Skip to content
EN

Lesson 5 of 8 · 6 min · +100 XP

Why does AI make things up?

An AI hallucination is a confident answer that is false or invented, such as a made-up fact, quote or source. It happens because a chatbot predicts likely-sounding words rather than checking facts, so when it lacks information it still produces something fluent. Be most careful with names, numbers, quotes, citations and recent events, and verify them elsewhere.

▶ Try it

Why it matters

A language model always produces likely words, even when it has no real answer. The result can be a polished, confident and completely invented fact, quote or source.

In 6 minutes you'll be able to

  • AI hallucinations happen because chatbots predict likely words, not checked facts.
  • Names, numbers, quotes, citations and recent events are the hot spots.
  • Verify important details in a trusted source, not by asking the chatbot again.

What you'll practise

  1. Why it happens

    A hallucination isn't a glitch or a lie. It comes from how chatbots write.

    🔮 It predicts, it doesn't look up
    A chatbot writes the most likely next words. There's no built-in step that checks them against the facts.
    🕳️ Gaps get filled
    When the training text says little about a topic, the model still produces something that fits the pattern of an answer.
    😌 Always sounds sure
    The tone is the same whether the answer is right or wrong. Confidence tells you nothing.
    📌 Hot spots
    Errors are most common in names, dates, numbers, quotes, citations, niche topics and recent events.

  2. Spot the inventions

    Noor knows the real story: Lumen Lab was started by three students in 2019 as a small research lab. Tap every invented detail in Lumi's answer.

    Lumi
    Great question! Lumen Lab is a research lab that studies how machines learn. It was founded in 1987 by Dr Elena Varga, who also invented the first solar-powered calculator. She once said: “Light is the cleverest teacher.” You can read more in her book Bright Machines (2004), page 112

  3. Risky or fairly safe?

    Which question is more likely to get a made-up answer?

    1. Which of these should you double-check most?

      • “Give me three studies, with page numbers, about my topic” (Answer)
      • “Suggest a title for my holiday photo album”

    2. Lumi gives you the exact population of a tiny village. What's the smart move?

      • Check it in an official or trusted source (Answer)
      • Trust it, the number is so precise
      • Ask Lumi “Are you sure?” and accept its reply

  4. Habits that catch inventions

    Choose all the habits that help.

    • Look up names, dates and numbers in a trusted source
    • Open every cited source and check it says that
    • Search the exact quote to see if it really exists
    • Be extra careful with recent events and niche topics
    • Trust answers that sound confident and detailed
    • Accept a source because it has a page number

Remember

Fluent isn't true. Check names, numbers, quotes and sources outside the chat.

Fluent isn't true. Check names, numbers, quotes and sources before you rely on them.

Key facts

  • In June 2023, a US federal judge fined two lawyers and their firm $5,000 for submitting a court filing that cited court decisions invented by an AI chatbot. · Source: US District Court, Southern District of New York, Mata v. Avianca, Inc., opinion and sanctions order, 22 June 2023
  • The US National Institute of Standards and Technology lists “confabulation”, confidently stated but false content, as one of the main risks of generative AI. · Source: NIST, AI 600-1, Artificial Intelligence Risk Management Framework: Generative AI Profile, 2024
  • Some researchers prefer the word “confabulation” to “hallucination”, because the model isn't seeing things that aren't there; it is filling gaps with plausible-sounding content.
  • Asking a chatbot whether its own answer is correct is not a reliable check, because the reply is generated the same way as the original answer.

Updated:

Questions people ask

What is an example of an AI hallucination?

A chatbot inventing a book, study or court case that doesn't exist, complete with a realistic title, author and page number, or giving a precise but wrong date or statistic.

Why do AI chatbots hallucinate?

They are trained to produce likely-sounding text, not to check facts. When the model lacks information, it still generates a fluent answer that fits the pattern, even if it is false.

Can AI hallucinations be fixed?

They can be reduced, for example by connecting a chatbot to search or trusted documents and by better training, but no current method removes them completely. Checking important facts remains your job.

How can I tell if an AI is hallucinating?

You often can't from the text alone, because it sounds equally confident either way. Check names, numbers, quotes and sources in a trusted place outside the chat.

Related terms