How to check what AI tells you

What you will learn

• What a hallucination is and why it happens
• Which parts of an answer to check first: names, numbers, dates, quotes and links
• How to ask for sources and then check them yourself
• How to separate facts from the AI's guesses

The idea

AI tools write fluent, confident text whether or not the content is correct. When a tool produces something that sounds right but is false, such as a book that does not exist or a quote nobody said, it is called a hallucination: confident output that is wrong or made up. It happens because the model predicts likely text. It is not lying, and it does not know when it is wrong.

You do not need to check every word. Check the parts that would cause harm if they were wrong. Names, numbers, dates, quotes and links are the usual risk points, along with anything legal, medical or financial. For each one, go to a primary source: the original document, the official website, the law itself, or the person's own publication. A blog post that repeats the claim does not count as confirmation.

You can ask the AI for sources. That helps, but it is not the last step. A cited source may not exist, may be about something else, or may not say what the AI claims. Open it and find the sentence that supports the claim. If you cannot find it, treat the claim as unconfirmed.

Finally, ask the AI to label its own answer: which parts come from a source or from material you gave it, and which parts are general knowledge, guesses or opinions. The labels are not always right, but they show you where to look first.

Example

A small business owner asks:

When is the deadline for filing my quarterly sales tax return, and
what is the late fee? Give me a link to the official rule.

The answer gives a date, a fee and a link. It reads well. She checks each part:

• The link opens a page on the tax office website, but it is about a different form.
• The tax office's own page for her return lists a different deadline for her region.
• The late fee is not mentioned on any official page she can find.

She asks a follow-up:

Which parts of your answer come from a source you can point to, and
which parts are general knowledge or a guess? Put them in separate
lists.

The AI moves the fee into the guess list. She takes the deadline from the official page and calls the tax office about the fee. The AI helped her see what to look up, but every fact she acted on came from the primary source.

Practice now

1. Ask any AI tool a factual question about something you know well, such as your town, your job or a hobby.
2. Ask it to include at least two sources with links.
3. Underline every name, number, date and quote in the answer.
4. Open each source and look for the exact sentence that supports each underlined item.
5. Mark each item as confirmed, wrong or not found.

At the end you should have a short list of claims with a status for each, and a first sense of how much this tool needs checking in your area. For longer research, continue with Research with AI and check citations.

Copy this template

Answer this question: [your question].

Then split your answer into three lists:
1. Facts with a source I can check. Give the source name and link
   for each one.
2. General knowledge you did not take from a specific source.
3. Guesses, assumptions or opinions.

If you are not sure about a name, number, date or quote, say so next
to it. Do not invent sources.

Asking AI to flag uncertainty does not make its answer correct. Keep checking the parts that matter.

Check the output

• Does every link open, and is it the original or official page?
• Does the source say what the AI claims, in its own words?
• Are names spelled correctly, and do dates and numbers match the source exactly?
• Are quotes word for word, and said by the person named?
• Is the source current enough for your purpose? Look for a date on the page.
• Are guesses clearly separated from facts?

Common mistakes

• Checking only that a link works. A real page can still be about something else. Find the sentence that supports the claim.
• Asking the same AI whether it is sure. It may agree, or change its answer, without any new evidence. Check a primary source instead.
• Treating a confident tone as a sign of accuracy. Tone tells you nothing about truth. Judge each claim by its source.
• Checking everything with the same care. You will run out of time. Start with what would cause harm if wrong: money, health, legal points, and anything you will publish or send.