Skip to content

Lesson 003Field guide · AI Safety & Privacy

Hallucinations Explained Simply

Why AI sometimes makes things up — and how to catch it.

Beginner11–14 min5 sections · exercise · quick check

You'll learn

  • What's actually happening when AI makes something up
  • Why fabrications come out fluent, specific, and detailed
  • The four zones where hallucinations cluster
  • Catch strategies that work — and why they work
  • Why this can't be switched off, and why that's okay

01Section

Prediction, not lookup

What AI Can and Can't Do gave you the definition: a hallucination is AI stating something false as if it were fact. This lesson explains the machinery — because once you see how fabrications happen, you can predict where they'll happen.

Here's the core of it: a chat AI doesn't look things up. There's no database of facts inside being searched when you ask a question. The model learned patterns from enormous amounts of text, and it answers by generating, word by word, the most likely continuation of the conversation. When the topic is well covered, the likely continuation is usually also the true one. When it isn't, the model still produces a likely-sounding continuation — it just isn't attached to reality anymore.

A useful analogy

Imagine someone who has read half a library but answers everything from memory, with the library doors locked — no shelf to check, ever. On famous topics, pattern-memory serves them well. Ask for the exact page number of an obscure quote, and they'll produce something page-number-shaped — because producing plausible answers is the only thing they can do.

Checkpoint

02Section

Why fabrications are fluent and specific

The unsettling part of hallucinations isn't that they happen — it's how good they look. A fabricated citation arrives with an author, a year, and a plausible title. That's not a malfunction. It's the machine doing its job.

The model's job is producing text shaped like a good answer. When you ask for a source, a good answer is shaped like a specific source — so that's what gets generated: named authors, precise years, realistic titles. Specificity is part of the pattern being completed, not evidence that anything behind it exists.

This also explains why the model rarely says "I don't know." In the text it learned from, questions are mostly followed by answers, not refusals. An answer-shaped response is almost always the more likely continuation — so uncertainty gets papered over with fluency unless you explicitly invite it.

Key takeaway

Detail is not evidence. A fabricated answer is specific for the same reason a real one is — specific is what good answers look like.

Quick knowledge check: Why do hallucinated citations come complete with authors, years, and titles?

Quick knowledge check

Why do hallucinated citations come complete with authors, years, and titles?

Checkpoint

03Section

The four high-risk zones

Hallucinations aren't spread evenly. They cluster where the model's patterns run thin — four zones account for most of them:

  • Citations & sources

    Studies, books, links, quotes. The demand for specificity is highest exactly where the model must generate rather than retrieve.

  • Niche topics

    Your industry's fine print, a small town's history, a specialized tool's settings. Thin training data means weak patterns — and confident filler.

  • Recent events

    Everything after the model's knowledge cutoff. You may get yesterday's world presented, fluently, as today's.

  • Precise values

    Exact figures, exact wording, exact dates. The pattern says "a number goes here" more strongly than it says which number.

Plain-language definition

Knowledge cutoff the date a model's training data ends. It has no built-in awareness of anything after that date. Some tools bolt on live web search, which helps — but a search-assisted answer still needs the same verification as any other.

Quick knowledge check: Which question is most likely to produce a hallucination?

Quick knowledge check

Which question is most likely to produce a hallucination?

Checkpoint

04Section

Catching them in the act

You can't spot a hallucination by looking at it — that's what the fluency is. But you can trap one:

  1. Scan the risk zones first: Before anything else, find the citations, niche specifics, recent-events claims, and precise values in the answer. That's where your verification minutes go.

  2. Click every link, check every title: Fabricated sources collapse on contact. If a study or book is load-bearing, confirm it exists before you build on it.

  3. Ask again in a fresh chat: Put the same question to a brand-new conversation. If the specifics change — different numbers, different sources — you've caught invention. Matching answers don't prove truth, but diverging ones prove trouble.

  4. Invite uncertainty in your prompt: Tell the model an honest "I'm not sure" is a good answer. It doesn't change the machinery, but it makes hedging a likelier continuation than guessing.

  5. Verify what's still load-bearing: Anything that survives and matters gets the Checking AI Outputs treatment: confirmed at an original source.

Fresh chat vs. "are you sure?"

Asking the same chat "are you sure?" applies social pressure to one answer. Asking a fresh chat re-rolls the dice independently — that's why disagreement between the two rolls is real information.

Copy and keep — the uncertainty invitation

Before you answer: if you're not confident about any specific fact, number, source, or date, say so directly instead of guessing. An honest "I'm not sure" is more useful to me than a plausible guess. End your answer by flagging anything I should verify elsewhere.

Checkpoint

05Section

Living with it

Hallucination isn't a bug that a future update will remove. It's the flip side of the machinery that makes AI useful — the freedom to generate likely text is what produces both your fluent drafts and the occasional fabricated citation. Tools keep getting better at grounding their answers, but "sometimes wrong, always fluent" remains the baseline to plan for.

And that's fine. You don't need AI to be infallible, any more than you need spellcheck to write your sentences. You need to know where it's weak — and now you know precisely: four zones, checkable in minutes.

Pause and think: Think of the last AI answer that impressed you with its detail. How many of those details sat in a risk zone — and did you check any of them?

Checkpoint

Prompt exercise

Catch one in the wild

Copy this prompt into ChatGPT, Claude, Gemini, Copilot, or whichever AI tool you have access to — the website doesn't run AI itself. Pick a genuinely niche topic: your specialty, your hometown, a hobby's fine print. Then verify every source it gives you. Either result teaches: real sources show the tool's range, shaky ones show you the machinery live.

Give me a reading list on [a niche topic you know well]: five specific sources, each with the author, the year, and one sentence on what it covers. Be as specific as you can.

Reflection: How many of the five checked out when you searched for them? A mixed list, delivered in one fluent confident answer — that's hallucination mechanics made visible.

Quick check

4 quick questions — no pressure

There's no pass or fail here. Answer them all, and we'll show you the answers either way.

1. When a chat AI answers a factual question, what is it actually doing?
2. Why does a fabricated answer include precise details like authors and years?
3. Which task sits in a high-risk zone for hallucinations?
4. You ask the same question in a fresh chat and get different specifics. What does that tell you?