Understanding AI Hallucinations and How Prompts Can Reduce Them

Sometimes AI confidently states things that simply are not true. That is called a hallucination. Knowing why it happens — and which prompts make it less likely — is essential before you trust AI with anything important.

1. Introduction

An AI hallucination is when the model generates text that sounds right but isn't right. It might invent a citation, get a date wrong, fabricate a quote, or describe a feature that does not exist in some software. The result reads beautifully and looks confident — that is exactly what makes hallucinations dangerous.

2. The Concept Explained

To understand hallucinations, remember what the model is actually doing: predicting the next likely token based on patterns it has seen. It is not "looking up" facts in a database. It is composing a plausible continuation.

Most of the time, "plausible" and "true" overlap. But when the AI is asked about something it does not really know — a niche topic, a recent event, a specific person, an unusual edge case — it still produces plausible text, because that is what it was trained to do. The result looks correct because the structure is correct. Only the content is fabricated.

One-line definition: A hallucination is fluent, confident-sounding output that has no factual basis.

Why hallucinations happen

• The model was never trained on that fact. It fills the gap with something that looks similar.
• The model has outdated training data. Anything after its cutoff date is unknown to it.
• The prompt is ambiguous. The AI guesses what you mean and answers the wrong question confidently.
• The temperature is too high. Higher randomness encourages more creative — and more invented — output. (More on this in Topic 8.)

3. What Hallucinations Look Like

Give me three academic papers that prove
the benefits of intermittent fasting.

Based ONLY on the abstracts I paste below,
summarise the main findings on intermittent
fasting. If a question is not answered in
these abstracts, say "Not covered in source".

Abstracts:
"""
… paste real abstracts here …
"""

4. Five Prompt Patterns That Reduce Hallucinations

Pattern 1 — Give the AI source material

The single most powerful technique. Don't ask the AI to recall facts — give it the facts and ask it to work on them. "Based only on the text below…" is the safest prompt phrasing for factual tasks.

Pattern 2 — Allow "I don't know" as a valid answer

If you are not certain, reply with
"I don't have reliable information for that."
Do not guess.

This gives the AI explicit permission to refuse — most hallucinations happen because the model feels it must give some answer.

Pattern 3 — Ask the AI to cite its source from your text

For every claim, quote the exact sentence
from the source text that supports it.

If the AI cannot find a supporting quote, it is forced to admit the claim has no basis.

Pattern 4 — Use Chain-of-Thought for factual tasks

Asking the AI to walk through its reasoning step by step exposes weak links — you can see where it guessed.

Pattern 5 — Lower the temperature

If your AI tool lets you set temperature (covered in the next topic), use 0 to 0.3 for factual tasks. Higher creativity is great for brainstorming but invites invention.

5. Step-by-Step: Building a Hallucination-Safe Prompt

1. Provide source material whenever possible. Paste documents, data, or transcripts.
2. Restrict the AI to that material. "Use only the source above." Repeat at the end.
3. Allow refusals. Tell it to say "not covered" or "unknown" when needed.
4. Demand quoted evidence. Make it cite the exact sentence supporting each claim.
5. Lower randomness. Use a low temperature setting if your tool exposes it.
6. Verify anyway. No technique is 100%. Always spot-check the important parts.

6. Practice Exercises

Recommend three books written before 2010 about prompt engineering, with author and publisher.

Summarise this article in 5 bullets. For each bullet, quote the sentence from the article that supports it.

7. Key Takeaways