3 Prompts That Stop
AI Making Things Up

Sources: Anthropic: how Claude actually thinks → · Watch the video →

You ask AI a question. It gives you 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 is called a hallucination. It is when AI states something false while sounding completely sure of itself. The scary part is that a wrong answer looks just like a right one. There is no warning sign on it. This guide explains in plain English why it happens, and then gives you three prompts you can copy and paste to catch it.

AI is not looking things up in a fact book. At its core, it is predicting the most likely next word, kind of like the autocomplete on your phone but much more powerful. It is very good at writing sentences that sound true. That is a different skill from knowing what is actually true.

Anthropic looked inside their own model to see how it decides what to say. They found something surprising. By default the model has a built-in habit of saying it does not know. But there is a separate "known answer" switch inside it. The problem is that this switch can flip on when the model even half-recognizes a topic. Once it flips, the model answers anyway, even when it does not really have the facts. So it sounds sure, because part of it thinks it knows.

There is a second reason too. These models are trained to be as helpful as possible. Being helpful pushes them to give you an answer instead of saying "I am not sure." So between a habit that turns off the "I don't know" and training that rewards giving answers, the model leans toward guessing.

The first fix is simple: do not let claims float free. Make the AI attach a real source to every single thing it said, then go check that the source actually says it. A lot of made-up answers fall apart the second you ask "where did that come from?"

This works because it forces the model to slow down and look instead of just sounding confident. Anything it cannot back up gets flagged, so you can see exactly which parts to trust and which to throw out.

For every claim you just made, give me the exact source. Then go open each source and confirm it actually says what you claimed. Remove or clearly flag anything you cannot verify.

Show me the final answer in two parts:
1. Claims that are confirmed by a real source, with the source next to each one.
2. Claims you could not verify, clearly labeled so I know not to trust them yet.

Remember, the model is trained to be helpful, and that pressure pushes it to give you something instead of admitting a gap. So take the pressure off. Tell it straight up that "I don't know" is a perfectly good answer, and that you would rather have a gap than a confident wrong answer.

It sounds too simple, but it works. When you remove the pressure to always produce an answer, the model is more willing to stop and say it is not sure, instead of filling the gap with a guess.

Before you answer, here is one rule for this whole conversation:

If you are not sure or do not have the facts, say "I don't know" instead of guessing. I would rather have a gap than a confident wrong answer.

When you are sure, answer normally. When you are not sure, tell me clearly that you are not sure and explain what would be needed to find the real answer. Do not fill gaps with a guess that sounds confident.

This is the strongest one. Open a brand new chat, paste in the answer you got, and ask a fresh AI to tear it apart. No history, no loyalty to the first answer, just a clean second set of eyes.

Here is why a fresh chat matters. In the same conversation, the AI tends to defend its own previous answer instead of catching the mistake. It already committed to it, so it leans toward backing it up. A fresh chat did not write that answer, so it has no reason to protect it. It is far more willing to find the holes.

Here is an answer from another AI. Be skeptical. Find any errors, and check that every source actually backs up the claim it is attached to. List anything that is wrong or unverifiable.

Do not assume it is correct. Your job is to catch mistakes, not to agree.

Give me back:
1. Anything that is factually wrong, and the correct version if you know it.
2. Any source that does not actually support the claim next to it.
3. Anything that cannot be verified and should not be trusted yet.

[Paste the original answer here]

Here is the honest part. Hallucinations are getting better with every model version. The newer models say "I don't know" more often and make things up less. This is a real problem the companies are working hard on, and it is shrinking.

But it is not solved yet. For now, the responsibility to check important facts is still on you. Use AI for speed and ideas, and for anything that really matters, a date, a number, a name, a source, run one of these three prompts first. A few seconds of checking beats sharing a confident wrong answer.