An AI response can be fluent, specific, and wrong. Critical thinking does not mean distrusting everything; it means matching the strength of verification to the consequences of error. A brainstorming suggestion and a production security decision should not receive the same level of confidence or review.
Identify the actual claim
Separate facts, interpretations, predictions, and recommendations. Rewrite the central claim in one sentence. Note dates, versions, locations, quantities, and implied assumptions. A broad answer often contains several smaller claims that need different sources or tests.
Ask what evidence would change your mind
Before searching for support, define the evidence that would confirm or reject the claim. Prefer primary documentation, direct measurements, source code, official records, or reproducible tests. Multiple articles repeating the same unsupported statement are not independent confirmation.
Check freshness and scope
Software versions, prices, policies, and product capabilities change. Verify the date and applicable environment. A technique that works in Node may not work in a browser; an Apps Script example may depend on authorization scopes or deployment settings. Confirm that the answer applies to the situation you actually have.
Look for missing alternatives
Ask which reasonable explanations or approaches were omitted. Compare tradeoffs, not only features. A generated recommendation may optimize for speed while ignoring maintenance, privacy, accessibility, or cost. Request the strongest objection and the conditions under which another option is better.
Scale verification to risk
For reversible experiments, run a small test. For consequential decisions, seek authoritative sources and qualified review. Record uncertainty instead of converting it into false precision. The goal is not perfect certainty; it is a transparent decision based on evidence proportionate to what could go wrong.
Use a claim ledger
For a research-heavy answer, create a small table with claim, importance, source, publication date, scope, confidence, and unresolved question. Verify high-impact claims first. Mark which conclusions are directly supported and which are your inference. This prevents a fluent paragraph from hiding one weak premise and makes it easier to update the answer when newer evidence appears.