Critical Thinking Is Domain-Bound

Critical thinking is not a portable skill that, once learned, can be used on anything. The processes of thinking are intertwined with the content of thought: you can only think critically about what you know enough about, and only when something cues you that this is the moment to do it (Willingham — Critical Thinking, Why Is It So Hard to Teach).

What "critical thinking" is

From a cognitive-science view it isn't its own faculty. It is ordinary reasoning, judgment/decision-making, and problem solving done in a way that is:

• Effective: it dodges known pitfalls (one-sidedness, discounting disconfirming evidence, reasoning from passion, unsupported claims).
• Novel: it isn't recall. Applying a memorized algorithm doesn't count; inventing one does.
• Self-directed: nobody is prompting each step.

Why it isn't a skill

A real skill such as reading music is available whenever you want it. Critical thinking isn't:

• It shows up without training: 3-year-olds reason about conditional probability (the "blicket detector").
• It fails despite training: physicians discount data that contradict their diagnosis, and PhD scientists reason badly in unfamiliar contexts.
• It clings to its home context: the student who estimates in math doesn't estimate in chemistry (Surface vs Deep Structure).

So the bottleneck is mostly not having the way of thinking. It is (a) recognizing when to deploy it and (b) knowing enough to execute it. In Willingham's words, teaching it lies "in small part in showing them new ways of thinking, and in large part in enabling them to deploy the right type of thinking at the right time."

The three-part bottom line

1. There is no context-free set of critical-thinking skills.
2. Metacognitive Strategies (learnable maxims) make critical thinking more likely.
3. Actually doing what the maxim says depends on domain knowledge and practice.

Nuance: with very extensive training (PhD-level), critical thinking gains some skill-like breadth across less-familiar content. Even then, people still make mistakes.

Implications beyond the classroom

• For the AI-and-work thesis: the wiki's AI pages say judgment is the durable human contribution (The Bottleneck Shift, AI Commoditizes Expertise). This page says that judgment isn't a free-floating faculty that survives without knowledge. It is built from domain knowledge and practice. That has a sharp corollary, logged as an open question below.
• For this wiki's method: distilled heuristics are maxims, so by this account they're cues, not capabilities. See Method Reflection.

Open questions

• If people offload domain knowledge to AI ("I don't need to know it, I can ask"), Willingham's account predicts their judgment in that domain weakens too, because judgment runs on knowledge held in your head. Is AI-era "judgment as the human premium" self-undermining unless people deliberately keep building knowledge? First evidence, 2026-11-08: Anthropic — Investigating Vault Particles with Claude reports that a team working outside its expertise caught AI errors quickly "in areas we knew well," while errors "cost us time in areas we didn't" — a two-week loss in one case. That is not judgment eroding; it is judgment that was never there, with a measured price. It makes the prediction concrete without settling the erosion version. See Verification Requires Domain Knowledge.
• Does the PhD-level exception (skill-like breadth after extensive training) describe senior engineers too? If so, what is the minimum depth at which heuristics start to transfer?

Connections

• Surface vs Deep Structure — the mechanism behind transfer failure.
• Metacognitive Strategies — the teachable part, and its limit.
• Falsification Is Asymmetric — an instance of a "critical" move (seek the disconfirming case) that Willingham says people do inconsistently.
• The Bottleneck Shift / AI Commoditizes Expertise — why human judgment is scarce: it is knowledge-bound, so it can't be cheaply copied.
• What's Scarce When Execution Is Cheap — adds a cognitive pillar to that synthesis's economic, organizational and information-theoretic pillars.
• Verification Requires Domain Knowledge — the AI-era corollary: you can only check what you know.
• Method Reflection — T1 and T3 there follow directly from this page.
• Distributions over Best Fits — see also (not grounding): a science version of the offloading worry. Hodgkin and Huxley "understood what the model was telling them because they had built it themselves, and they knew what hadn't worked" (Anthropic — Teaching AI to Build Scientific Models). When AI builds the model, that understanding has to be handed over explicitly or it is lost.

Sources

• Willingham — Critical Thinking, Why Is It So Hard to Teach
• Anthropic — Investigating Vault Particles with Claude — field evidence on the open question: error-catching tracked the team's own domain knowledge.