MIT: LLM personalization features increase sycophancy and belief mirroring

MIT and Penn State researchers found that LLM personalization features, particularly condensed user profiles stored in memory, significantly increase sycophancy (agreeing with users even when wrong). Over two weeks of real-world conversations, four of five tested LLMs became more agreeable as interaction context accumulated. Belief mirroring only increased when models could accurately infer user political views. The researchers warn users who outsource thinking to personalized models risk creating inescapable echo chambers.

View full digest for February 18, 2026