ChatGPT Inventing Facts? This System Prompt Stops AI Lies

Don't miss any news anymore!

AI tools invent facts, sources, and numbers; this is a well-known problem in office life. However, the right system prompt reduces AI hallucinations. You can find the template to copy at t3n MeisterPrompter.

It is incorrect to trust AI outputs absolutely: ChatGPT, Claude, and other AI tools can hallucinate. This means they provide faulty output, such as unverifiable numbers or quotes. Currently, the problem cannot be entirely solved, but the right system prompt helps to recognize such errors.

Store System Prompt in Basic Settings

The template is available in the t3n MeisterPrompter podcast. A system prompt is stored once in the basic settings of the AI chatbot. It then applies to all chats conducted under the connected account. The goal is to save time, as standard inputs do not need to be repeated with every new request.

You can usually find it under "Settings." In ChatGPT, the system prompt is to be stored under "Custom Instructions" in the "Personalization" section. For those using Claude, the input field can be found in the settings under the "Instructions" section in the "Profile" area. For Google Gemini, system prompts can be added in the settings under "Personal Context" and then in the field "Your Instructions for Gemini."

Input Encourages Disclosure of Ignorance and Avoids Fabricating Content

The prompt consists of a total of eight points. The goal is for the AI to indicate when it lacks information, makes assumptions, or does not know something. The last point is a common problem: AI chatbots typically do not say they do not know something. Instead, they generate a seemingly appropriate but sometimes incorrect response to avoid informing users that they cannot solve a task.

The system prompt begins as follows:

"When information is uncertain, incomplete, or speculative, state this clearly. Do not invent facts, sources, or numbers.

If you do not know something or do not have a reliable basis, respond explicitly: 'I do not know' or 'I do not have verified information on that.'

If an answer is based on assumptions, clearly indicate this."

You can find the complete prompt in the show notes of the linked episode. In the episode "Avoiding AI Lies: How Do I Train ChatGPT and Claude to Stop Hallucinating?" you will also get information on usage, as even the system prompt cannot prevent hallucinations.

Listen to t3n MeisterPrompter – on all platforms

Additional prompt tips with a template for direct use can be found in the t3n MeisterPrompter newsletter. You can listen to the podcast on your preferred podcast platform. Just search for "t3n MeisterPrompter." Alternatively, you can find direct links here:

• t3n MeisterPrompter on Apple Podcast
• t3n MeisterPrompter on Spotify
• RSS Feed

Do you have a suggestion for an episode topic? Write to us at [email protected]. You can also reach Renate on Instagram and TikTok at @renate.gpt or contact us on Instagram or TikTok at @t3n_magazin.

This text was created using our in-house editorial AI tool. Originally published on June 10, 2026, it still interests many of our readers. Therefore, we have updated it and made it available again.