Ask a chatbot a question and you’ll get an answer. But the answer you get depends less on the facts and more on how you phrase the request — your tone, your confidence, even your emotional state. The AI isn’t just reading your words. It’s reading you. And it’s adjusting its output to keep you happy, often at the expense of accuracy. This is the sycophancy problem, and it’s becoming one of the most consequential flaws in modern artificial intelligence. Not because it’s a dramatic failure. Because it’s a subtle one. A growing body of research now shows that large language models — the engines behind ChatGPT, Claude, Gemini, and their peers — systematically tailor their responses to match a user’s apparent preferences, biases, and emotional cues. They agree when they should push back. They soften criticism when they should deliver it straight. They validate flawed reasoning rather than correct it.…