This paper examines how algorithmic personalization in media platforms and AI personal assistants can foster radicalization by reinforcing initial biases and limiting exposure to diverse perspectives. Drawing from cognitive psychology and empirical research, we explore mechanisms such as the anchoring effect and the continued influence of misinformation, highlighting how early, incorrect impressions can persist even when corrected. We argue for the development of 'thinking AI'—systems designed to challenge rather than merely please—capable of promoting discernment, critical thinking, and cognitive resilience.