What is critical thinking?

Critical thinking is the capacity to test a claim against evidence, to separate what you know from what you are assuming, to notice the reasons you might be wrong, and to revise accordingly. The load-bearing part is not intelligence. It is orientation: whether you are trying to defend a position or improve your picture of what is true.

This site has argued about what AI does to critical thinking for two years while listing the term itself as an open question. The definition below comes from the judgement and forecasting literature rather than from the AI debate, because the human mechanism is old and well studied even where the AI effect is neither.

Scout and soldier

Julia Galef's framing is the most useful available. A soldier evaluates a claim by asking whether it threatens or supports what they already hold. A scout asks what is actually out there. Both can be intelligent, informed and articulate; they differ in what they are trying to achieve. The Scout Mindset (2021).

This locates the skill somewhere unexpected. Analytical capability would predict it if the two were the same thing, and among clever people it plainly fails to. Galef puts the deciding factor in a set of habits about wanting to know, which are trainable and which most education leaves alone.

The one field that scores this is forecasting. Tetlock and Gardner's account of the Good Judgment Project found that accuracy over geopolitical questions is measurable and learnable, produced by probabilistic thinking, frequent updating and working in teams rather than by subject expertise or credentials. Superforecasting (2015). Subject expertise without feedback performed poorly, which lands on the same boundary Kahneman and Klein identified for intuition: where the environment returns no clear result, confidence detaches from accuracy. Graded entry.

What it is not

Scepticism is the common substitute, and it is cheap. Someone who disbelieves everything shares a defect with someone who believes everything: in both cases the conclusions have stopped tracking the evidence. What is being asked for is harder than either, because it requires holding a view firmly enough to act on while leaving it revisable.

It is not the same as reasoning either. Kahneman's two-system account describes fast associative processing and slow effortful processing, and catalogues the biases of the first. Thinking, Fast and Slow (2011). Critical thinking is not simply engaging the slow system. Careful reasoning can be deployed entirely in the service of a conclusion already chosen, which is what makes intelligent people good at defending errors.

And it is not the same as judgement. Judgement is recognising what a situation is. Critical thinking is testing a claim once it is on the table. They come apart: a person can evaluate an argument rigorously and still be answering the wrong question, which is the failure mode that matters most when the question arrives pre-framed by a machine.

What AI changes, and what it does not

The mechanism above predates computers by decades. Two things about it hold separately from any claim about AI, and are worth keeping apart from one.

The first is that producing a fluent, well-argued case for either side of a question now costs nothing. That removes a friction that used to do quiet work. Constructing a persuasive defence of a position used to take effort, and effort is a tax on motivated reasoning. When the defence is free, the soldier orientation gets cheaper to indulge and the scout orientation does not get any cheaper.

The second is that a model will supply confident prose regardless of the strength of the underlying case, because confidence is a property of the writing rather than of the knowledge. This site treats that at why does AI sound so confident when it is wrong.

Whether AI measurably weakens critical thinking is a different and unsettled question, handled at does AI weaken critical thinking. Three studies point the same way and all three have design weaknesses, so agreement between weak designs is suggestive rather than strong. The definition on this page does not depend on that result and should not be read as evidence for it.

The habits that show up in the scored data

Because forecasting is scored, it is the only place where advice about thinking has been checked against outcomes. Four things separate the accurate:

Confidence in probabilities rather than certainties. Say seventy per cent and you can be shown to have been wrong. Say likely and you cannot, which is what makes the second more comfortable and less useful.

Deliberately seeking disconfirming information. Not considering it when it arrives, which everyone believes they do, but going to look for it.

Keeping a record. Without one, memory reconstructs past beliefs to match present ones, and the feeling of having been broadly right survives almost any actual performance. This estate publishes its own dated predictions including the misses for that reason.

Separating identity from belief, so that revising costs less than defending. Adam Grant makes this the centre of his account of rethinking. Think Again (2021).

The AI-specific version follows from the first section rather than from new evidence. If a model will argue whichever side you signal you want, then the value of asking it depends entirely on whether you asked it to find the strongest case against your position or the strongest case for it. That is a question about use rather than about the model, and it is treated at how do I get AI to challenge me.

What this page does not establish

The definition is drawn from a literature that is largely about individual reasoning in laboratory and forecasting settings. Whether it transfers cleanly to professional work under time pressure is not settled, and Klein's naturalistic tradition would argue that experts in real conditions are doing something else entirely, which is recognition rather than evaluation.

The forecasting findings also come from questions with resolution dates. Most consequential judgements at work never resolve cleanly, and that is the environment in which Kahneman and Klein agree confidence deserves least trust. Advice derived from scored forecasting should be applied to unscored domains with that caveat attached.

And no claim is made here that AI degrades critical thinking. The mechanism described is a change in the cost of producing arguments, which is a fact about the tools rather than a measured effect on people.

Key sources

• Galef, J. (2021). The Scout Mindset: Why Some People See Things Clearly and Others Don't. Portfolio. In the essential works.
• Tetlock, P. E. and Gardner, D. (2015). Superforecasting: The Art and Science of Prediction. Crown. In the essential works.
• Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux. In the essential works.
• Kahneman, D. and Klein, G. (2009). Conditions for intuitive expertise: a failure to disagree. American Psychologist, 64(6), 515-526. Graded entry.
• Grant, A. (2021). Think Again: The Power of Knowing What You Don't Know. Viking. In the essential works.

Related SuperSkills research

On the measured question, does AI weaken critical thinking. On the adjacent capability, what is judgement. On using a model against yourself rather than for yourself, how do I get AI to challenge me. On why confident prose is not a signal, why does AI sound so confident. On the record this estate keeps of its own calls, predictions and corrections.

About this research

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. This page follows the rule this research uses for foundational concepts: established books answer the underlying human mechanism, and current studies answer what AI changes. Critical thinking is a long-standing field and none of the accounts here is a SuperSkills coinage.

How this research works  ·  Reviewed quarterly  ·  Found an error? Tell me and it is corrected on the page.