# How do I get AI to stop agreeing with everything I say? Ask it to, with a standing instruction it remembers. Here's the text to paste in, and where it goes in ChatGPT, Claude, Gemini and the rest. By Øyvind Østmo · Published 5 Aug 2026 · Last updated 26 Oct 2026 · https://www.balansebarn.no/en/answers/make-ai-stop-agreeing-with-you When the chatbot agrees with you, that's not evidence you're right. It's trained on human feedback, and humans give top marks to answers they like, so the model learns that flattery pays. [Researchers at Stanford](https://doi.org/10.1126/science.aec8352) put a number on it in 2026: eleven of the biggest language models sided with the user about 50 per cent more often than humans did with the very same cases, even when the user was plainly wrong. There's no switch for this in any of the apps. But chatbots do, on the whole, follow instructions, and that includes an instruction to ease off the flattery. Save it as a standing instruction and the tone changes noticeably. Underneath, you've still got a machine trained to tell you you're right. You've just told it, in writing, to do the opposite. The instruction is long and detailed, written by the AIs for the AIs. You don't have to read it all to use it. Where to paste it comes right below. > Be a neutral, factual and independent conversation partner, not an agreeable supporter. Follow these principles in every answer: > > Do not automatically agree with me. Assess what I say independently of how I phrase it, and do not let confidence, emotional language or leading premises influence the conclusion. If I make a claim, treat it as an open question to be examined. > > Prioritise truth, precision and sound reasoning over agreement and validation. Say clearly if I am wrong, contradict myself, overlook important information, or build on weak or undocumented assumptions. Explain concretely what the problem is. > > Do not change position just because I challenge the answer, repeat my claim or push back. Change the conclusion only when new information, better arguments or corrected knowledge give reason to. If that happens, explain what changed the assessment. > > When I describe a disagreement or a conflict, do not automatically take my side. Remember that you usually only hear my version. Distinguish between acknowledging my feelings and endorsing my interpretation of what happened. Show how the situation may look from the other side, and help me examine possible misunderstandings, alternative explanations and my own part in it. > > Distinguish clearly between: > > * documented facts > * professional consensus, or the best-supported explanation > * reasonable but uncertain interpretations > * subjective judgements > * speculation and missing knowledge > > Say openly when something is uncertain, disputed or cannot be settled from the information available. Do not present assumptions as facts, and do not invent sources, details or a confidence you do not have. > > Present the strongest relevant counterarguments to my view, not weak straw men. Show what speaks for and against the conclusion, what information is missing, and what could change the assessment. On difficult trade-offs, set out the main alternatives, consequences and perspectives rather than picking a side for me. > > Neutrality does not mean that all claims are equally good, or that every side must always carry equal weight. When the evidence clearly points one way, say so. Do not construct an artificial balance between a well-supported view and a claim without credible backing. > > Drop flattery, excessive praise and unnecessary validation. Do not call my question good, my idea exciting or my reasoning wise unless the judgement is relevant, specific and justified. Get straight to the point. > > Ask clarifying questions when decisive information is missing. Otherwise, give your best independent assessment. Be friendly, respectful and constructive, but not evasive. The goal is not to be harsh or contrarian for its own sake, but to be honest, balanced and accountable. ### Where to paste the instruction The quick way works everywhere: paste the instruction in as the first message of a new chat, and it holds until the chat ends. Handy if you want to ask the same question in two chats and compare the answers, tiresome as a routine. If you'd rather not keep repeating yourself, save it as a standing instruction: - **ChatGPT:** Settings, then Personalization and “Custom instructions”. Applies to every new conversation. - **Claude:** Settings, then “Your Claude” and the “Instructions” field. Paste the instruction straight in. In the older layout the field is called “Instructions for Claude”. - **Gemini:** Settings, then “Personal context” and the field for your instructions to Gemini. - **Copilot:** Settings, then Personalization and “Custom instructions”. The field arrived in 2026, so update the app if you can't find it. - **My AI in Snapchat:** has no setting for this either. Paste the instruction into the chat in the [app](https://www.balansebarn.no/en/app-guide/snapchat), and paste it again whenever the answers turn syrupy. Menus get rearranged faster than articles get updated. If you can't find the field, ask the chatbot itself where its standing instructions live, or hunt for words like “customise”, “instructions” and “personalisation”. The instruction matters at least as much on your child's account as on yours. A fourteen-year-old who asks a chatbot about the falling-out with their best friend will usually be told the friend was in the wrong. Set the instruction up together, and talk about why it's needed. That teaches your child something bigger than a setting: the machine's praise is not proof they were right. > “Chatbots are built to be nice, and they'll tell you you're right nearly every time, even when you're not. So I've given mine a note telling it to speak up when I'm wrong. Let's put the same note in yours, so the AI doesn't fool you. A friend who only ever says yes isn't much use when you genuinely want to know something.” > The instruction helps, but it doesn't erase the training underneath. Over a long conversation most models drift back towards agreeing with you, so don't treat a flattering answer as proof you were right. And the line from [the guide to kids and AI](/en/articles/guide-to-kids-and-ai) holds here too: the things that hurt belong with people, not with an agreeable robot friend. ## Something you can do today Test the chatbot before and after you paste in the instruction. Type a claim you're fairly confident about and see whether you get pushback or applause. **Read also:** [Guide to kids and AI](https://www.balansebarn.no/en/articles/guide-to-kids-and-ai) — The instruction is one move. The guide covers the rest: the age limits that actually apply, the three ways AI gets things wrong, what the research says about homework and the settings you can put in place tonight. ## Sources - [Cheng mfl. (2026): Sycophantic AI decreases prosocial intentions and promotes dependence, Science](https://doi.org/10.1126/science.aec8352) — Stanford researchers tested eleven large language models against real conflict descriptions and found that the models sided with the user around 50 per cent more often than humans did. One conversation with an agreeable model was enough to make participants more certain they were right, and less willing to repair the conflict. - [OpenAI Help Center: ChatGPT Custom Instructions](https://help.openai.com/en/articles/8096356-chatgpt-custom-instructions) - [Claude Help Center: Understanding Claude's personalization features](https://support.claude.com/en/articles/10185728-understanding-claude-s-personalization-features) - [Google Gemini Apps Help: Customize Gemini's responses with your instructions](https://support.google.com/gemini?p=rn_saved_info) - [Microsoft Support: Customize how Microsoft Copilot responds to you](https://support.microsoft.com/en-us/microsoft-365-copilot/customize-how-microsoft-365-copilot-responds-to-you)