When AI Gets It Wrong

What a hallucination is

A hallucination is the industry's word for a moment when an AI states something false as if it were fact. Not a typo, not a vague answer — a confident, well-written, completely invented claim. The tone is exactly the same as when it's right, which is precisely what makes it tricky.

What does that look like in practice? A few real-world flavours:

Notice the pattern: hallucinations cluster around specifics — names, numbers, dates, titles, quotes. The general explanation around them is usually fine; it's the precise-sounding details that get invented.

Why it happens

This isn't a bug that will be patched next month. It comes from how these tools work at the core. As How AI Works explains, a chatbot is a next-word prediction engine: given everything so far, it produces the most plausible continuation, one word at a time. Its entire skill is completing patterns convincingly.

Most of the time, the most plausible continuation is also the true one — that's why these tools are so useful. But when the model doesn't actually "know" the answer, the machinery doesn't stop. It keeps generating what an answer would sound like. Ask for a citation and you'll get something shaped exactly like a citation: surname, initial, year, plausible journal. The pattern is perfect. The fact is fiction.

Two more facts complete the picture. First, the model has no internal fact-checker — there's no second system inside comparing its sentences against a database of truth before they reach you. Second, models are trained to be helpful and complete, so they rarely volunteer "I don't know" — unless you explicitly invite it, which we'll do below. The good news: newer models hallucinate less than early ones, and web-search modes help a lot. The honest news: no current tool is at zero, so the habit of checking stays.

Where the risk lives — and where it doesn't

Here's the empowering part: hallucination risk isn't spread evenly. It's concentrated in a few zones you can learn in one minute.

Why is the left side so safe? Because there, the AI isn't recalling anything — the raw material is in front of it, in your message. Rewriting your email, summarising the PDF you uploaded, brainstorming names for a café: there are no facts to invent. Why is the right side risky? Because you're asking it to pull precise details from memory. Its "memory" is patterns, not records — and patterns are thinnest exactly where topics are niche, local, or newer than its knowledge cutoff (the date its training data ends). Ask about a famous novel and it's on solid ground; ask about a bus schedule in your district town, and it may quietly start composing fiction.

The verification toolkit

You don't need to fact-check everything — that would make the tool useless. You need five cheap habits, applied where the spectrum says the risk is:

• Ask it to audit itself. "How confident are you in that answer? What parts might be wrong?" This works surprisingly well — the model will often flag its own weak spots, downgrade a claim to "commonly believed but disputed", or admit a number was an estimate. It's not a guarantee, but it's a ten-second filter that catches a lot.
• Ask for sources — then actually open them. The second half is the part everyone skips. A hallucinated source looks identical to a real one until you click it. If the link is dead, the book doesn't appear in any bookshop, or the article says something different — you've caught a fabrication in under a minute.
• Use web-search mode for factual questions. Most chatbots now have a mode where they search the live web and answer with linked sources. For anything recent, local, or numeric, switch it on. You get current information plus links you can check — this is the backbone of Research with AI.
• Cross-check important answers. Ask a second AI the same question, or spend thirty seconds on a search engine. Two independent tools agreeing doesn't prove truth, but disagreement is a loud alarm bell that saves you from acting on a fabrication.
• The golden rule: verify anything you'll act on or publish. Using AI's answer to satisfy your curiosity about Roman history? Relax. Using it in a report with your name on it, a payment, a medical decision, or advice to someone else? Verify first. The moment an answer is about to leave the chat window and touch the real world is the moment to check it.

Invite honesty — it works

Here's a trick most people never discover: models are far more willing to say "I'm not sure" when you make it clear that's a welcome answer. A single line in your prompt changes the behaviour:

I'm going to ask you a factual question. Rules:
- If you're not sure about something, say so clearly instead of guessing.
- Mark anything you can't verify as "unverified".
- "I don't know" is a perfectly good answer.

My question: [your question here]

And when you already have an answer in hand — especially one you're about to use — run it through this self-check before trusting it:

Review your previous answer as a strict fact-checker.
1. List every specific claim (names, numbers, dates, titles, quotes).
2. For each one, rate your confidence: high / medium / low.
3. Tell me which claims I should independently verify before acting on this, and where I could verify them.

Try that second prompt after any research-style conversation. Watching an AI mark its own homework — and honestly downgrade three of its claims to "low confidence" — is the fastest way to internalise everything on this page.

One more reason this page matters: fabricated content isn't always an accident. The same fluency that produces innocent hallucinations also powers deliberate fakes — that's the subject of Scams, Deepfakes & Fake Content. And before you paste sensitive material into a chat to verify it, read Privacy — What Not to Share.

Key takeaways

• A hallucination is a confident, fluent, invented "fact" — the tone gives no warning, so the tone proves nothing.
• It happens because AI completes patterns rather than looking up records: "sounds right" is its skill, and there's no internal fact-checker.
• Risk is highest for specifics — numbers, citations, quotes, laws, medicine, niche or recent topics — and lowest when it works on text you supplied.
• The toolkit: ask it to rate its own confidence, open the sources it cites, use web-search mode for facts, cross-check important answers.
• Golden rule: verify anything you'll act on or publish. Everything else can stay relaxed.
• Invite honesty explicitly — "if you're not sure, say so" — and the model becomes noticeably more trustworthy.