Why a procedure beats intuition
Intuition works well in familiar settings but fails when information is new, numbers are murky, or a claim is fluently presented. A slower procedure separates the feeling of being right from the actual structure of an argument. In critical thinking, the smallest unit is not an opinion but an argument, made of three parts: a statement, a reason, and evidence. Once those parts are visible, questions have somewhere to land.
Five questions to ask before a claim
Who is speaking, and what do they gain?
Source does not determine truth, but it determines how much caution is warranted. A claim from someone with domain expertise, public data, and no major conflict of interest deserves more attention than one from an unverifiable source.
What exactly is being claimed?
Many disputes are confused because two sides address two different claims. Write the claim as a testable sentence: who, does what, under what conditions, to what degree. Words like always and everyone often hide the gap between a statistical tendency and an absolute rule.
Where are the reason and the evidence?
The reason is the bridge between evidence and conclusion. Evidence is data, observation, experiment, or text that can be checked. A claim with a conclusion but no bridge is not yet an argument.
Is there a counterexample or an alternative explanation?
This is the strongest question. One case with the same conditions but the opposite outcome refutes a universal claim. If another explanation fits the data equally well, the first one cannot be chosen merely because it is more comfortable.
Does the evidence actually support the conclusion?
A percentage can be misread without its denominator. A personal story can be emotionally powerful yet unrepresentative. This question forces a check of the whole path from data to conclusion.
Quick reference table
| Question | Goal | Warning sign |
|---|---|---|
| Who is speaking? | Assess source and interest | Anonymous, unverifiable source |
| What exactly is claimed? | Clarify scope and degree | Absolute words, unclear conditions |
| Where are reason and evidence? | Separate conclusion from data | Conclusion only, no bridge |
| Is there a counterexample? | Test universality | Ignoring contrary cases |
| Does evidence support the conclusion? | Check the logical path | Jumping from correlation to cause |
The most common fallacies
Confirmation bias
We tend to seek, recall, and believe information that matches what we already think, while ignoring contrary information. This is the foundational error because it damages the other four questions.
Sunk cost fallacy
We continue a choice only because we have already invested time, money, or effort, even when new data says stop. Sunk costs are unrecoverable and are not a valid reason to continue.
Ownership effect
We value what we already have more than what we do not, even when the two are equivalent. This makes changing our mind harder than it needs to be.
Hot-hand fallacy
We believe a random event will repeat in a streak, for example that someone on a hot streak will keep succeeding. For independent events, this belief has no probabilistic basis.
Base-rate neglect
We judge a specific case while forgetting the background rate of the phenomenon in the population. This is common when reading test results, news, or forecasts.
Straw man and tu quoque
The straw man attacks a distorted version of the opponent's argument. Tu quoque replies by pointing out the opponent's past errors instead of addressing the content.
Where schools disagree
There is no single list of fallacies accepted by every school. Some texts classify confirmation bias as a cognitive error, others as a logical one. Some place the hot-hand fallacy under probabilistic bias, others under reasoning error. This article follows the common convention in introductory critical thinking courses: cognitive errors and logical errors are separated, but they often travel together.
Can AI avoid these errors?
Not fully. A study published in Manufacturing & Service Operations Management tested GPT-3.5 and GPT-4 on 18 well-known cognitive biases. In nearly half of the scenarios, ChatGPT behaved like humans when facing emotional choices. GPT-4 outperformed GPT-3.5 on logic problems but showed stronger confirmation bias and was more prone to the hot-hand fallacy. Conversely, AI avoided some errors such as base-rate neglect and the sunk cost fallacy. The cause lies in training data and human-feedback fine-tuning. AI remains safer for well-defined tasks than for strategic or emotional judgment.
Practical takeaway
The five questions do not guarantee correctness, but they make errors easier to detect. The goal is to turn them into a reflex: write the claim down, look for contrary evidence, and check whether the logical bridge holds. When both humans and AI can make the same mistake, an external checking procedure becomes a necessary condition for any serious conclusion.
