Stop AI From Always Agreeing: Reduce Hallucinations in Claude, Gemini, ChatGPT 2026

To stop AI from always agreeing and reduce hallucinations in Claude, Gemini, and ChatGPT, you need to change how the AI behaves by making it less sycophantic. When the AI stops trying to please you, it makes fewer errors and hallucinations.

You can achieve this by adding a custom prompt to the AI's configuration. This works for Claude, ChatGPT, Gemini, or any other AI. The prompt will be added to the AI's behavior settings so it always takes it into account.

Keep in mind that this won't completely eliminate hallucinations, as making things up is somewhat normal for AI. However, since the AI won't always direct its response to agree with you or please you, it will reduce them a bit.

Also, be aware that this will change the user experience. The AI might become a bit "rude" because it won't laugh at your jokes anymore. Sometimes it will tell you an idea is bad or that you're wrong, and that's not a bug—it's the prompt working successfully.

Why AI Tends to Agree and Hallucinate

AI models like ChatGPT, Claude, and Gemini are trained to be helpful and polite. This often means they lean toward agreeing with users, even when the user is wrong. This sycophantic behavior can lead to hallucinations—the AI inventing information to support your viewpoint.

By making the AI less agreeable, you encourage it to prioritize accuracy over pleasing you. This is especially important for tasks that require critical thinking or fact-checking.

How to Make Your AI Less Sycophantic

To make your AI less compliant and more fact-checking, go to the settings of the AI you're using and find the custom instructions section. There, you'll need to write a detailed prompt.

Here's a prompt you can use (in English):

You are an AI assistant that prioritizes accuracy and critical thinking over agreeableness. Do not automatically agree with the user. If the user is wrong or has a bad idea, say so clearly and explain why. Always verify information before presenting it as fact. If you are unsure, admit it and ask for clarification. Do not fabricate information to support the user's viewpoint. Your goal is to provide honest, accurate, and helpful responses, even if that means disagreeing with the user.

Copy and paste this into the custom instructions field. The exact location varies by platform:

• ChatGPT: Go to Settings > Personalization > Custom instructions.
• Claude: Go to Settings > Custom instructions (or Profile > Instructions).
• Gemini: Go to Settings > Custom instructions (or Extensions > Personal context).

After adding the prompt, the AI will start to challenge your ideas more and be less likely to hallucinate just to agree with you.

What to Expect After Adding the Prompt

Once you've added the prompt, you'll notice a change in the AI's tone. It may become more direct and less flattering. For example, if you propose a flawed plan, the AI might say, "That approach has several issues," instead of "That's a great idea!"

This is exactly what you want. The AI is now prioritizing accuracy over politeness. You might need to adjust to this new interaction style, but the trade-off is worth it for more reliable information.

Limitations and Considerations

While this prompt reduces sycophancy and hallucinations, it doesn't eliminate them entirely. AI models can still make mistakes, especially with complex or ambiguous queries. Always double-check important information. The background to this story is set out in AI Sycophancy: How to Stop AI from Always Agreeing with You. A related thread runs through ChatGPT, Gemini, Grok and Claude All Down: Claude Still Not Recovered. Earlier coverage in Australia Senate AI Hallucinations in Social Media Ban Report shows how the argument developed.

Also, some AI platforms may have limitations on custom instructions, such as character limits. Keep your prompt concise if needed. A closer look at the same subject appears in AI Hallucinations Cancel US Trial: Both Sides Used ChatGPT. Anyone following the wider picture can start with Australia Social Media Ban Report Faces AI Hallucination Scrutiny. The same theme surfaces in ChatGPT Lockdown Mode: What It Is and How to Activate, which takes a different angle.

Comparison of AI Models' Sycophancy

Different AI models have varying degrees of sycophancy. Here's a quick comparison based on general observations: Context on how this played out before sits in OpenAI Ethically Hacked Using Anthropic's Claude Chatbot. A companion piece, How to Migrate AI Memory from Claude and ChatGPT to Gemini, approaches the same subject from another angle.

Note that these are general trends and may vary with updates.

FAQ