Trust, but Click the Source

AI chatbots are useful. I'm not going to pretend otherwise. They're great for drafting, explaining a concept you half understand, turning messy notes into something readable, or getting unstuck on a bug. But they have one habit that causes almost every horror story you've heard: they sound exactly as sure of themselves when they're wrong as when they're right.

There's no hesitation, no "um," no change in tone. A made-up court case reads just like a real one. A function that doesn't exist in a library looks just like one that does. That's the part that burns people. Not that the tool is wrong sometimes (people are wrong sometimes too), but that nothing in the answer warns you.

So the skill isn't avoiding these tools. It's knowing when to check, and how.

The things most likely to be wrong are specifics. Exact numbers, dates, quotes, statistics, citations, URLs, version numbers, legal and medical details, and anything about recent events. If the answer contains something you'd repeat to another person as a fact, that's the part to verify. The general explanation around it is usually more reliable than the precise details inside it.

How to check, in practice:

Click the sources. If the tool gives links, open them and confirm they actually say what the answer claims. Sometimes the link is real but says something different. Sometimes it doesn't exist at all. If there are no sources, ask for them, and still check them.

Go to the primary source for anything important. The official docs for a piece of software. The actual text of a law or a policy. The real paper, not a summary of it. Your own doctor, lawyer or accountant when it matters.

Run the code. AI-written code can look clean and still be wrong, or call things that don't exist, or quietly do something insecure. Read it, run it, test the edge cases. Be especially careful with anything that touches authentication, user input, file deletion, or database queries. Don't paste in a command you don't understand, particularly one that runs as root or deletes things.

Ask it to argue the other side. "What would make this answer wrong?" or "What are you least sure about here?" It won't catch everything, but it often surfaces the shaky parts.

Watch for answers that are a little too convenient. If you asked a leading question and got exactly the answer you were hoping for, be suspicious. These tools have a tendency to go along with how you framed things.

There's a second way to get burned that has nothing to do with accuracy: what you paste in. Don't put passwords, API keys, private keys or other secrets into a chatbot. Be thoughtful about personal details, other people's information, and anything your employer would consider confidential. Read the privacy settings for whatever tool you use, including whether your conversations can be used for training and how to turn that off if you want to. That varies between services and plans.

And one about your own skills. If you're learning something, letting the AI do all the thinking feels productive and leaves you with nothing. Use it to explain, quiz you, or check your work. Then write the thing yourself.

The way I think about it: an AI answer is a draft from someone smart who never double-checks. Excellent starting point. Your name goes on the final version, though, so you're the one who double-checks.