01
Why models hallucinate
An appliance company has just launched a support chatbot. A customer asks about its new product, the K2 electric kettle, and the chatbot replies with four smooth, confident sentences. In the scene below, check each sentence against the product manual.
- Capacity is 1.7 L
- Not checked yet
- It switches off when it boils
- Not checked yet
- The warranty is 2 years
- Not checked yet
- The body is dishwasher-safe
- Not checked yet
Question: "Tell me about the K2 electric kettle"
The model's answer looks smooth and confident. Turn on the check against the manual.
When a model states something false or unsupported as if it were fact, that's called a hallucination. There are two kinds. The answer can contradict facts about the world, or it can contradict — or go beyond — the source material you gave it. In the scene, "2-year warranty" contradicts the manual, and "dishwasher-safe" is something the model made up that the manual never says.
The reason lies in how LLMs produce text. An LLM doesn't look facts up before answering. It picks the most plausible words to come next. Descriptions of electric kettles often mention things like "2-year warranty" or "dishwasher-safe", so those phrases feel like a natural fit. The less the model knows — a product it never saw in training, news from after its training ended, a vague question — the more likely it is to fill the gaps with something made up.
If a question says… → pick this
- “Generates plausible content that is not true”Hallucination
- “Contradicts the provided documents, or adds things they don't say”Ungrounded answer (hallucination)
- “Cites papers or sources that don't exist”Hallucination (fabricated citations)
- “The answer sounds confident, so it must be right”Wrong — you can't tell from the tone
You can't get rid of hallucinations entirely. So how far can you reduce them?
02
Reducing hallucinations
There's no single fix. You reduce hallucinations by stacking several safeguards on top of each other. In the scene below, add them one at a time.
- Wrong answers
- 28%
Share of wrong answers as you add safeguards to a kettle support chatbot (illustrative)
With no safeguards, the model answers from what it knows and gets about three in ten wrong (28%).
| Method | How it helps |
|---|---|
| Ground answers with RAG | Instead of making up what it doesn't know, the model reads the retrieved material and answers from it |
| Tell it to say "I don't know" | Keeps it from inventing a plausible answer when the material doesn't cover the question |
| Citations and grounding checks | Ask for a source behind each sentence, and automatically flag sentences that don't match the sources |
| Ask it to think step by step | Cuts down on mistakes in multi-step calculations and reasoning |
| Human review | A person checks answers that would be dangerous if wrong before they go out |
Two misunderstandings come up a lot. First, lowering the temperature makes answers more consistent, not more true. Second, fine-tuning a model to teach it facts isn't a good fix. The model won't always recall those facts accurately, and whenever the information changes you have to train it again. Use RAG to supply facts, and save fine-tuning for style and format.
If a question says… → pick this
- “Reduce hallucinations by grounding answers in up-to-date company data”RAG (grounding)
- “Don't make things up when the material has no answer”Instruct it to say "I don't know"
- “Automatically check that each sentence is backed by the source”Grounding check
- “Setting temperature to 0 eliminates hallucinations”Wrong
- “Answers in a field where mistakes are dangerous”Human review (human-in-the-loop)
Hallucination is the model making mistakes on its own. But there's another problem: users who deliberately try to pull harmful answers out of it, and a model saying things it never should.
Pro only
That’s the free part of this guide
The other 5 chapters are open to Pro subscribers. Each one continues with 3D visuals and “if a question says…” tips, and the last chapter lets you check yourself with a short quiz.
- 03Guardrails
- 04Content filters
- 05Protecting personal data
- 06Red teaming and monitoring
- 07Choosing a safeguard
Advertisement
PasslyExam
You have read the concept. Now test it on exam questions
PasslyExam offers original practice questions, explanations and full mock exams built around each certification’s exam guide. The explanations cover why the right option is right and why the look-alike options are wrong. Sign up to practice for free every day.