You ask a store assistant, “Does the other branch still have this berry-red bag?”
She never checks the computer or calls anyone, yet answers smoothly: “Yes. There is one left. I can reserve it for you.”
You travel across town and discover that branch stopped carrying the color three months ago.
The problem is not that she did not know.
The problem is that she did not know, but sounded completely certain.
That is a useful entrance to the idea of an AI hallucination.
A smooth answer is not proof that the AI checked anything
Generative AI is very good at producing text that looks like a natural continuation of what came before.
That ability makes answers fluent and useful. It also creates a risk: a polished, detailed and confident sentence can still be wrong.
Ask about a paper that does not exist and a model may invent a plausible title, author list or year. Ask about an old limited-edition product and it may combine nearby facts into an answer that sounds perfectly believable.
The first safety idea is simple:
Confidence in tone and correctness in fact are two different things.
Why not just say “I don’t know”?
A Large Language Model (LLM) is fundamentally built to generate a likely continuation from the available context. It is not automatically a database that verifies every claim against a trusted record before speaking.
When reliable information is missing, the model may still generate something that has the shape of a good answer.
That is like a charismatic store assistant guessing a release year from memory. The sentence can be elegant even when the date is wrong.
What should you verify more carefully?
Extra checking is especially important for:
- medicine and health,
- finance, investment and tax,
- legal rules,
- prices, dates and current information,
- exact statistics,
- papers, news, URLs and citations,
- anything you are using for an important decision.
AI can still help with these topics. The cost of an error is simply higher.
Does asking for sources solve the problem?
Not automatically.
A model that can generate a false fact can also generate a source that merely looks real. Open the link. Check that the source actually says what the answer claims. Check the date.
Search tools and trusted databases can improve reliability, but it is still useful to distinguish between information the system actually retrieved and text the model generated itself.
You do not need to stop using AI
Knowing about hallucination is not a reason to distrust everything.
It is more like knowing online product photos can have color differences: for an important purchase, you check more photos and confirm the measurements.
Understanding the limitation helps you use the tool with more confidence, not less.
One thing to remember today
An AI hallucination is incorrect or misleading content produced by a model, often in language that sounds fluent and convincing.
After six lessons, you now have a useful foundation: AI, models, training, tokens, prompts and hallucination.