Hallucinations and privacy
Lesson 8 of 23 in the course “AI basics”. Free, with animations and a quiz.
AI can make things up
Sometimes the model invents facts, numbers, quotes and even links to articles that don't exist. These made-up bits are called hallucinations. Example: the model writes “according to a 2019 study…”, but the link goes nowhere because it invented the study. Hallucinations look just as convincing as the truth. Any model can slip like this.
Why AI makes things up
The model doesn't look for the truth; it picks a plausible continuation of the text. If it doesn't have the exact answer, it will still write something that looks like one. It can say “I don't know”, but not always. Made-up stuff shows up most in rare facts, exact numbers, dates, quotes and links. That's exactly where to be extra careful.
How to check
Always check numbers, dates, laws, health and money. The original source is where a fact first appeared: a government website, the actual research paper or an official document. Ask the model for a source, then open the link yourself and make sure it says the same thing. Link doesn't open, or it says something else? Then you can't rely on that fact.
Don't overshare
Never paste passwords, card numbers, seed phrases or other people's personal data into a chat. A seed phrase is usually 12 or 24 words that give full access to a crypto wallet. Whatever you send goes to the company's servers and may be stored there. So it's no longer only yours. Simple rule: don't type into a chat what you wouldn't show a stranger.
If you already sent it
Don't panic, but act fast. Sent a password? Change it right away and turn on two-factor authentication, meaning login with a code from an app or SMS. Card number? Call your bank and get the card reissued. Seed phrase? Create a new wallet and move everything there; treat the old one as lost. Deleting the chat helps, but a copy may already be saved.
Check yourself
- The AI gave you a link to a study. What now?
Correct answer: Open it and check it says the same thing. Links can be made up too. Asking the model isn't enough: it may confidently confirm its own invention. - What is an AI hallucination?
Correct answer: A made-up fact presented as true. The model picks what sounds plausible, not necessarily what is true. That's why invented facts look real. - What is fine to paste into an AI chat?
Correct answer: Your own history notes to review. Your own study notes are fine. Passwords, seed phrases, card details and other people's data stay out of the chat.
← How a chatbot is trained: 3 stages How to tell if an AI is good →