Using AI Safely: Hallucinations, Privacy and Sources

Why AI makes things up

Language models write the most likely next words, not facts looked up in a database. When they lack good information, they can still produce smooth, specific text, complete with names, numbers and references that do not exist. This is called hallucination, and every major assistant does it sometimes, even as newer models improve.

Warning signs include very precise figures without a source, citations you cannot find, quotes attributed to real people, and answers about recent events or niche topics. Errors are also common in arithmetic, dates and legal or medical details.

How to check an answer

Ask for sources, then open them and confirm they exist and actually say what the answer claims. Prefer original sources: the official site, the published study, the government page, the product manual. Cross-check important facts against a second independent source. If the tool has a search or citation feature, use it, but still click through.

For health, legal, tax and financial questions, use AI to prepare good questions and understand terms, then confirm with a qualified professional. Never rely on an unchecked AI answer for anything where a mistake could hurt someone.

Protect your privacy

Think of a chat box like an email to a company. Depending on the service and your settings, conversations may be stored, reviewed for safety, or used to improve future models. Read the privacy settings, turn off history or training use if you prefer, and use workplace accounts that your employer has approved for work material.

Do not paste passwords, bank or card numbers, government ID numbers, medical records or other people's private details. Strip names from documents before asking for a summary when you can. Be careful with browser extensions and apps that ask for broad access to your email or files.

Watch for AI-powered scams

Scammers use the same tools. Voice cloning can imitate a relative asking for urgent money, and AI can write convincing phishing messages without the spelling mistakes people used to look for. Agree on a family code word, call back on a number you already know, and slow down when a message pushes urgency or secrecy. A little skepticism, used consistently, is the best safety feature you have.

← How to Write Good AI PromptsAI Tools by Job: A Practical Guide →