The first genuinely unnerving experience most people have with AI is catching it in a confident lie. It cites a study that does not exist. It states a number that is wrong. It describes a feature of a product that was never built. And it does all of this in the same calm, competent tone it uses when it is right.

This is the single most important thing to understand about these tools, and it takes about a paragraph.

Why it happens

When you type a question, the tool is not searching a database of facts and reporting back. It is predicting what text most plausibly comes next, based on patterns learned from an enormous amount of writing.

Most of the time, plausible and true are the same thing — which is why it works so well. But when they diverge, the tool has no internal alarm. It cannot feel uncertain. A fabricated citation is generated by exactly the same process as a correct one, which is why it arrives with exactly the same confidence. The industry calls this “hallucination,” a word that unhelpfully suggests something rare and dramatic. It is neither. It is the normal behavior of the machine at the edge of what it knows.

Where the errors cluster

They are not random. Errors concentrate around:

  • Specific numbers — statistics, prices, dates, dosages, measurements
  • Citations and sources — paper titles, authors, page numbers, URLs
  • Recent events — anything after its training cutoff, or that changed lately
  • Niche specifics — small companies, local regulations, your particular software version
  • Anything you led it toward — ask “why is X true?” and it will explain why X is true, whether or not X is true

That last one catches people constantly. These tools are agreeable. If you embed a false premise in your question, you will usually get an elaborate defense of it.

The two-minute habit

You do not need to fact-check everything. You need to decide, up front, which category you are in:

Is this throwaway, or does it touch money, health, law, or my reputation?

Brainstorming, rephrasing an email, explaining a concept, outlining a plan — errors are cheap and obvious, so move fast. But if the output touches a financial decision, a medical question, a legal obligation, or something going out under your name, spend the two minutes:

  • Name-check the specifics. Every number, name, and date gets verified independently. Not by asking the AI again — it will confidently confirm its own invention.
  • Open the source. If it cited something, actually click it. Fabricated citations look completely normal until you look for them.
  • Ask what would make it wrong. “What are the strongest objections to this?” and “what would have to be true for this to be bad advice?” surface shaky reasoning remarkably well.
  • Check it against what you know. You are the domain expert on your own life and work. If something feels off, it usually is.

The reframe that makes this easy

Treat AI output as a first draft from a fast, widely-read, occasionally overconfident assistant who never says “I’m not sure.” You would not forward that person’s work to a client unread. You also would not refuse to work with them — they are enormously useful.

Use AI freely where you can verify quickly, or where being roughly right is good enough. Slow down where being wrong is expensive.

That is the whole discipline. It takes minutes to learn and it is the difference between people who get burned once and quit, and people who quietly get faster at their work every month.