AI will hand you a wrong answer with total confidence, and it looks exactly like a right one. Here are three copy-paste prompts that catch the made-up stuff before you trust it.
You ask AI a question and get a clean, confident answer with a date, a stat, maybe even a source. It sounds right, so you use it. Then you find out it was never true.
That’s a hallucination: AI stating something false while sounding completely sure. A wrong answer looks just like a right one, with no warning sign on it. Here’s why it happens, then the three prompts I use to catch it.
AI isn’t looking things up in a fact book. At its core it predicts the most likely next word, like your phone’s autocomplete but far more powerful. Writing sentences that sound true is a different skill from knowing what’s true.
Anthropic looked inside Claude to see how it decides what to say. Its default is to say it doesn’t have enough information. When it recognizes a topic, a "known entity" signal switches that default off so it can answer. The trouble is when Claude recognizes a name but doesn’t actually know anything else about it. The switch can still flip, and it answers anyway. Anthropic also made a short video explaining this.
The second reason is how models get graded. OpenAI’s researchers compare it to a multiple-choice test: a guess might score points, while "I don’t know" is guaranteed zero. When tests mostly reward accuracy, guessing wins.
When It Happens Most
Exact stats, dates, names, quotes and sources, plus very niche or very recent topics the model saw little of in training. The narrower or newer the question, the more you should double-check.
01
Make it back itself up
Don’t let claims float free. Paste this as your next message, in the same chat, right after any answer with facts, numbers or sources. A lot of made-up answers fall apart the second you ask where they came from.
A source only counts if you can open it and read the line yourself. A dead link, or a page that doesn’t say what was claimed, is a red flag.
02
Give it permission to not know
Take the pressure to answer off the table. Paste this at the start of any chat where being right matters more than being fast. Once "I don’t know" is allowed, the model is more willing to stop instead of filling the gap with a guess.
Want it on every chat? In Claude, add the rules to Settings, then Your Claude, under Instructions. In ChatGPT, add them under Personalization in Settings.
03
The fresh-chat check
This is the strongest one. In the same conversation, AI tends to defend the answer it already gave. A brand new chat didn’t write that answer, so it has no reason to protect it.
Open a new chat (a different AI is even better), paste this prompt, then paste the answer you want checked where it says to.
AI companies are working hard on this, and OpenAI’s own researchers say it’s still a core challenge for every model. So the job of checking important facts is still yours.
My Rule Of Thumb
Low stakes, like a brainstorm or a rough draft? Just go. High stakes, like something you’ll publish, send or decide on? Run the fresh-chat check first.