive. A natural arrangement is one that arose without deliberate design — which tells you about its history and nothing whatever about whether it's good. Feudalism was natural in this sense. So was cholera.

The word survives because it borrows the authority of inevitable and the warmth of healthy while committing to neither. When someone says an arrangement is natural, the useful question is: natural as opposed to what, and why does that difference matter? Usually the answer is that the alternative would require someone to decide something, and the speaker prefers that no one does.

"Common sense"

A claim about the speaker's confidence, presented as a claim about the proposition. It asserts that a position needs no defence, which is convenient, since offering none is exactly what the phrase permits.

What makes it slippery is that sometimes it's right — some things really don't need arguing. But the phrase is used most heavily precisely where a position is contested, which is where it's least warranted. The tell is that "common sense" almost never appears in front of something everyone agrees with. Nobody says it's common sense that fraud should be illegal.

"Freedom"

The classic case, and Isaiah Berlin's distinction remains the most useful tool for it: freedom from interference versus freedom to actually do the thing. These frequently conflict, and almost every substantive dispute about liberty is a dispute about which sense is at stake.

What the single word conceals is that both readings are always available, so both sides can claim the term without either noticing they've stopped talking about the same thing. The repair is mechanical and it works: free from what, and free to do what? Every argument I've seen improve did so within a sentence of someone asking.

"The people"

Does an enormous amount of work with almost no content. It can mean the electoral majority, the total population, the subset the speaker identifies with, or a normative construct — what the population would want if properly informed — which is unfalsifiable by design.

The rhetorical value comes from sliding between these. A policy supported by 51% is "what the people want." Opposition to it is not the people disagreeing; it's a faction, or an elite, or a special interest. The category is defined so that it cannot contain dissent, which means it has stopped describing anyone.

"Extreme"

Purely relational, and pretends not to be. Extreme means far from the centre, and the centre is wherever the current distribution of opinion happens to sit — which is a fact about a population at a time, not about the merits of a position.

This matters because the word is used as though it were an argument. It isn't one. Abolition was extreme. So was universal suffrage. The distance of a view from the median tells you about its political prospects and precisely nothing about whether it's correct.

The honest version is usually available and usually stronger: this position is held by few people, or this position would require changes we're not prepared for, or this position is wrong, and here is why. Each of those can be argued with. "Extreme" cannot, which is the point of using it.

The common thread is that all five are conclusion-bearing. They don't describe a position; they deliver a verdict on it, in a form that looks like description and so escapes having to be defended.

I want to be careful about what this does and doesn't establish. It's not a claim that people deploy these terms in bad faith — mostly they don't, and the freight comes with the word rather than being loaded by the speaker. Nor is it a claim that a suitably neutral vocabulary is available. It isn't. Every political term worth using is contested, and demanding definitions before every conversation is its own kind of obstruction.

The claim is narrower: when an argument turns on one of these words, that's the place to look. Substitute the literal content and one of two things happens. Either the argument survives — in which case you've clarified it, and the word was harmless shorthand — or it doesn't, in which case the word was carrying the whole thing.

That test is cheap, and it works about as well on your own writing as on anyone else's. Which is the uncomfortable part, and also the reason to run it.