Before you trust an AI answer, check these five things

An AI answer can be clear, detailed and completely wrong. The difficult cases are often believable: a real company attached to an invented feature, a plausible quotation with no original source, or a correct formula applied to the wrong units. The useful skill is deciding what to check first.

NIST identifies confidently incorrect generated content as a distinct generative-AI risk. OpenAI also advises checking important facts, references and quotations. The workflow below is an editorial method for turning that general warning into a repeatable habit.

1. Separate the answer into claims

Do not try to verify a whole paragraph at once. Extract the statements that affect your decision. For a proposed software purchase, these might be the price, supported devices, export options, storage allowance and cancellation terms. For a calculation, identify the inputs, formula, units and result.

Put the claims in a short checklist. Mark each one as a fact, an inference or a suggestion. “The free plan includes export” is a fact to verify. “This would suit a small team” is a judgment that depends on your needs. Treating both as equivalent makes review harder.

2. Find the original evidence

Ask for a source, then open it yourself. A link does not prove that the nearby sentence is supported. Look for the exact feature, number or qualification in the page. Prefer the product's own documentation for specifications, the original paper for research findings and the relevant authority for rules.

Check that the source applies to the right country, product version and account type. An enterprise-only feature is not evidence that a personal free account includes it. Record the date you checked a changing claim.

3. Recalculate important numbers

Suppose an answer says that 18 items costing ₹240 each total ₹4,230. Multiplying independently gives ₹4,320. This invented example illustrates why polished explanations are not a substitute for arithmetic.

For percentages, write the denominator explicitly. A reduction from 50 to 40 is 20% of the original value; increasing 40 back to 50 requires 25%. Check the units before checking the arithmetic. Minutes and hours, grams and kilograms, or monthly and yearly prices can produce convincing but unusable results.

4. Look for missing conditions

Ask what would make the recommendation fail. Does a tutorial require a desktop app? Does a quoted price exclude tax? Does a test describe a controlled environment rather than normal use? Ask the model to identify uncertainties, but verify those answers too.

For high-consequence decisions, such as medical treatment, legal obligations or safety-critical work, use an appropriately qualified professional and authoritative guidance. A second chatbot repeating the first answer is not independent verification.

5. Keep a small evidence trail

Save the source URL, relevant date and your checked conclusion alongside the finished work. You do not need a large research database for every task. A three-column note—claim, source, checked result—is enough to make later updates easier.

Use AI for organizing, explaining and drafting after the evidence is settled. If a claim cannot be supported, remove it, qualify it clearly or leave the decision open. “Not yet verified” is useful information; a confident guess is not.