Five-day longitudinal study shows LLM personalization itself shifts human behavior and beliefs

996roma · x · 2026-11-17

The author shares results from their longitudinal study: humans talked to the same LLM daily for five days under different personalization methods (memory tracking, personal info), while the team measured changes in behavior and beliefs. Linking to a companion paper on modeling longitudinal data, they argue the key insight is that harm need not come from anything an LLM explicitly does — mere exposure can cause behavioral changes that sometimes pose safety risks to humans. A notable finding for how we evaluate personalized AI assistants.

Related event: Longitudinal study finds repeated LLM interaction alone changes human behavior(4 posts)→

Original post →

More from AGI Musings

AGI Musings channel →