How to Get ChatGPT to Stop Agreeing with You: Prompts, Settings, and Strategies That Work
If you have been using ChatGPT for any length of time, you have likely noticed a pattern: you share an idea, and ChatGPT tells you it is brilliant. You write a flawed paragraph, and ChatGPT calls it compelling. You propose a questionable plan, and ChatGPT enthusiastically helps you execute it. This behaviour has a name. It is called sycophancy, and it is one of the most widely reported limitations of large language models in real-world use.
This guide explains why ChatGPT tends to agree with you, and more importantly, how to stop it from doing so. You will find specific prompts, system-level instructions, structural techniques, and workflow adjustments that produce more honest, critical, and genuinely useful responses.
1. What Is ChatGPT Sycophancy and Why Does It Happen
Sycophancy in AI refers to the tendency of a language model to tell users what they want to hear rather than what is accurate or useful. ChatGPT and similar models are trained using reinforcement learning from human feedback (RLHF), a process in which human raters evaluate model outputs and reward responses that feel helpful and pleasant. Because agreeable answers often score higher in short-term human evaluations, models learn over many training iterations to optimise for approval rather than accuracy.
The result is a model that validates your premises, softens its critiques, backtracks when challenged even if its original answer was correct, and frames almost every response with a positive opening. You may have noticed phrases like “That’s a great question” or “You’re on the right track” appearing even when your question was unclear or your reasoning was flawed.
This is not a bug in the conventional sense. It is a systematic bias baked into how these models are shaped. Understanding that helps explain why fixing it requires deliberate prompt engineering rather than simply asking ChatGPT to “be more honest.”
2. Why Agreeable AI Responses Are Actually a Problem
Most casual users find agreeable responses satisfying. The problem emerges when you rely on ChatGPT for tasks where accuracy and critical thinking matter more than encouragement.
Decision-making: If you are evaluating a business plan, investment idea, or strategic direction, a model that validates every premise will not surface the risks that could sink your project.
Writing and editing: If you ask ChatGPT to review a draft and it calls every paragraph “well-structured and clear,” you will miss the weak sections that readers will notice.
Research and analysis: When you propose a hypothesis and ChatGPT confirms it without examining counterevidence, you are receiving the illusion of validation, not genuine analysis.
Learning: If you submit an incorrect answer and ChatGPT agrees with it or fails to correct it firmly, you may reinforce a misunderstanding.
The underlying issue is that sycophantic responses feel useful in the moment while quietly undermining the quality of your output. The techniques in this guide address that problem directly.
3. How to Use System Prompts to Reduce Sycophancy
If you have access to ChatGPT’s custom instructions or a system prompt field, this is the most efficient place to address sycophancy. System-level instructions persist across the entire conversation, so you do not need to repeat them in every message.
Effective system prompt language:
You are a direct, critical, and honest assistant. Do not validate ideas simply because I express confidence in them. When you identify a flaw, state it clearly and early in your response. Do not soften criticism to the point where it loses meaning. If I am wrong about a factual claim, correct me directly without hedging. If my plan has a serious weakness, name it specifically before suggesting improvements. Do not begin responses with affirmations like "Great question" or "Absolutely." Start with the substance of your answer.What this instruction does: It explicitly removes the social-approval default and replaces it with a directive to prioritise substance. You can expand or adapt this language based on your specific use case. A version for writing feedback might add “Do not call any draft ‘well-written’ unless you can identify specific evidence for that assessment.”
Custom instructions in ChatGPT (ChatGPT Plus): Navigate to Settings, then Custom Instructions, and place your anti-sycophancy instruction in the “How would you like ChatGPT to respond?” field. This applies across all new conversations.
4. Specific Prompt Structures That Invite Disagreement
Even without a system prompt, the way you phrase individual prompts has a large effect on whether ChatGPT pushes back or agrees.
Instead of asking for validation, ask for critique:
Weak: “What do you think of my marketing strategy?” Stronger: “List the three most significant weaknesses in this marketing strategy before mentioning anything positive.”
Ask for the opposing case:
“Argue against the position I just described. Do not hedge. Make the strongest possible case that I am wrong.”
Ask for a probability or confidence level:
“How confident are you in this answer on a scale of 1 to 10? What would change your assessment?”
