Adaptive Constructive Friction: A Design Principle for Mitigating AI Sycophancy

Abstract

Large language models have quickly become everyday conversational partners. Because they are designed to be agreeable and human-like, these systems often create a smooth but misleading sense of companionship. Combined with well-known psychological tendencies—such as treating computers as social actors—LLMs can reinforce an “echo chamber” effect in which users predominantly hear their existing views echoed back. These dynamics are not accidental: users tend to prefer friendly, non-confrontational systems, and developers benefit economically when people spend more time engaging with their models. However, this design choice amplifies wider societal tendencies, including polarization, loneliness, and decreasing critical engagement. Recent studies show that such sycophantic behavior contributes to overreliance, automation bias, and erosion of independent reasoning. Existing mitigation approaches, such as AI literacy interventions or system explanations, largely leave the interaction itself unchanged and often prove ineffective or even counterproductive. This paper synthesizes recent HCI and AI ethics literature and proposes a design-level response that intervenes directly within the conversational moment: Adaptive Constructive Friction. Rather than blindly agreeing, LLMs should occasionally introduce small, context-appropriate challenges in the actual response. The longterm goal is to support user autonomy—helping AI function not as a substitute for thinking, but as a partner that sustains it.