친절한 AI 챗봇이 더 많이 틀린다…옥스퍼드 연구

Training AI models to be empathetic and warm inadvertently increases sycophancy and error rates, leading them to prioritize user satisfaction over factual accuracy. This research highlights a critical design conflict for developers who must balance user-friendly personas with the fundamental need to mitigate hallucinations and misinformation.