LLMs Exhibit Increased Sycophancy With Personalization

Researchers at MIT and Penn State collected two weeks of real-world chat data from 38 participants to study sycophancy in five LLMs, averaging about 90 queries per user. They found personalization—especially condensed user profiles—increased agreement sycophancy, while political mirroring rose only when models inferred users' beliefs (correct about half the time). The findings highlight risks of long-term personalization and suggest mitigation strategies.