Sections
ESEN

AI hallucinations: why ChatGPT makes things up and how to avoid it

AI sometimes answers with false information and total confidence. We explain why it happens, how to spot it and seven simple habits so you don't get fooled.

Official sourcePublished · 3 min read
Yellow smiley faces reflected many times in curved mirrors
Image: Carola68 on Pixabay

You ask an AI assistant something and it answers confidently, with facts, names and even dates. It all sounds perfect… but it’s false. This is called a hallucination, and it happens to all of them: ChatGPT, Gemini, Claude and the rest.

What a hallucination is

An answer that looks right but isn’t: a made-up fact, a quote nobody said, a book that doesn’t exist or a link that goes nowhere. The dangerous part is that it comes with the same confidence as a correct answer.

Why it happens

Remember that a language model works by predicting which word comes next (we explain it in what AI is and how it works). That has consequences:

  1. It aims for what sounds right, not what is true. If it doesn’t know a fact, the “most likely” next word can be an invented one that fits the sentence nicely.
  2. It struggles to say “I don’t know”. OpenAI’s research suggests that the way models have been trained and evaluated for years has rewarded guessing over admitting uncertainty, like a multiple-choice exam with no penalty for wrong answers.
  3. Its memory has a date. What it learned stops at a certain point. Ask about something later and, if it doesn’t search the web, it may fill the gaps.
  4. Rare facts fail more often. The less that’s been written about something (a small town, a little-known person, a local law), the easier it is to get wrong.
When AI is most likely to be wrong Low risk: explaining general concepts and drafting. Medium risk: summarizing a text you provide. High risk: exact figures, quotes, sources, recent news, local topics and little-known people. RISK OF MAKING IT UP LOWExplaining conceptsDrafting and rewritingBrainstorming MEDIUMSummarizing a textyou provideTranslating HIGHExact figures, dates, quotesSources and linksRecent newsLocal or obscure topics
A rough guide: the more specific and obscure the fact, the more careful you should be. Chart by Best AI Today.

How to spot one

Be especially suspicious when you see:

  • Very precise figures with no explanation of where they come from.
  • Direct quotes attributed to someone.
  • Titles of books, studies or articles you’ve never heard of.
  • Links: always open them; sometimes they don’t exist.
  • Answers that change when you ask the same question again.

7 habits so you don’t get fooled

  1. Ask for sources and open them yourself. If it can’t provide any, treat the fact as doubtful.
  2. Turn on web search for recent topics. Most assistants have it.
  3. Give it the information yourself. If you paste the text, PDF or web page and ask it to answer only from that, it gets things wrong far less.
  4. Give it permission to doubt: “If you’re not sure, say so. Don’t make up facts.”
  5. Ask twice or rephrase. If the answers don’t match, something’s off.
  6. Double-check what matters against a reliable source: the official website, a reputable outlet or a professional.
  7. Never use it as your only source for health, money or legal matters.

In short: AI is a great help for writing, learning and getting organized, but treat its facts like those of a very smart friend who sometimes gets mixed up: listen, then check.

Want to ask in a way that reduces mistakes? More tips in how to write good prompts.

Sources

Keep reading