Steelmanning and The Problem with Making Everyone Sound Rational

Respond to the arguments people actually make, not hypothetically better arguments

Photo by Jerry Kavan on Unsplash

Picture yourself stepping out of a time machine into an elegant political salon in 1845 America. Crystal chandeliers cast warm light over heated debates about the burning questions of the day: Should women be granted the right to vote? Is the enslavement of Africans morally justified? Should children continue working with dangerous machinery in factories for twelve hours a day?

You encounter a man arguing that women should not be allowed to vote¹ and you begin responding that this is just rank bigotry against women. Just then, a proto-Rationalist steps in and argues you should be more generous to your interlocutor. After all, you aren’t responding to the best version of their argument.

In a scolding tone, you are told that the best version of their argument is an elaborate bank shot about economic growth, political stability, and the notion that slow social change in society is beneficial for stability.

Something feels deeply wrong about this intervention, but what exactly?

When the steelman does not compute

I really can’t stress enough how often political discussions take the form of misrepresenting someone’s arguments to make them weaker. It’s a scourge when trying to understand the world, and it frequently occurs when the stakes are high and involve important moral questions. But steelmanning–responding to the “best” possible version of someone’s argument rather than their stated argument–can lead to engaging with phantom advocates that don’t behave like your “improved” versions do. This is especially plain when there are big disagreements between people on norms or what a good and just society would look like.

Sometimes people have repulsive views. They may or may not be willing to state those views, but steelmanning as a habit will tend to guide one away from understanding their likely actions in such cases. The best version of an argument rarely contains malicious intent or poorly informed bigotry, but many people do have malicious intentions (particularly towards outgroups) and are poorly informed bigots. Steelmanning the arguments of these people would lead you just as astray as someone steelmanning the case for African slavery would have in the 1800s. It wouldn’t help you better understand either what the person actually believes, their behavior, or the policies they concretely support.

Nor would it necessarily even be convincing to the person you are steelmanning. If you told an advocate of African slavery that the best reason to continue African slavery was some story about economic growth when what they really thought was “I deserve to rule based on [archaic racial takes]” they may agree with the outcome, but persuading them the economic benefits were overstated wouldn’t actually change their mind about whether African slavery was worth doing.

An academic exception

There are still contexts where steelmanning makes sense. In certain academic settings, a habit of steelmanning can be useful, particularly as a pedagogical exercise. When everyone agrees on basic moral premises and is genuinely committed to evidence-based reasoning, steelmanning can help illuminate genuine areas of uncertainty and help open people to changing their minds. But even in academic contexts, this should be labeled clearly as an intellectual exercise rather than an attempt to understand actual actors. There's nothing wrong with exploring hypothetical positions held by imaginary people—just don't confuse those imaginary people with real ones making real decisions that affect real lives.

Out in the real world, I take the bold stance that the best way to understand a political opponent is to model them accurately. This requires being generous and presenting a genuine version of their arguments, rather than a strawman. But in these cases where you’re engaged in discussion because someone has an important politically influential belief you think is false, you should not use steelman, you should address their actual view of either the person and/or the audience you’re engaging. Being convincing in these circumstances likely means not making up an even “better” version of their position and responding to that view.²

You might counter that I’m intentionally selecting an unrepresentative sample of issues, picking things that are largely settled moral topics. But that’s the point. Some currently active arguments are driven largely, if not entirely, by prejudice. We obviously don’t know perfectly in real time which arguments those are (and sometimes not even in retrospect), but sometimes that’s the real reason for behavior.

Powering down the steelman

My position here might be best articulated by this Sep 2024 Bluesky thread from the philosopher John Holbo:

“If you are never arguing civilly with people you regard as so wrong that it is getting annoying, you are not swallowing annoyance enough, for the sake of maybe learning something. But: tolerance never requires interpretive fiction - pretending your bad interlocutor is in good faith, steelmanning.”

“You never need to invite actual Smith to be on your panel, even though he would send you to the camps and is a menace to honest debate, just because you can conceive an alt-version of Smith, more intriguingly debateable. Actual invites are for actual people not possible people.”

