Avoiding Hallucinations & Fact-Checking

Reviewed & published by Brayan K

AI models can state false things with total confidence — this is called a “hallucination”. They predict plausible-sounding text, not verified truth, so a wrong answer can arrive as smoothly as a right one. Knowing this changes how you use AI.

Part of the free Prompt Engineering course at LearnCodingFast — hands-on lessons with worked examples and the output they print, plus practice exercises and a quick quiz.

This lesson explains why models make things up, the prompts that reduce it, and the habit that matters most: verifying anything that counts.

What You'll Learn in This Lesson

• What a hallucination is and why models make things up
• Giving the AI permission to say 'I don't know'
• Which facts are riskiest: dates, stats, recent events
• Asking for sources — and why you must verify them
• Checking high-stakes claims with authoritative sources
• Treating AI answers as a draft to verify, not gospel

1 Why AI makes things up

An AI model generates the most plausible next words — not checked facts. If a wrong answer looks likely, it can come out just as confidently as a true one. There’s no built-in “am I sure?” alarm.

Give me the exact population of
this small town in 2023 and the
name of its mayor.

Obscure, precise, and recent — the perfect storm for a confident made-up answer.

What's the population of this town?
If you're not confident, say so and
tell me where I could verify it.
Don't guess at exact numbers.

You invite honesty and a path to verify, instead of forcing a guess.

2 Which answers to double-check

The more specific, recent, or consequential the fact, the more it deserves a second source.

3 Ask for sources — then verify them

Asking “what’s your source?” is good practice — but models can invent realistic-looking citations and URLs. Treat sources as leads to check, not proof:

• Click through — does the source exist and actually say that?
• Cross-check — does a second, independent source agree?
• Ask for confidence — “How confident are you, and what might make this wrong?”
• Go to a pro — for medical, legal, or financial decisions, confirm with an authoritative source or a professional.

📋 Copy-paste fact-checking template

Answer my question, but:
- If you're not confident, say so plainly.
- Don't invent numbers, names, or citations.
- List any sources I can check myself.
- Note how confident you are and what could
  make this answer wrong.

Try it for real

You cannot stop a model inventing things, but you can refuse to ship a claim no source supports — and that check is about fifteen lines. Watch the fourth claim especially: it cites a real policy, confidently, and that policy says nothing of the kind. Asking for citations is half the job; verifying them is the half that catches this.

# You cannot stop a model inventing things. You CAN refuse to ship a claim that
# no source supports. That check is code, and it is short.

sources = {
    "policy-1": "Refunds are available within 30 days of delivery.",
    "policy-2": "Faulty goods are collected free of charge.",
}

# A reply where every sentence carries the source it came from.
claims = [
    ("Refunds are available within 30 days.", "policy-1"),
    ("Collection of faulty goods is free.",   "policy-2"),
    ("Store credit never expires.",           "policy-3"),   # invented source
    ("Refunds take 24 hours to process.",     "policy-1"),   # real source, absent fact
]

def key_terms(text):
    stop = {"are", "is", "the", "of", "to", "a", "within", "take", "never"}
    return {w.strip(".,").lower() for w in text.split()} - stop

for text, src in claims:
    if src not in sources:
        print(f"REJECT  {text!r}\n        cites {src}, which does not exist")
        continue
    overlap = key_terms(text) & key_terms(sources[src])
    if len(overlap) >= 2:
        print(f"OK      {text!r}\n        supported by {src} (shared: {sorted(overlap)})")
    else:
        print(f"REVIEW  {text!r}\n        cites {src} but shares little with it: {sorted(overlap)}")

print()
print("The third claim invents a source, which is trivially catchable. The fourth")
print("is the dangerous one: a real source, confidently cited, that does not")
print("actually contain the fact. Requiring citations is only half the job —")
print("checking them is the other half.")

# 🎯 YOUR TURN — finish the citation checker.

sources = {
    "faq-1": "Delivery takes 3 to 5 working days.",
    "faq-2": "We ship to mainland UK only.",
}

claims = [
    ("Delivery takes 3 to 5 working days.", "faq-1"),
    ("We ship worldwide.",                  "faq-2"),
    ("Next day delivery is free.",          "faq-9"),
]

def check(text, src):
    # 1) A citation to a source you do not have is an automatic reject.
    if src ___ sources:
        # 👉 replace ___ with  not in
        return "REJECT (no such source)"

    words_claim  = {w.strip(".,").lower() for w in text.split()}
    words_source = {w.strip(".,").lower() for w in sources[src].split()}

    # 2) Support means the claim and its source share real wording.
    overlap = words_claim ___ words_source
    # 👉 replace ___ with  &

    return "OK" if len(overlap) >= 3 else "REVIEW (weak support)"

for text, src in claims:
    print(f"{check(text, src):<24} {text}")

# ✅ Expected output:
# OK                       Delivery takes 3 to 5 working days.
# REVIEW (weak support)    We ship worldwide.
# REJECT (no such source)  Next day delivery is free.

Common Mistakes (and the fix)

• Trusting a confident tone. — Sounding sure isn’t being right; verify the facts.
• Accepting citations at face value. — Click through; models can invent sources.
• Forcing an exact answer. — Let the AI say “I don’t know” instead of guessing.
• Not verifying high-stakes facts. — Medical, legal, and financial claims need an authoritative check.

⏱ Test Yourself — Timed Quiz

10 quick questions, 12 seconds each. Instant feedback — beat the clock!

🎉 Lesson Complete

• ✅ A hallucination is a confident but false answer
• ✅ Models predict plausible text, not verified truth
• ✅ Add “say you don’t know rather than guessing”
• ✅ Verify dates, stats, quotes, and recent events
• ✅ Ask for sources — then check they’re real and say what’s claimed

Practice quiz

Continue this course

• Previous: Iterating & Refining Prompts
• Next: Reusable Prompt Templates & Cheat Sheet — Copy-paste templates and a one-page recap of every technique

Frequently asked questions

What is an AI 'hallucination' and why does it happen?

A hallucination is when the AI states something false while sounding completely confident. It happens because these models generate the most plausible-sounding text, not verified truth. If a wrong answer looks likely given the words around it, the model may produce it just as smoothly as a correct one — there's no built-in 'am I sure?' check.

How do I get the AI to admit when it doesn't know?

Give it permission explicitly: add 'If you're not certain, say so instead of guessing' or 'Only answer if you're confident; otherwise tell me you don't know.' Without that nudge, models tend to fill the silence with a confident-sounding guess.

Can I trust the sources an AI gives me?

Treat them as leads, not proof. Asking for sources is good practice, but models can invent realistic-looking citations, authors, and URLs that don't exist. Always click through and confirm the source is real and actually says what the AI claims.

Which answers most need fact-checking?

Specific names, dates, statistics, quotes, legal or medical or financial claims, and anything about very recent events. Those are exactly where hallucinations cluster. Common, well-known facts are usually safe; precise or obscure details deserve a second source.