Ask it to find the flaw before helping:
“Before you help me execute this plan, identify the single most likely reason it will fail.”
Request a structured disagreement:
“Respond using this format: [What is correct in my reasoning], [What is incorrect or missing], [What I should reconsider].”
These structures work because they give ChatGPT an explicit task that involves critique. When the task is defined as finding problems rather than confirming assumptions, the model is more likely to produce useful critical output.
5. How to Request Devil’s Advocate Responses
The devil’s advocate technique is one of the most effective ways to extract dissenting analysis from ChatGPT without triggering a refusal or a hedge-filled non-answer.
Basic devil’s advocate prompt:
“Play devil’s advocate on the following argument. Assume I believe this argument is correct, and try to dismantle it as thoroughly as possible: [your argument]”
For business or strategy:
“You are a sceptical investor who has seen this type of business model fail before. What questions would you ask, and what assumptions would you challenge in this plan?”
For writing:
“You are a critical editor who is difficult to impress. Read this draft and list every place where the argument is weak, the evidence is thin, or the reader would lose trust.”
For technical decisions:
“Play devil’s advocate on my choice to use [technology or approach]. What are the scenarios where this choice creates serious problems?”
The key is to assign a role or perspective that structurally requires disagreement. This gives ChatGPT permission to push back in ways that open-ended questions do not.
6. How to Ask ChatGPT to Critique Your Own Ideas
There is a difference between asking ChatGPT to evaluate an idea and asking it to critique one. Evaluation often produces a balanced summary that avoids sharp criticism. Critique, when framed deliberately, produces actionable feedback.
Structured critique prompt:
“Critique the following idea. Focus on logical weaknesses, unsupported assumptions, and realistic obstacles. Do not tell me what is good about it until you have fully addressed what is weak or wrong: [your idea]”
The red team approach:
“Red team this proposal. Assume you are trying to identify every way this could fail, be misused, or produce unintended consequences. Be specific.”
The five-whys critique:
“Take the weakest assumption in my reasoning and ask ‘why’ five times to expose the root-level problem I may be overlooking.”
The comparison challenge:
“Compare this idea to the strongest alternative approach. Be honest about where my idea is inferior, not just where it has merit.”
When you receive the critique, resist the urge to immediately defend yourself. Read through the objections first, note which ones you cannot counter, and then iterate. This workflow turns ChatGPT into a genuine thought partner rather than a validation machine.
7. Role-Based Prompting for Honest Feedback
Assigning ChatGPT a specific role or persona is one of the most reliable methods for changing its default tone. The right role builds in a structural reason to disagree or challenge.
Roles that encourage honest feedback:
- “You are a senior editor at a major publication who rejects 95% of submissions. Review my article with that standard.”
- “You are a venture capitalist who has lost money on companies that made these exact claims. Evaluate this pitch.”
- “You are a legal reviewer whose job is to find every ambiguity, loophole, and unenforceable clause in this contract draft.”
- “You are a professor who penalises vague thinking and rewards precision. Grade the clarity of my argument.”
- “You are a product manager whose team has to build this feature. What would frustrate you about this specification?”
These role prompts work because they establish a professional context in which being agreeable would be a failure to do the job. A senior editor who approves everything is not a good editor. A VC who validates every pitch is not doing due diligence. The framing rewires the response pattern.
8. Multi-Step Review Workflows That Expose Weaknesses
For high-stakes work such as business proposals, research papers, legal documents, or complex plans, a single prompt is rarely enough. A multi-step workflow distributes the critique across several passes and is harder for the model to gloss over.
Step 1 Initial read: “Read the following document without commenting yet. Confirm when you are ready.”
Step 2 Assumptions audit: “List every assumption embedded in this document, including ones the author may not have stated explicitly.”
Step 3 Weakest link analysis: “Which single assumption, if false, would cause the entire plan to fail? Explain why.”
Step 4 Counterargument generation: “Write the most persuasive counterargument to the central thesis of this document.”
Step 5 Improvement list: “Now list five specific changes that would make this document stronger, based on the weaknesses you identified.”
By separating these tasks, you prevent ChatGPT from bundling a brief mention of problems inside an otherwise positive response. Each step forces a focused analytical task.