“If an actual person is horrible you could possibly send him a possible invite to a possible panel at a possible conference on a possible world possibly closely counterpart to ours, if he has a counterpart in the world who deserves to be on that panel. Modal epistemic justice!”

You should make every effort to ensure you have an accurate model of the arguments against any (important) position you hold. But this never requires lying to yourself about what someone believes or why they believe it. There’s no obvious trick here about when to read good faith into others’ arguments, and when not to. There are often many smart people who disagree with you,³ but those people being smart isn’t a reason to think the reasons they hold a position are so free from prejudice or ignorance to preemptively rule out those potential avenues when considering their actual arguments.

If you consistently steelman, you will likely encounter in your mind many more people than exist who are hyperrational and happen to have a peculiar view of the economy or how society responds to change, who also think those people deserve less legal protections than the in-group.

Yes, in some settings, where you are quite certain of everyone’s motives and takes, steelmanning might help you see a different perspective. But political discussions about rights and legal protections are often not one of those settings, even if the person making the argument is otherwise a reasonable person.

By all means, when in conversation–particularly when disagreeing with someone–be kind to others and ask for clarification when you are uncertain about what someone meant. Grant premises if they aren’t central to the concern you are raising. But fundamentally, address the arguments that are actually made, not hypothetical better or worse arguments.

Accurate modeling over charitable fiction

What I am proposing doesn't mean abandoning all charity in interpreting others' arguments. Needlessly assuming the worst possible motivations without evidence leads to its own problems. The goal should be accurate modeling rather than either extreme charity or hostility.

Here's what this looks like in practice:

Listen to what people actually say, not just the most sophisticated version of what they might mean. If someone consistently uses certain coded language or focuses on particular concerns, take that seriously rather than translating it into more palatable academic concepts and terminologies.

Pay attention to revealed preferences, not just stated arguments. What policies do they actually support? How do they behave when they think no one is watching? What happens when their stated principles conflict with their group loyalties?

Consider the full range of human motivations, including unflattering ones. People are sometimes driven by selfishness, prejudice, tribal loyalty, and simple ignorance. Sophisticated reasoning exists, but it's not the only or even primary driver of beliefs.

Distinguish between different types of disagreement. Good faith disagreements about complex empirical questions deserve different treatment than conflicts rooted in fundamental value differences or prejudice.

1. No, Peter Thiel has not gone back in time with you. Someone steelmanning Thiel might note when he infamously argued since 1920 the increase in “welfare beneficiaries and the extension of the franchise to women” led him to be more pessimistic about the compatibility of democracy and freedom, he wasn’t necessarily arguing against women being allowed to vote. After all, the more charitable interpretation is merely women being allowed to vote have made the US much worse off but that they should continue to be allowed to vote even though this has severely limited “freedom” in his view. This is like how when someone in the US argues that the country went downhill since the Civil Rights Act of 1964, you can not be literally certain they support racial segregation and discrimination. After all, it could be a big coincidence of timing.

2. A really common circumstance where people needlessly steelman is when they are trying to stop someone from facing social repercussions of being “canceled.” The typical setup is someone says something inflammatory on Twitter, people become outraged, and then others try to erect a steelman because they don't think that the original poster should be labelled as a "Bad Person" (not because they think the poster’s views match the steelman). In this case, the whole debate is fake, because we're really just arguing about who should be accepted in polite society.

3. A 2016 Ozy Brennan post “Against Steelmanning” made this point about not having to invent arguments because “you have many smart opponents!” But, as if to make the point about how I’m not strawmanning steelmanning as a concept–and that steelmanning really causes big problems when there are value differences–someone in the comments apparently made a genetics/race and IQ comment and some people took this as evidence they might support segregation. They found this ridiculous and insulting but in a huge amount of cases support of such policies is exactly why these types of comments are made. The demand to steelman in this manner that preemptively rules out background knowledge of favoring bigotry is basically a demand to treat each claim someone makes in isolation of all other knowledge you have about people’s how people’s motivations, beliefs, and statements interact.