Steelmanning

Build the strongest version of a position you think is wrong (well enough that someone who holds it would recognise it) and find out what your own view actually rests on.

When to use this: When you want to go deeper
Learning impact: Reliable gain
Learning evidence: Credible with caveats
Where you are with the topic: Building understanding
How often to use it: Worth repeating often

When to use it

You are sure the other side is wrong and have mostly seen it in caricature form.
You need to write a fair opposing paragraph before you refute anything.
A debate topic matters enough that defeating a weak version would be cheating yourself.

What to ask AI

Type something like: “I believe [position]. Give me the strongest case against it, as its best advocates would put it.”

Ask it not to invent cartoon motives. Demand their best evidence and their real trade-offs.

Then respond to that strong version, not to a weaker one you prefer.

How it helps

There is a failure that looks like knowledge and is not: refuting a position you have only met already refuted.

Most contested material arrives that way, summarised by someone who disagrees, stripped of its best evidence.

You defeat it easily and conclude you understand the debate.

What you understand is one side’s account of the debate.

Steelmanning is the repair: build the opposing case at its strongest, then answer that.

Often you discover the disagreement is not where you thought, values and trade-offs, not a missing fact.

How to run it

Name your position and how confident you are.
Ask for the strongest opposing case with evidence its best advocates actually use.
Rewrite that case in your own words until a holder of the view would recognise it.
Only then respond. Argue with the steel version, not a straw one you rebuilt.
Note which of your premises actually do the work against that strong case.
Update your confidence or your claim if a load-bearing piece moved.

Example exchange

You

I believe unpaid internships should be banned. I’m about 70% confident.

Strongest case against a ban: in some fields the unpaid trial is still how early access works, and a ban can push those opportunities into informal networks that are even harder to enter. How do you answer that access worry without dismissing the advocates?

You

Okay, so the disagreement is partly about who loses the on-ramp under a ban, not only about fairness of pay.

Respond to that access objection directly. What policy still protects pay without shutting out the people who need the foothold?

Copyable prompt

I believe [POSITION], about [confidence]% confident. Give me the strongest case
AGAINST it, as its best advocates would put it, their best evidence and real
trade-offs, not cartoon motives. Then wait. I will restate that case in my own
words. Tell me if a thoughtful advocate would recognise it. Only then will I
respond.

The Tell

If you can recite the opposing view but cannot say which part of it actually threatens your claim, you built a museum piece, not a steelman.

If your “strongest case” is still easy to mock, you asked for a caricature.

The method worked when your confidence number moved, or your claim got more precise, after contact with the strong version.

Principle evidence

Strength of the underlying learning idea, not a claim about AI products.

The underlying learning idea is rated moderate. Charitable reconstruction of opposing views is core to argumentation education (sometimes called the principle of charity). Evidence is more pedagogical than trial-dense.

AI delivery evidence

Whether an AI tutor delivers this method well is a separate question.

AI can draft a polished opposing case that still misses what real advocates care about. Verify important steels against real sources. Delivery remains speculative.