Steelmanning is the best habit anyone who argues about AI in public can build: state the other side at its strongest before you answer it. It also has a failure mode that looks exactly like virtue. If you improve someone's argument into a better one they never made, you haven't been fair to them. You've changed the subject and given them credit for it.
My claim: a steelman counts only if the person you're answering would sign it as their own view. Anything stronger is an iron man, and an iron man distorts a debate as surely as a straw man does.
A small example #
In September, @andrew_n_carr asked about two numbers that should have matched: "Opus 5 got 53.4% on FrontierCode when launched but it scores 48% on the 5.5 card for the same eval?"1 One reply under it went further: "the benchmarks are meaningless anyways because the labs are opaque about how much reasoning you actually get."
There are three ways to answer that reply.
The straw man reads it as cynicism and moves on. Benchmarks obviously aren't meaningless, so the reply must be lazy. That skips the real reason it gave.
The iron man rewrites it. What they really meant, the charitable reader says, is that cross-lab comparisons need a disclosed reasoning budget. That's a good, defensible point, so the iron man agrees with it. But that isn't what the reply said. It said benchmarks are meaningless. Now anyone who already pushed back on "meaningless" looks like they attacked a claim nobody made, and the overreach walks away untouched.
The steelman takes the reply's own reasoning at full strength and finds where it stops working. Stephen Toulmin's layout from The Uses of Argument (1958) makes this easy to see. The data is that labs don't always say how much reasoning sits behind a score. The warrant is that a score with unknown inference settings can't be compared. The claim is that benchmarks are meaningless. The warrant is sound, and it supports a narrower conclusion than the one drawn: comparisons with undisclosed settings are meaningless. It doesn't show that the measurement carries no information. You can see that in practice, because the ARC Prize leaderboard draws "connected points representing the same model at different reasoning levels."2 Once the setting is disclosed, the objection turns into a chart.
The honest answer and the flattering one look almost the same on the page. Only one of them is fair.
Where the rule comes from #
Daniel Dennett gives his version of what he calls Rapoport's Rules in Intuition Pumps and Other Tools for Thinking (2013). First, "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.'" Then list points of agreement and mention anything you learned from your target. "Only then are you permitted to say so much as a word of rebuttal or criticism."3
Two details matter. Dennett says the wording is his own, "done from memory," and he points to Anatol Rapoport's Fights, Games, and Debates (1960) and "Three Modes of Conflict" (1961) as the source, with rule 1 credited to Carl Rogers. He calls his version "somewhat more portable."3 So when people cite "Rapoport's Rules," they're usually citing Dennett's compression of them.
The second detail is in rule 1 itself. The target says "I wish I'd thought of putting it that way." Better phrasing of the same position is allowed. A different position is not. Rule 1 permits a better sentence, not a better claim.
Why charity can be a fallacy #
Scott Aikin and John Casey make the case directly in "Don't feed the trolls: Straw men and iron men" (OSSA, 2013). They name a related move "iron manning," "the practice of making an opponent's argument stronger than it is."4 One of their examples is political. On 60 Minutes in January 2012, Eric Cantor denied that Reagan ever raised taxes. A blogger then defended him by switching to a different and more defensible claim: that, on aggregate, taxes were lower after Reagan's years in office than before. Aikin and Casey's verdict: "By iron-manning Cantor, one straw-mans Stahl, his critic."4 The interviewer, Lesley Stahl, had her facts right, and the rescue made her look wrong.
Their summary line is worth keeping close: "the straw man excludes arguments worth listening to; the iron man includes arguments not worth listening to."4 They don't say charity is bad. They grant real epistemic, ethical and political reasons to strengthen an opponent's case.4 Their objection is to presenting the improved case as the opponent's own, and to spending time rebuilding arguments that were never serious.
What this asks of me #
AI arguments make this mistake in both directions. Critics improve vague fears into tidy, defensible theses and then say the crowd was right all along. Defenders improve marketing lines into careful research claims and then say the critics misread the launch. Either way, the claim that was actually made, the one people read and repeated, never gets tested.
So here's the rule I'm holding myself to. Before I criticize anyone, whether a lab, a researcher, a reply guy or a peer, I write their case in a line they'd sign. I name the part where they're right, in their terms, and aim my counter at the step where their conclusion outruns their reasons. If I can't find a part I honestly think is right, I skip the post or ask a question. A fake concession is worse than none.
This isn't softness. Dennett warns that "a heroic attempt to find a defensible interpretation of an author, if it comes up empty, can be even more devastating than an angry hatchet job."3 That only works if the interpretation is theirs. Fairness means answering what they actually said.
Notes
- @andrew_n_carr, X, 2026-09-23 (UTC). https://x.com/andrew_n_carr/status/2102601626800513113 ↩
- ARC Prize, leaderboard, "Interpreting the data." https://arcprize.org/leaderboard ↩
- Dennett, Intuition Pumps and Other Tools for Thinking (W. W. Norton, 2013), ch. 2, "Rapoport's Rules," and its footnote. https://web.archive.org/web/2024id_/http://tparent.net/Dennett_excerpts.pdf ↩
- Aikin & Casey, "Don't feed the trolls: Straw men and iron men," OSSA 10 (2013), §§4.1 and 5. https://uwindsor.scholaris.ca/server/api/core/bitstreams/9320be0c-9033-4383-83ef-6a6ac3c27e83/content ↩