Study finds LLMs fail to simulate human preferences

New study tested LLMs against thousands of real users and found they systematically fail to capture human preference distributions. The results cast doubt on the growing trend of replacing human feedback with synthetic user data.