How to check an AI answer before you rely on it

AI assistants are fluent, fast and sometimes wrong in a very convincing way. Here is how to tell which is which.

Guides · 4 min read · Written by the Taufa team · Updated 30 October 2026

An assistant writes the most likely next words given everything it has learned and everything you have told it. Usually that produces a correct, useful answer. Sometimes it produces something that reads exactly like a correct answer and is not: a made-up statistic, a law that does not exist, a book the author never wrote, a link that goes nowhere. The tone is the same in both cases, so tone tells you nothing. You need a routine.

Know where mistakes cluster

Specific facts: dates, figures, prices, names, addresses, phone numbers and the wording of laws and rules.
Quotations and citations: titles, authors, page numbers, court cases and web addresses can be invented while looking real.
Recent events: an assistant's built-in knowledge stops at some point. Without a live search, it may describe things as they were, not as they are now.
Multi-step calculations and logic puzzles: one early slip carries through to a wrong final answer.
Niche subjects: the less written about a topic, the more the assistant fills gaps with plausible guesses.
Your own situation: it does not know your local rules, your contract or your medical history unless you tell it.

A five-step routine

Decide how much it matters. A recipe idea needs a glance. A legal deadline, a dosage or a figure in a report needs real checking.
Find the checkable claims. Pull out the facts, numbers and names that your decision depends on. Ignore the rest.
Check them in a source that is not the assistant: an official website, a textbook, the original document, a reputable news report. Prefer the primary source over a page that repeats it.
Use two independent sources for anything important. If they disagree, find out why before choosing one.
Note what you could not confirm, and either leave it out or say clearly that it is unverified.

Ask the assistant to help you check it

The assistant can make checking easier, as long as you do not treat its own review as proof. Ask it to separate what it is sure of from what it is guessing, and to tell you what to verify.

A prompt to adapt

List every factual claim in your last answer that I should verify before using it: figures, dates, names, rules and quotations. For each, tell me how confident you are and what kind of source would confirm it. Do not invent sources.

Be wary of asking "Are you sure?". Assistants often change a correct answer when challenged, or defend a wrong one. A better question is a specific one: "What is the exact source for the 2019 figure?" If it cannot name a source you can find, treat the figure as unconfirmed.

Use research mode with sources, and still look

Some assistants can search the web and show where each statement came from. That is much better than an answer from memory, because you can click through. But a source being listed is not proof that it supports the claim. Open one or two of the links, find the passage, and confirm it says what the answer says. Check that the page is current and from a trustworthy publisher.

Check numbers by doing the sum

Redo the arithmetic yourself, or with a calculator, and compare.
If the assistant ran code on your data, ask to see the code and read what it filtered and counted.
Check units and currencies. Kilometres versus miles, lakh versus million and rupees versus dollars are classic traps.
Ask whether the figure is an estimate, a forecast or a measured value.

High-stakes topics need a person

For health, legal, tax, safety and financial decisions, use the assistant to understand the topic and to prepare good questions, and then confirm the answer with a qualified professional or the official authority. An assistant is not a doctor, lawyer or accountant, and it does not know your full circumstances. If someone is in danger or having a medical emergency, contact local emergency services right away.

Habits that make you harder to fool

Ask for the answer and the reasoning, then judge the reasoning.
Keep your question narrow. Broad questions invite more filler and more errors.
Give the assistant the document or data to work from instead of asking it to recall.
When you find a mistake, tell the assistant. It helps in the rest of that chat, though it will not carry over to other chats unless the service remembers it.
Keep a note of claims you verified, especially if you publish or present them.

Related guides

AI can be wrong, so check anything that matters. About Taufa · Start free