Fact-checking an AI answer: 10 steps against wrong information

The AI always sounds confident. Even when it's inventing a number, a quote or a study right then. That's exactly the trap: the tone stays self-assured whether the answer is right or made up. If you work with AI answers, on the job, in an application, in an offer to a client, you need a fixed routine for checking facts. Not every answer, but every one that counts.

Here's such a routine. Ten steps you can go through even without programming knowledge. You don't always have to do all ten. For a casual question a quick glance is enough. For something that touches money, reputation or a decision, you go further down the list.

Step 1: Ask for the source first

Before you believe anything, ask the simplest control question: "How do you know that?" A good model names you a source or honestly says it isn't sure. A bad sign is when the source suddenly sounds just as invented as the answer itself. A vague answer like "that's common knowledge" isn't a source yet. It's a hint to look more closely.

Step 2: Check numbers and dates separately

Numbers are the most common place where AI gets it wrong. Percentages, years, prices, statistics. Take every concrete number from the answer and check it individually. The more precise a number sounds ("according to a 2024 study it's 37.4 percent"), the more suspicious you should be. Invented numbers often carry a false precision meant to look real.

Step 3: Cross-check names, titles and quotes

Invented book titles with real authors, misattributed quotes, papers that never existed. That's a classic hallucination pattern. Copy the name or the quote and search for it yourself once. If you find nothing on a supposedly famous quote except the AI answer itself, it probably doesn't exist.

Step 4: Open the links yourself

When the AI gives you links, click them. Not just look, open. URLs that look plausible but lead nowhere are a standard error. And even when a link works: check whether the page really says what the AI claims. Sometimes the source exists but says something completely different.

Step 5: Ask the same question differently

Ask the same question a second time, but with different words. If the answer stays the same at its core, that's a good sign. If you get two different, contradictory answers, you've found a spot where the model is guessing. This consistency test costs you a minute and reveals surprisingly much.

Step 6: Have a second AI cross-check

Take the answer and put it to another AI tool. Don't ask "is this right", but "what about this statement is wrong or inaccurate". Different models were trained differently and don't make the same mistakes. Where two independent models agree, the probability is higher that it's correct. Where they contradict each other, you know where you have to check yourself.

Step 7: For expert topics, a human source

Law, medicine, finance, taxes. For anything where a mistake gets expensive, the AI is a starting point, not the answer. Use it to formulate the right questions, but get the binding information from a human who's liable for it. The AI knows no individual case and no current legal situation in your region.

Step 8: Watch the timeliness

An AI model knows the world only up to its training cutoff. For anything that changes fast, current prices, new versions, the latest legal changes, events of the last weeks, the model can be out of date without knowing it. When in doubt, ask directly: "Up to when does your knowledge reach?" And always check time-critical facts against a current source.

Step 9: Ask about uncertainty directly

Models can be surprisingly honest when you ask them. "How sure are you about this statement, on a scale of one to ten?" or "Which part of this answer is the least backed up?" That forces the model to mark the weak spots itself. Don't take this self-assessment as proof, but as a map of where you should look more closely.

Step 10: Set your own rule of thumb

You won't completely fact-check every answer, that would be too much. So decide for yourself from when you check. A usable rule: anything you pass on, publish or take as the basis for a decision gets checked. Anything that's just for you to brainstorm may stay unchecked. Once you draw this line clearly, checking becomes a habit instead of a burden.

What comes next

If you want to understand why the AI even gets things so confidently wrong, read the lesson on hallucinations in Level 1. It explains the pattern behind the mistakes. And if you research regularly, look at how you give the AI real, current sources through a research tool instead of relying on its memory. Fact-checking then gets much faster, because the sources come along right away.