Why AI makes mistakes

You practiced catching hallucinations: outputs that sound polished and specific but aren’t grounded in reality — fake quotes, invented book titles, wrong dates, plausible APIs that don’t exist. You also brushed the wider family of misses: outdated knowledge, misread prompts, and overconfident guessing when the model should say “I’m not sure.” The scary part isn’t that AI fails; it’s that it often fails while sounding helpful.

Why this happens (plain English):

1. Predicting next words isn’t the same as looking something up. Generative models often continue patterns that look like answers. A real citation may or may not exist.
2. Training averages the web’s habits. The web mixes facts, fanfic, and marketing. The model learns fluent structure from all of it.
3. No built-in embarrassment. Humans hedge when unsure; models can narrate fiction with the calm of a news anchor.
4. Your prompt can trap it. Ask for “exact quote from chapter 7” of a book it doesn’t have, and it may invent rather than refuse.

Everyday disaster tape:

• Trip planning: confident museum hours that closed last year.
• Work: a “source” link that 404s.
• Code: a perfect-looking function for a library method that never shipped.

Detective habits: ask for sources you can open; test one factual claim; try a second tool or human; watch for too-specific numbers with no trail.

• Health, money, legal, or safety decisions — verify outside the chat before you move.
• School essays — treat citations as suspects until checked.
• Anything you’ll paste to a boss or client under your name.

Tools with live search reduce some stale-fact errors; they don’t erase spin, bad sources, or creative fabrication in the drafting layer. Also, a wrong answer can still be partly useful (good outline, bad statistic) — edit, don’t worship.

Ask any chat AI for a specific fact you’ll verify in 60 seconds (a local store’s Sunday hours, a public holiday date). Compare tone vs truth. Note the gap.