← Explainers
2 min read

What is an AI hallucination?

Why AI sometimes makes things up with total confidence, and how to protect yourself.

BuzzwordsAI basics
Watch it in about a minute

AI can write a brilliant answer, and then invent a fact, a quote or a source that doesn’t exist. This is called a hallucination, and it’s one of the most important things to understand before trusting AI.

Term
Hallucination
Confident, but wrong

When an AI model gives an answer that sounds confident and believable, but is wrong or made up.

Think of it like this

A student in an exam who never writes “I don’t know”. Faced with a question they can’t answer, they write something that sounds right, in the right style, with total confidence.

What it looks like

You Can you give me a source for that statistic?
AI Of course. It comes from Thompson & Reyes (2019), 'Customer Patience in Digital Channels', Journal of Service Research, vol. 22.
This source is invented for this example. It looks real, and that's the problem.

Why it happens

  1. 1It predicts, it doesn't look up — an LLM writes the most likely-sounding next words
  2. 2No built-in fact check — sounding right and being right are different things
  3. 3Gaps get filled — when it doesn't know, it can fill the gap with something plausible

How to reduce it

What helps
  • Give it trusted sources (RAG)Most effective
  • Ask it to say when it's unsureHelps
  • Ask for sources, then check themHelps
  • Verify anything important yourselfAlways

Where to be most careful

Names, dates, numbers, quotes, citations, legal or medical details, and anything about very recent events. These are where made-up answers are easiest to miss.

A common misunderstanding

“It’s lying.” Lying means knowing the truth and hiding it. The model has no sense of true or false; it produces plausible text. That’s exactly why checking matters.

In the real world

In 2022 Air Canada’s website chatbot told a customer he could claim a bereavement discount after his trip. The airline’s own policy said he could not. In February 2024 a Canadian tribunal ruled that the airline was responsible for what its chatbot said and ordered it to pay the difference.

LLM · RAG · Temperature · Evals