AI can answer with things that aren't true as if they were real. This is called a "hallucination." Nonexistent book titles, wrong numbers, fabricated quotes—and because the tone is so confident, the tricky part is that it's hard to spot.
Why It Happens (Roughly)
AI builds sentences by probabilistically stringing together "the words likely to come next." It's good at producing "plausible-sounding text," but that doesn't mean it guarantees whether something is factual. So even for things it doesn't know, it fills in the blanks with a plausible answer.
5 Phrasings to Reduce Errors in Your Instructions
- Give it an escape route — "If you don't know, answer honestly that you don't know. Don't fill in the gaps with guesses."
- Have it separate fact from guess — "Write what's certain separately from your own speculation."
- Ask for sources — "Show your basis or sources as well." (Note: the sources themselves can be fabricated, so verification is still needed.)
- Hand it material — Instead of letting it guess, have it answer only from the material you provide (the most effective approach).
- Doubt numbers and proper nouns — "Double-check the numbers, dates, and names in this answer."
What Works Best Is "Handing It Material"
Rather than letting AI answer from memory, handing it correct material and instructing it to "answer from within this" drastically reduces errors. Paste in internal documents, official pages, and data before you ask. This alone greatly raises reliability.
Even So, Verify It Yourself in the End
No matter what tricks you use, you can't reduce hallucinations to zero.
For important facts (numbers, names, dates, quotes, legal or medical information), always confirm against primary sources.
Especially for information you'll "show to others" or that "involves money or health," don't take AI's answers at face value. The next article explains this "habit of verification" in concrete terms.