How agents work · Hallucination
Why AI makes things up (and how to catch it)¶
Sometimes an AI states something false with total confidence — a fact, a quote, a source that doesn't exist. People call this hallucination. It's not lying, and it's not a glitch you can turn off. It's how the tool works, so the fix is a checking habit, not a setting.
Why it happens, in plain words¶
An AI writes by predicting what text should come next, based on patterns it learned. It's very good at producing text that looks right. It has no built-in step that asks "is this actually true?"
So when it doesn't know something, it doesn't feel a gap the way you would. The pattern-machine keeps going — and fills the gap with something plausible. Plausible is the problem: wrong answers arrive in the same confident tone as right ones.
Where it bites hardest¶
- Sources and citations. Invented links, invented titles, real authors attached to papers they never wrote.
- Numbers and dates. Close-but-wrong figures that survive a skim.
- Quotes. Things a real person plausibly could have said, but didn't.
- Niche details. The less material exists about a topic, the more the AI fills in.
- Long jobs. Late in a big task, with a full context window, small inventions creep in.
How to catch it¶
- Click the sources behind anything you'll act on. A source you can't open isn't a source. (More in answers you can check.)
- Check names, numbers, and dates against the original material.
- Ask the same important question a second time, fresh. Two different confident answers means at least one is made up.
- Tell it up front: "say what you don't know — never invent a source." Banning the guess out loud helps.
- Keep the Before you trust it checklist from the research playbook for anything that matters.
What doesn't work¶
- Asking "are you sure?" It will apologize and produce a new confident answer — which may also be wrong.
- Trusting tone. Confidence and correctness are unrelated. That's the whole trap.
- One giant prompt saying "be accurate." Helpful, not sufficient. Checking is still your job.
You're done when¶
You treat every AI answer as a draft until the parts you'll act on are checked — and checking takes you minutes, because you asked for sources up front.
If this helped
If this saves you from acting on one made-up fact, it earned its keep. A small tip on Ko-fi keeps every guide here free. Tip on Ko-fi