Why AI Makes Things Up and How to Avoid Being Fooled

Once I asked an AI about a book and it gave me the title, the author, the publisher, and even the year of publication. All with complete confidence. The problem: that book did not exist.

That’s called a hallucination, and it’s probably the AI trait that confuses first-time users the most. Not because it happens constantly, but because when it does, the AI gives no warning signal. It says the wrong thing just as confidently as when it’s right.

Why AI “makes things up”

To understand this without diving into theory, think of it this way: a text AI doesn’t “know” things the same way you do. Instead of looking things up in a reliable database, it predicts which word or idea is most likely given what you asked.

It’s like asking someone who has read thousands of books but doesn’t remember exactly where each piece of information came from. They can complete the pattern very well, but sometimes they fill in the gaps with what “sounds right,” not with what’s actually real.

That’s a hallucination: the AI completed the pattern with something plausible, but false.

The signals that should trigger your attention

Not every AI response is made up. In fact, the vast majority is useful and fairly accurate. The problem lies in specific types of information where the risk goes up considerably:

• Verbatim quotes: if the AI says “So-and-so said exactly this in 2018,” verify it. This is one of the most common errors.
• Statistics and studies: percentages, study names, and specific figures are high-risk territory.
• Dates, proper names, and current prices: the AI can mix up years, confuse people, or lack recent data.
• URLs and sources: it can invent a link that looks real but doesn’t exist.
• Very specific or technical information: the more niche the topic, the more likely it is to fill gaps with guesses.

There’s also a behavioral signal: when the answer is too perfect, complete, and detailed about something very specific that hasn’t been widely documented, that deserves more scrutiny.

Ways to verify without going crazy

You don’t need to look up every piece of data an AI gives you. But it’s worth developing a sense for when verification matters.

1. Ask the AI for its source. “Where does that data come from?” or “Can you cite the source of that study?” If it gives you a URL, copy it and check it in your browser. If it says it can’t provide exact sources, you already know that data needs confirmation.

2. Look up the key fact directly. If the AI mentions a study, search for it on Google Scholar or the institution’s official site. If it talks about a law, look it up in the official text. You don’t need to verify everything, just what you’re going to use in something important.

3. Ask the question from a different angle. If an AI says something that seems too convenient or vague, ask again differently: “Are you sure about that? Is there any chance you’re estimating?” With modern models, this sometimes triggers a self-correction.

4. Use AI for drafts, not for facts. For writing, organizing ideas, summarizing what you already know, or reviewing your text: AI excels. For data you’re going to cite in a report, a presentation, or an article: always verify before publishing.

The prompts that reduce the problem

How you ask the AI also affects how much it “guesses.” Some prompts that help:

• “If you’re unsure about a piece of data, tell me explicitly.” You give it permission to admit uncertainty instead of making something up.

• “Don’t invent sources or studies. If you don’t have a real source, say so.” Very useful when looking for citations or statistics.

• “Only give me what you know with high confidence about this topic.” Reduces responses padded with guesses.

• “Which parts of your response are confirmed facts and which are estimates?” You ask it to separate facts from assumptions.

These prompts don’t eliminate hallucinations 100%, but they reduce them considerably, especially with the most recent models.

Keeping it in perspective

Hallucinations are a real problem, but they’re not a reason to stop using AI. They’re a reason to use it with judgment: as a powerful tool that has limits, just like any other.

If you’ve already read When AI Makes Mistakes, you now have the full picture: there you learned how to detect the errors; here you learned why they happen and how to reduce them through prompting.

The most useful AI isn’t the one that never makes mistakes. It’s the one you use with enough judgment to know when to trust and when to verify. And that’s something you can learn, faster than you’d think.