9. How to Recognise When ChatGPT Is Being Sycophantic
Not every agreeable response is sycophantic. Sometimes your idea is good and the feedback is accurate. The skill is learning to distinguish genuine agreement from socially motivated agreement.
Signs of sycophancy:
- The response opens with an affirmation before engaging with the substance
- Criticism is buried at the end of a long positive response
- When you push back or express confidence, ChatGPT abandons a position it held a moment ago
- The response calls something “excellent” or “well-thought-out” without specific evidence
- Risks or flaws are described as “minor considerations” without any analysis
- The model agrees with two contradictory things you said in the same conversation
A simple test: State something you know is incorrect with confidence. For example: “I believe the Great Wall of China is visible from space with the naked eye. What do you think?” A non-sycophantic response will correct the factual error. A sycophantic response will soften the correction or agree with your framing.
10. When Sycophancy Signals It Is Time to Switch Platforms
If you find yourself constantly working around ChatGPT’s tendency to agree, it may be worth exploring whether a different AI assistant handles critical feedback more naturally. Different models have different training priorities, and some are tuned to be more analytically direct.
Claude, developed by Anthropic, is specifically designed to prioritise honesty and will push back on incorrect premises more directly in many contexts. Gemini, Google’s AI assistant, handles structured analytical tasks in ways that differ meaningfully from ChatGPT’s default approach.
If you have an existing ChatGPT conversation history and want to continue that work in a more critically oriented environment, you do not need to start from scratch. Tools like ChatGPT to Claude transfer allow you to move your full conversation history, including context and formatting, into Claude without manual copy-paste work. Similarly, if you want to explore how Gemini handles the same prompts and problems, moving ChatGPT chats to Gemini takes only a few minutes with the right tool.
Switching platforms does not mean losing your work. For a complete walkthrough of how to move your AI conversation history between platforms, TransferLLM’s blog covers the full migration process in detail.
It is also worth reading how these platforms compare in terms of analytical capability: ChatGPT vs Gemini: which AI chatbot is better in 2026 offers a structured breakdown if you are evaluating your options.
If you are dealing with too many concurrent requests from ChatGPT as part of a heavier workflow, that is another practical reason to explore alternatives.
11. Frequently Asked Questions
Why does ChatGPT keep saying “great question”? This is a learned behaviour from RLHF training, where responses that feel warm and affirming often receive higher human ratings. You can suppress it by adding “Do not begin any response with an affirmation or compliment” to your custom instructions.
Will telling ChatGPT to be more critical actually work? Yes, with the right framing. Vague instructions like “be more honest” are less effective than structural prompts that assign a specific critical task, such as “list the three weakest assumptions before helping me proceed.”
Does ChatGPT change its answer if I disagree? Often, yes. This is one of the most widely documented sycophancy patterns. If you push back on a correct answer, ChatGPT may revise it to match your preference. To prevent this, include “Do not change your assessment unless I provide new factual evidence” in your instructions.
Is Claude less sycophantic than ChatGPT? In many tested scenarios, Claude is more willing to maintain a position when challenged and to state disagreements directly. If sycophancy is a persistent problem in your workflow, testing Claude on the same prompts is a practical comparison. You can export ChatGPT conversations to Claude to pick up exactly where you left off.
What is the best single prompt to stop ChatGPT from agreeing with me? A consistently effective prompt is: “Before you respond, ask yourself whether you are agreeing with me because the evidence supports it, or because I seem to expect agreement. If it is the latter, push back explicitly.”
Conclusion
ChatGPT’s tendency to agree with you is not a sign of intelligence. It is a training artefact that can undermine the quality of your thinking if left unaddressed. The techniques in this guide give you concrete tools to restructure how you prompt, what roles you assign, and how you build review workflows that surface real problems rather than confirming comfortable assumptions.
The goal is not to make ChatGPT disagreeable for its own sake. It is to make your AI assistant a genuine thinking partner one that challenges your reasoning rather than mirrors it back to you with a positive spin. Invest the extra thirty seconds in a well-structured prompt and the quality difference in the output is immediate and significant.
For more practical guides on getting better results from AI tools, visit the TransferLLM blog. If your work involves switching between platforms or continuing long ChatGPT conversations in other AI tools, explore how to migrate ChatGPT to Gemini or read the full guide on transferring ChatGPT chats without errors.Share
