BEFORE ANYTHING ELSE

You ask for the steps to change a setting in a program you use every day. The answer is clear, numbered, and confident. Step 4 says to click a menu option that doesn't exist, and nothing hinted it was any less certain than steps 1 to 3.

WHAT WENT WRONG

Nothing, from the tool's side. It produced the most likely-sounding steps, and you treated "sounds right" as "is right."

WHAT CATCHES IT

Trying the steps somewhere safe first, or checking the program's own help.

You ask for five good books on growing tomatoes, with authors and years. Three are real. One has the wrong author. One doesn't exist, though its title sounds exactly like a real book.

WHAT CATCHES IT

Looking up any one title before buying or recommending it. Thirty seconds.

You start with "Keep every answer under 100 words, I'm on my phone." The early answers are short. Forty minutes and many questions later they're long again. You never changed the rule. It faded.

WHAT FIXES IT

Repeat the rule, or start a new chat with the rule at the top.

You spend an evening explaining your small business to an AI tool, open a new chat the next morning, and type "Okay, continue with the plan." The reply is generic, or it asks what plan.

WHAT FIXES IT

Keep your important context somewhere you control and paste it back in. Some tools can remember between chats. Unless you know yours does, assume it doesn't.