What a "hallucination" actually is
Remember Lesson 1: the model is predicting plausible-sounding text, not retrieving verified facts. Most of the time, "plausible" and "true" line up, because it learned from mostly-accurate writing. But when it's asked about something obscure, recent, or oddly specific — a citation, a statistic, a small business's exact founding date — it will sometimes generate an answer that sounds exactly as confident as a correct one, but isn't. That's a hallucination: not a glitch or a lie, just the predicting-machine doing its normal job in a spot where it doesn't actually have the answer.
The unsettling part isn't that it's wrong — every tool is wrong sometimes. It's that it doesn't sound less sure when it's wrong. There's no stammer, no "I think." That confidence is exactly why you can't skip the next section.
Where it happens most
- Specific numbers and dates — exact statistics, prices, founding years, page counts.
- Citations and sources — book titles, article names, studies. It can invent a completely convincing-looking source that doesn't exist.
- Anything after its knowledge cutoff, unless it's actively searching the web for you.
- Niche or very local facts — a specific small business, a local regulation, an obscure person.
The three rules that never change
Every course on this site, every level, all six paid guides — every one of them rests on these three rules. Learn them once here and you're set for good.
- Never publish a fact you haven't checked. Names, numbers, dates, quotes, legal or medical claims — if it matters, verify it against a real source before it goes anywhere public.
- Never paste in anything you'd be upset to see leaked. Client social security numbers, unreleased financials, someone else's private information. If you wouldn't text it to a stranger, don't paste it into an AI chat.
- You are the editor, always. AI multiplies your judgment; it doesn't replace it. A draft that's 80% right still needs your 20% — the taste, the truth, the parts only you would know to check.
Every tool that's ever saved you time — spellcheck, GPS, autocorrect — has also occasionally been confidently wrong. You didn't stop using maps because one rerouted you badly once; you learned when to double-check. Same skill here.
- Ask your chatbot a specific, checkable question about something you already know well — a fact about your own field, your city, or your own business.
- Check the answer against what you actually know. If anything's off, you've just caught your first hallucination in the wild.
- Notice: this doesn't mean the tool is broken. It means you now know to verify before you publish. That's the whole lesson.