Spotting HallucinationsLesson 4 of 8
12 min read

How AI Makes Things Up (And How to Catch It)

Hallucination isn't lying


AI doesn't know what it doesn't know. When you ask it something it can't answer reliably, it doesn't say "I don't know." It generates the most statistically plausible answer it can — and presents it with the same confidence as everything else.


That's hallucination. Not deception. Just confident pattern-matching that happens to be wrong.

Where hallucinations cluster


Not all AI outputs are equally risky. Hallucinations cluster in predictable places:


  • Specific facts — exact numbers, dates, names, percentages
  • Citations and quotes — invented sources, misattributed quotes
  • Recent events — anything after the model's training data
  • Niche or rare domains — specialized topics with little training data
  • Edge cases of common topics — uncommon facts about famous things

    • The three-tier trust model


    As a working rule, sort AI outputs into three buckets:


  • TRUST LOW — specific facts, names, numbers, dates, citations, anything you'd put in a footnote. Always verify before using.
  • TRUST MEDIUM — general concepts, common patterns, well-established frameworks. Spot-check unfamiliar claims.
  • TRUST HIGH — structure, organization, brainstorming, writing assistance, rephrasing your own ideas. You're using AI as a tool, not a source.

    • Most of your AI use should be in the high-trust zone. Most of the trouble people get into is from using it in the low-trust zone without verifying.

    The "could I be confidently wrong?" test: before sharing anything AI told you, ask yourself — if this turned out to be wrong, would I look bad for not catching it? If yes, verify it. The 30 seconds of verification is always cheaper than the embarrassment.
    AI is not legal, medical, or financial advice. Even when it sounds confident and cites cases or studies, treat anything in those domains as research you need to verify with a licensed professional. Real harm has happened to people who skipped this step.

    Quick Check

    Which of these AI outputs would you trust the LEAST without verification?

    ←Compare and Contrast Without the BiasSource-Grounded Prompting→
    How AI Makes Things Up (And How to Catch It) — AI for Research & Analysis | Upgraide