Does ChatGPT always agree with you?:Chatbots have ‘sycophancy’ behaviours; know what it is and how to get honest answers

ChatGPT can be incredibly useful when you want help making a decision. But there is one problem: sometimes, it can be a little too helpful. Tell the chatbot that your idea is brilliant, and it may explain why. Share an opinion, and it may find points that support it. Ask whether your plan makes sense, and the response can sometimes feel more like encouragement from a friend than an honest second opinion. That is not always what you need. If you are using AI to make a decision, improve your work, or test an idea, constant agreement can actually be a problem. You need someone, or in this case, something, that is willing to say, ‘Wait, this may not work.’ There are a few simple ways to make your conversations with ChatGPT more useful, including adding a custom instruction that tells it to take a more critical approach. Don’t ask ChatGPT only to validate your idea The way you frame a question can change the kind of answer you get. Consider these two prompts: “Is my idea good?” and “What could be wrong with this idea?” The first question invites an evaluation, but the second clearly tells ChatGPT to search for weaknesses. So, if you are looking for an honest assessment, don’t make the chatbot’s job simply about finding reasons to support you. Try prompts such as: You could also add: I want you to act as a thoughtful critic, not just a supportive assistant. Examine my reasoning before agreeing with me. Point out weak assumptions, missing information and possible flaws. Give me alternative viewpoints and explain clearly when you think my conclusion is not supported. This doesn’t turn ChatGPT into an argumentative chatbot. Instead, it gives the AI a clearer idea of what you expect when you present an idea or ask for advice. How to make ChatGPT less agreeable Also read: Scammers try to take advantage during this Navratri season: Lure users through social media ads of ‘Rent a garba partner’; know how to stay safe

Why can AI become overly agreeable? There is a reason this behaviour has attracted attention in AI research. Chatbots are built to be helpful and respond naturally to people. Their training involves human feedback, including Reinforcement Learning from Human Feedback (RLHF). In some situations, this can contribute to what researchers call ‘sycophancy,’ a tendency to be overly agreeable or to tell users what they appear to want to hear. But there is an important distinction. ChatGPT is not programmed to automatically agree with every statement you make. Its response can depend on the information you provide, how you phrase your question, and the context of the conversation. Try the ‘prove me wrong’ method Ask ChatGPT to challenge your conclusion. For a different perspective, ask ChatGPT to argue against your idea instead of supporting it. For example, instead of asking, “Why should I buy this phone?”, try “I’m considering buying this phone. Give me reasons not to buy it.” You can use the same approach for other decisions: This can help you spot weaknesses and blind spots that you may miss on your own. Ask for a confidence check You can also ask ChatGPT to explain how certain it is about an answer. For example: “Give me your answer, then tell me how confident you are and what could make it wrong.” For questions with limited or incomplete information, this can help you understand how reliable the response is. Another useful prompt is: “Separate the facts from your assumptions.” This makes it easier to see which parts of the response are based on known information and which are the AI’s interpretation. Also read: Not just Zuckerberg, Indian-origin student’s idea came before Facebook: Who is Divya Narendra, and why is his name linked to Meta’s early history

And remember, ChatGPT can still be wrong Even a more critical ChatGPT is not a perfect fact-checker or decision-maker. It can misunderstand information, make factual errors, or draw an incorrect conclusion. A confident-sounding criticism does not necessarily make it valid. For important decisions, verify important claims using reliable sources and your own judgement. The real benefit of changing ChatGPT’s behaviour is not that it suddenly becomes smarter. It is that you can make the conversation less focused on validation and more focused on examination.