Steelmanned honesty

The strawman is a confession of weakness

The easiest way to win an argument is to fight a weak version of the other side. You restate your opponent's position in its clumsiest form, defeat the clumsy form, and declare victory — and the reader, if they share your priors, applauds. This is the strawman, and it is everywhere, because it works on anyone already inclined to agree. It is also a confession: a writer who attacks the weak version of an objection is telling you, without meaning to, that they are not confident their thesis could beat the strong one. Steelmanned honesty is the inversion of this. It is the discipline of constructing the strongest version of every objection — stronger, often, than any actual critic has bothered to make — and meeting that head-on. The rule is severe: a thesis is only worth holding if it survives its own steelman. If it can only beat the strawman, it does not get hedged into safety; it gets abandoned.

This is why ENSI can take adversarial positions on dangerous terrain without sliding into advocacy. Human-AI power dynamics: the gradual disempowerment problem is, in form, an extended steelman — it takes the strongest version of the pessimistic case about human agency under advanced AI and maps it in full, refusing the comfort of dismissing it as alarmism. The alignment problem: aspects to align decomposes a contested problem without pretending the hard parts are solved, giving each difficulty its strongest statement before responding. The trustworthiness these pieces earn is a direct function of how seriously they take the opposing case. A reader can feel the difference between an argument that has beaten a real adversary and one that has only beaten a puppet, and the method is built to produce the former.

Honesty about what you do not know

Steelmanning is one half of intellectual honesty; the other half is candor about the limits of your own claims. The method sorts every assertion by how much it can bear — what was measured, what was inferred, what is predicted — and writes each at the altitude its support allows, rather than inflating all of them to the same confident register. Superalignment: 40 techniques for aligning superintelligent AI is honest in exactly this way: it presents many techniques as approaches rather than guarantees, because the evidence supports cataloguing and assessing them, not declaring them solved. Logic: the limits of scientific inference is an entire essay about this discipline — the precise boundary between what an inference can and cannot license — and epistemology: from discovery to justification draws the line between having an idea and being entitled to assert it. The method lives on the justification side of that line. Predictions are written as conditional cascades — if the cost curve holds, then — rather than as prophecy, because that is both more honest and, not coincidentally, more defensible.

The hardest form of honesty is letting the evidence overrule the thesis. A writer who has built a beautiful argument is powerfully motivated to defend it against the inconvenient fact, and most of the machinery of bad reasoning exists to serve that motive — the selective citation, the buried caveat, the steelman quietly downgraded to a strawman when no one is looking. The method's defense is to construct the steelman before drafting, in good faith, and to let it have the authority to send the thesis back. If the strongest objection turns out to be stronger than the thesis can answer, the thesis is demoted or cut, not papered over. Regulation of externalities caused by AGI reads as credible precisely because it does not pretend the regulatory problem is easily solved; it concedes what must be conceded and argues only as far as the argument can actually reach. Honesty here is not a tone; it is a willingness to lose the argument to the evidence.

Conviction earned, not assumed

There is an apparent paradox at the center of the method: ENSI writes with great conviction — declarative, unhedged, axiomatic — while also insisting on meeting every objection and conceding every genuine limit. The resolution is sequence. The conviction comes after the steelman, not instead of it. You construct the strongest objection, you meet it head-on, you concede what it forces you to concede and rebut what you can rebut — and only the claims that survive that gauntlet earn the right to be stated flatly. The unhedged sentence is therefore not the opposite of intellectual honesty; it is its reward. Hedging everything is the failure of nerve that hides behind rigor's costume — a writer who claims nothing strongly can never be wrong and never be useful. The method rejects that as cowardice and demands the harder thing: verify adversarially in private, then assert confidently in public, but only the survivors.

This is why steelmanned honesty is the principle that makes the other five safe to wield. First-principles thinking, contrarian reframing, and systems synthesis are all powerful enough to be dangerous — they generate bold, consensus-cracking claims, and bold claims are exactly the ones most likely to be confidently wrong. The steelman is the governor on that engine. A reframe like bandwidth, not character is only allowed to stand because it has been made to beat the strongest defense of the character view; a systems claim is only allowed its sweep because it has answered the strongest charge of overreach. Without the steelman, the method would be a machine for producing brilliant errors at speed. With it, the same boldness becomes trustworthy, because every bold claim carries the invisible history of the objection it defeated.

The reader can feel the difference

The ultimate reason to steelman is that readers are not fooled for long. A reader who knows the field can feel when an objection has been handled honestly and when it has been dodged, and the dodge — once noticed — poisons trust in everything else the writer says. Conversely, when a piece meets the reader's own strongest objection before they can raise it, something rare happens: the reader stops defending and starts thinking alongside the author, because the author has demonstrated that they are not selling but reasoning. This is the texture of ENSI's pieces on contested questions — the disempowerment problem, the alignment aspects — where the reader's resistance is disarmed not by persuasion but by the visible fact that the hardest version of their objection has already been taken seriously.

The wager of this principle is that the strength of a thesis is measured by the strength of what it can defeat, not by the weakness of what it chooses to fight. A thesis that has only beaten strawmen is worthless however confidently it is stated; a thesis that has beaten its own steelman is worth holding even when it is uncomfortable. So the method builds the strongest objection on purpose, meets it head-on, concedes what is true and rebuts what is not, and keeps only what survives. Meet the strongest version, every time. If the thesis lives through that, it has earned its conviction — and if it does not, the honest thing, the only thing, is to let it go.

Further reading

• Human-AI Power Dynamics: The Gradual Disempowerment Problem
• The Alignment Problem: Aspects to Align
• Superalignment: 40 Techniques for Aligning Superintelligent AI
• Logic: The Limits of Scientific Inference
• Epistemology: From Discovery to Justification
• Regulation of Externalities Caused by AGI