Checking the answer
An AI answer can sound sure and still be wrong. Like a route planner that sends you down a closed road. You still look out of the window. Learn three ways to check, and when checking matters most.
First, what do you think?
You ask a chatbot "Are you sure?" and it says yes. Is the answer checked now? Have a guess.
What is checking the answer?
The whole story in three sentences. An AI answer is a prediction, so it can be wrong. Checking means comparing it with something outside the chatbot. There are three ways to do that. Click through, look in a second source, or test it.
Think of a route planner. Most of the time it gets you there. But sometimes it sends you down a road that's closed today. You don't drive into the barrier. You look out of the window.
- Click through to the source. Did the tool name a website? Open it, and read what it says there.
- Look in a second source. A reliable website, a book, an expert. Does it say the same?
- Test it yourself. Run the code. Do the sum again. Try the recipe on a small scale.
How much you check depends on what's at stake. A fun fact for a quiz? A quick look is enough. Your health, money or a school report? Check carefully.
What the predictor does
The predictor gives the most likely next words, not checked facts. That's the rule.
Researchers explain why it sometimes makes things up. Training rewards guessing over saying "I don't know". In one test, a model gave three different wrong dates for one researcher's birthday. Three chatbots each made up a wrong title for his dissertation.
So asking again is no check. You may get a different answer, also wrong. And a sure tone says nothing. The predictor sounds sure because sure-sounding text is likely text.
Makers say so themselves. Google writes that AI answers may include mistakes. About its research tool, it adds: check the answers.
The route planner is usually right, but you still look out of the window. Click through, look in a second source, or test it.
Try it yourself
Eight AI answers. Which check fits best?
Everything runs in your browser. Nothing you type is sent anywhere.
How do you check an answer?
Four steps. Use them whenever an answer matters.
Decide how much is at stake
Fun or important? Health, money, law and school work count as important. Then check carefully.
Click through to the source
A search tool can show sources. Open the link. Does the page really say this? Does the page exist at all?
Look in a second source
Find the fact somewhere else, not via the chatbot. Choose a source that knows: an official website, an encyclopedia, a doctor.
Test what you can test
Code, sums, instructions. Try them out. One study looked at people with an AI assistant. They wrote less secure code, and felt more sure of it.
In practice
Three people meet a closed road.
Case 1The history homework
Daan (14) asks a chatbot when a famous treaty was signed. He gets a year, and a quote from a king.
Copy both into the homework.
A year and a quote can both be made up. They sound as sure as true facts.
Look up the year in a second source. Search for the quote too. Can't find it? Leave it out.
Only checked facts go into his homework.
Case 2The list of sources
Rick (30) asks an AI tool for five sources for his report. He gets five titles, with writers.
Paste the list into the report.
Chatbots make up titles. In the study above, three chatbots each gave a wrong dissertation title.
Look up every title. Keep only what he can find and open.
His report doesn't cite books that don't exist.
Case 3The two medicines
Ellen (52) asks whether two medicines can be taken together. The chatbot says yes, very clearly.
Trust the clear answer.
A sure tone says nothing. Health is too important for a prediction.
Ask the pharmacist, or read the leaflet in the box.
The answer comes from someone who knows, and who is responsible.
When not to
- Don't ask the chatbot to check itselfIt's the same predictor.Check outside the chatbot.
- Don't trust a sure toneSure-sounding text is likely text.Look for a source.
- Don't let AI decide on health, money or lawA mistake costs too much.Ask a professional.
By situation
Check every fact you hand in. Name the source you checked, not the chatbot.
Check numbers, names and quotes before you send anything. Test code before it goes live.
Fun facts and ideas? A quick look is fine. Health, money or contracts? Ask someone who knows.
Myths and mistakes
| Claim | True? | Why |
|---|---|---|
| If it says "I'm sure", it's checked | No | It's still a prediction. Check outside the chatbot. |
| Asking again is a check | No | One model gave three different wrong dates. |
| A source list is always real | No | Chatbots sometimes make up titles. Look them up. |
| Checking is only for school | No | Code, health and money need it even more. |
Related topics
Read first
- Hallucinationswhy the predictor makes things up.
Next
- Privacy and your datawhat you type in, and what you don't.
Related
- Search and researchtools that show their sources.
Sources
- researchKalai e.a. (2025). Why Language Models Hallucinate.Language models hallucinate because training and testing reward guessing over saying 'I don't know'. Like a student on a multiple-choice test, guessing scores better. Example: asked for one researcher's birthday, a model gave three different wrong dates. Three chatbots each made up a wrong title for his dissertation.
- documentationGoogle. AI Overviews in Google Search.An AI Overview is an AI-generated summary above the search results, with links to sources. AI responses may include mistakes. You can't turn it off, but the Web filter shows only links.
- documentationGoogle. Gemini Deep Research.Deep Research searches hundreds of websites by itself and writes a report of several pages within minutes. Google adds: check the answers.
- documentationAnthropic. Web search tool.Web search lets a model answer with up-to-date information beyond its knowledge cutoff, with citations. It searches for recent news, current prices and scores; for stable facts, maths or writing it answers directly. Search results go into the context window.
- researchPerry e.a. (2023). Do Users Write More Insecure Code with AI Assistants?.People with an AI assistant wrote significantly less secure code, and were more likely to believe their code was secure. Those who trusted the AI less did better.
Test
You get 7 questions about the whole page. Get 5 right, and you've mastered the topic.
About the tools in this course. Every tool is described in the same way, in alphabetical order, without a recommendation. Names, prices and features change often: the maker's own site is always right. This site was built with the help of Claude, one of the tools described here.