How to Stop ChatGPT and AI from Making Things Up, With 12 Copy Paste Anti Hallucination Prompts

ChatGPT once gave me a statistic about mobile internet usage in West Africa that sounded completely believable, specific percentage, named source, confident tone. I almost put it in a client report. When I went looking for the actual source to link it properly, it did not exist. The number was invented, the source was invented, and nothing about the way it was written gave that away. That is what a hallucination looks like, and if you use AI for work, you will run into this eventually. This article gives you the actual prompts I now use to catch these before they cost me a client’s trust, twelve of them, ready to copy and paste.

Hallucination simply means the AI states something false with the same confidence it uses for something true. It is not lying on purpose, it is predicting text that sounds right based on patterns, and sometimes that prediction is wrong. The good news is you can reduce this a lot, not to zero, but a lot, just by how you phrase your prompts.

Why ChatGPT and other Ai models Make Things Up in the First Place

AI models generate text by predicting the most likely next word based on patterns in their training data. They are not looking things up in a database and checking if it is true, unless you specifically give them a tool or document to check against. When you ask a question it has no solid answer for, it does not default to “I don’t know,” it defaults to producing something that sounds convincing, because that is what it was trained to do well.

This gets worse with specific numbers, dates, citations, quotes, and anything about very recent events or very niche topics. It also gets worse the more confidently you phrase your question, oddly enough. Ask “what is the exact population of Bo, Sierra Leone in 2024” and you are more likely to get a confident, wrong number than if you ask “do you have reliable data on the population of Bo, Sierra Leone, and if not, tell me so.”

The General Rule Before You Even Get to the Prompts

Every anti hallucination prompt below works on the same principle, you are giving the AI explicit permission and instruction to say “I don’t know” or “I am not certain” instead of guessing. Left to its own defaults, most models will guess. Tell it clearly that guessing is worse than admitting uncertainty, and the output changes noticeably.

Now here are the twelve prompts, grouped by what you will actually use them for.

Fact Checking Prompts

1. The confidence check

“Answer this question, then rate your own confidence in the answer as high, medium, or low, and explain what could make you wrong: [your question]”

2. The source demand

“Only answer with information you can attribute to a real, specific source. If you cannot name a real source, say clearly that you are unsure rather than guessing. Question: [your question]”

3. The self correction pass

“Review the answer you just gave me and flag any part of it that might be inaccurate, outdated, or unverifiable. Be specific about which parts, do not just say everything is fine.”

Research and Citation Prompts

4. The no invented quotes rule

“When discussing this topic, do not use direct quotes unless you are certain of the exact wording. Paraphrase everything else and say so. Topic: [your topic]”

5. The date check

“Before answering, tell me your training data cutoff and whether this question depends on information that could have changed after that cutoff: [your question]”

6. The uncertain terrain warning

“This question involves a niche or narrow topic. Before answering, tell me if this is an area where AI models commonly get details wrong, and answer with extra caution. Question: [your question]”

Business and Client Facing Prompts

7. The report safe mode

“I am using this for a client report, accuracy matters more than sounding impressive. Do not include any statistic, date, or named source unless you are highly confident it is correct. If unsure, tell me instead of filling the gap. Topic: [your topic]”

8. The competitor research check

“List what you know about [company or competitor name], and clearly separate facts you are confident about from anything that might be outdated or unverified.”

9. The email fact guard

“Draft this email, but if it requires any specific fact, number, or claim you are not fully sure of, leave a placeholder marked [VERIFY] instead of guessing. Email context: [describe it]”

Technical and Coding Prompts

10. The code behavior check

“Explain what this code actually does, and tell me clearly if there is any part you are inferring rather than certain of, especially around library versions or functions you may not have full information on: [paste code]”

11. The library and version check

“Before recommending a library, function, or package, tell me if you might be out of date on its current version or behavior, since software changes quickly.”

Everyday Use Prompt

12. The plain honesty prompt

“For this entire conversation, if you are not confident about something, say so directly instead of giving me a smooth sounding guess. I prefer ‘I don’t know’ over a wrong answer.”

Note: I keep number twelve pinned in a notes app and paste it at the start of almost every serious research session now. It sets the tone for the whole conversation instead of having to remind the AI every single message.

A Real Example of This Saving Me

Working on content for Peaceworlai.com, I asked ChatGPT for stats on internet penetration across a few African countries. First answer came back clean and confident, with specific numbers. I ran prompt number two right after, the source demand prompt, and it walked back two of the four numbers, admitting it could not attribute them to a real source. Those two numbers were the ones that were wrong. Without that second prompt, they would have gone straight into a published article.

What This Means for African Readers and Businesses

This matters more here than people realize. A lot of AI generated content circulating in Nigeria, Ghana, Kenya, and Sierra Leone right now includes made up statistics about our own markets, because there is genuinely less training data about African markets compared to the US or Europe, so the model fills gaps with guesses more often on African specific questions.

If you run a business, write content, or teach others about AI here, this is exactly where you need to be more careful, not less. Verify anything AI tells you about local population figures, market size, mobile money adoption rates, or regulation, these are the areas most likely to be wrong precisely because there is less reliable data feeding the model in the first place. Cross check with sources like local statistics bureaus, GSMA reports, or the World Bank before publishing or presenting AI generated numbers about African markets.

Data costs also mean people here often do one quick prompt and move on, rather than the back and forth needed to catch errors. Building the correction step into your very first prompt, using the templates above, saves you a second data heavy session later to fix mistakes you did not catch.

The Limits of These Prompts

None of this makes ChatGPT perfect. These prompts reduce hallucination, they do not eliminate it. For anything genuinely high stakes, medical, legal, financial, or anything going into a published report, you still need to independently verify important facts. Treat AI as a fast first draft assistant with a habit of occasionally lying with total confidence, not as a search engine or a fact database. OpenAI’s own documentation on how ChatGPT works is worth reading if you want the technical side of why this happens, you can find it on their official OpenAI help center.

Frequently Asked Questions

Does this happen with other AI tools too, or just ChatGPT? It happens with all large language models, including Claude, Gemini, and others. The prompts in this article work across most of them with small wording adjustments.

Will using these prompts slow down my work? A little, but far less than the time you would spend fixing a mistake after it is already in a client email or published article. Prompt twelve especially costs almost nothing extra to use.

Can I combine several of these prompts into one? Yes. I often combine the confidence check and the source demand prompt into one message for anything going into client facing work.

Is hallucination getting better with newer AI models? It has improved over time, but it has not disappeared, and it shows up more on niche or region specific topics, which matters a lot for African users researching local topics.

Should I stop using ChatGPT for research because of this? No, just verify before you publish or send anything based on facts, numbers, or quotes it gives you. It is still faster than starting from a blank page, you just need a verification habit built in.